npm - oidn-web - Versions diffs - 0.1.0 → 0.1.1 - Mend

oidn-web 0.1.0 → 0.1.1

This diff represents the content of publicly available package versions that have been released to one of the supported registries. The information contained in this diff is provided for informational purposes only and reflects changes between package versions as they appear in their respective public registries.

Files changed (7) hide show

package/dist/oidn.mjs +602 -605
package/dist/oidn.umd.js +137 -137
package/lib/UNet.js +5 -2
package/lib/UNet.js.map +1 -1
package/package.json +1 -1
package/src/UNet.ts +14 -14
package/weights/rt_hdr_calb_cnrm.tza +0 -0

package/dist/oidn.umd.js CHANGED Viewed

@@ -1,4 +1,4 @@
-(function(Nt,qt){typeof exports=="object"&&typeof module<"u"?qt(exports):typeof define=="function"&&define.amd?define(["exports"],qt):(Nt=typeof globalThis<"u"?globalThis:Nt||self,qt(Nt.oidn={}))})(this,function(Nt){"use strict";var xv=Object.defineProperty;var Sv=(Nt,qt,Xe)=>qt in Nt?xv(Nt,qt,{enumerable:!0,configurable:!0,writable:!0,value:Xe}):Nt[qt]=Xe;var K=(Nt,qt,Xe)=>(Sv(Nt,typeof qt!="symbol"?qt+"":qt,Xe),Xe);function qt(n,t){for(var e=0;e<t.length;e++){const s=t[e];if(typeof s!="string"&&!Array.isArray(s)){for(const r in s)if(r!=="default"&&!(r in n)){const o=Object.getOwnPropertyDescriptor(s,r);o&&Object.defineProperty(n,r,o.get?o:{enumerable:!0,get:()=>s[r]})}}}return Object.freeze(Object.defineProperty(n,Symbol.toStringTag,{value:"Module"}))}class Xe{constructor(){K(this,"dims",[]);K(this,"paddedDims",[]);K(this,"layout","x");K(this,"dataType","Float32")}getByteSize(){let t=1;for(const e of this.paddedDims)t*=e;return this.dataType==="Float32"?t*=4:this.dataType==="Float16"&&(t*=2),t}}class wf{constructor(t,e){this.desc=t,this.data=e}}class xf{constructor(t){K(this,"offset",0);this._view=t}read(t){const e=this._view,s=this.offset;switch(this.offset+=t,t){case 1:return e.getUint8(s);case 2:return e.getUint16(s,!0);case 4:return e.getUint32(s,!0);case 8:return Number(e.getBigUint64(s,!0));default:throw new Error("unsupported read size")}}}function ga(n){const t=new Uint8Array(n),e=new xf(new DataView(n));if(e.read(2)!==16855)throw new Error("invalid or corrupted weights blob");const r=e.read(1);if(e.read(1),r!==2)throw new Error("unsupported weights blob version");const o=e.read(8);e.offset=o;const i=e.read(4),a=new Map;for(let l=0;l<i;++l){const u=new Xe,c=e.read(2),h=new TextDecoder().decode(t.subarray(e.offset,e.offset+c));e.offset+=c;const f=e.read(1);for(let b=0;b<f;++b)u.dims.push(e.read(4));u.paddedDims=[...u.dims],new TextDecoder().decode(t.subarray(e.offset,e.offset+f))==="oihw"&&(u.layout="oihw"),e.offset+=f;const p=String.fromCharCode(e.read(1));if(p==="f")u.dataType="Float32";else if(p==="h")u.dataType="Float16";else throw new Error("invalid tensor data type");const g=e.read(8),m=t.slice(g,g+u.getByteSize());a.set(h,new wf(u,m))}return a}/**
+(function(Nt,Ht){typeof exports=="object"&&typeof module<"u"?Ht(exports):typeof define=="function"&&define.amd?define(["exports"],Ht):(Nt=typeof globalThis<"u"?globalThis:Nt||self,Ht(Nt.oidn={}))})(this,function(Nt){"use strict";var xv=Object.defineProperty;var Sv=(Nt,Ht,Je)=>Ht in Nt?xv(Nt,Ht,{enumerable:!0,configurable:!0,writable:!0,value:Je}):Nt[Ht]=Je;var K=(Nt,Ht,Je)=>(Sv(Nt,typeof Ht!="symbol"?Ht+"":Ht,Je),Je);function Ht(n,t){for(var e=0;e<t.length;e++){const s=t[e];if(typeof s!="string"&&!Array.isArray(s)){for(const r in s)if(r!=="default"&&!(r in n)){const o=Object.getOwnPropertyDescriptor(s,r);o&&Object.defineProperty(n,r,o.get?o:{enumerable:!0,get:()=>s[r]})}}}return Object.freeze(Object.defineProperty(n,Symbol.toStringTag,{value:"Module"}))}class Je{constructor(){K(this,"dims",[]);K(this,"paddedDims",[]);K(this,"layout","x");K(this,"dataType","Float32")}getByteSize(){let t=1;for(const e of this.paddedDims)t*=e;return this.dataType==="Float32"?t*=4:this.dataType==="Float16"&&(t*=2),t}}class wf{constructor(t,e){this.desc=t,this.data=e}}class xf{constructor(t){K(this,"offset",0);this._view=t}read(t){const e=this._view,s=this.offset;switch(this.offset+=t,t){case 1:return e.getUint8(s);case 2:return e.getUint16(s,!0);case 4:return e.getUint32(s,!0);case 8:return Number(e.getBigUint64(s,!0));default:throw new Error("unsupported read size")}}}function ga(n){const t=new Uint8Array(n),e=new xf(new DataView(n));if(e.read(2)!==16855)throw new Error("invalid or corrupted weights blob");const r=e.read(1);if(e.read(1),r!==2)throw new Error("unsupported weights blob version");const o=e.read(8);e.offset=o;const i=e.read(4),a=new Map;for(let l=0;l<i;++l){const u=new Je,c=e.read(2),h=new TextDecoder().decode(t.subarray(e.offset,e.offset+c));e.offset+=c;const f=e.read(1);for(let b=0;b<f;++b)u.dims.push(e.read(4));u.paddedDims=[...u.dims],new TextDecoder().decode(t.subarray(e.offset,e.offset+f))==="oihw"&&(u.layout="oihw"),e.offset+=f;const p=String.fromCharCode(e.read(1));if(p==="f")u.dataType="Float32";else if(p==="h")u.dataType="Float16";else throw new Error("invalid tensor data type");const g=e.read(8),m=t.slice(g,g+u.getByteSize());a.set(h,new wf(u,m))}return a}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -28,7 +28,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function If(n){let t=n.length,e=0;for(;t>0;)e=Math.random()*t|0,t--,En(n,t,e)}function En(n,t,e){const s=n[t];n[t]=n[e],n[e]=s}function Af(n){let t=0;for(let e=0;e<n.length;e++)t+=n[e];return t}function w(n,t){if(!n)throw new Error(typeof t=="string"?t:t())}function Ef(n,t,e=""){w(Zt(n,t),()=>e+` Shapes ${n} and ${t} must match`)}function ya(n){w(n!=null,()=>"The input to the tensor constructor must be a non-null value.")}function z(n){if(n.length===0)return 1;let t=n[0];for(let e=1;e<n.length;e++)t*=n[e];return t}function Zt(n,t){if(n===t)return!0;if(n==null||t==null||n.length!==t.length)return!1;for(let e=0;e<n.length;e++)if(n[e]!==t[e])return!1;return!0}function oo(n){return n%1===0}function Us(n,t){return t<=n.length?n:n+" ".repeat(t-n.length)}function Cf(n,t){let e=1,s=-1;for(let o=0;o<n.length;++o)if(n[o]>=0)e*=n[o];else if(n[o]===-1){if(s!==-1)throw Error(`Shapes can only have 1 implicit size. Found -1 at dim ${s} and dim ${o}`);s=o}else if(n[o]<0)throw Error(`Shapes can not be < 0. Found ${n[o]} at dim ${o}`);if(s===-1){if(t>0&&t!==e)throw Error(`Size(${t}) must match the product of shape ${n}`);return n}if(e===0)throw Error(`Cannot infer the missing size in [${n}] when there are 0 elements`);if(t%e!==0)throw Error(`The implicit shape can't be a fractional number. Got ${t} / ${e}`);const r=n.slice();return r[s]=t/e,r}function os(n,t){const e=t.length;return n=n==null?t.map((s,r)=>r):[].concat(n),w(n.every(s=>s>=-e&&s<e),()=>`All values in axis param must be in range [-${e}, ${e}) but got axis ${n}`),w(n.every(s=>oo(s)),()=>`All values in axis param must be integers but got axis ${n}`),n.map(s=>s<0?e+s:s)}function kf(n,t){const e=[],s=[],r=t!=null&&Array.isArray(t)&&t.length===0,o=t==null||r?null:os(t,n).sort();let i=0;for(let a=0;a<n.length;++a){if(o!=null){if(o[i]===a&&n[a]!==1)throw new Error(`Can't squeeze axis ${a} since its dim '${n[a]}' is not 1`);(o[i]==null||o[i]>a)&&n[a]===1&&(e.push(n[a]),s.push(a)),o[i]<=a&&i++}n[a]!==1&&(e.push(n[a]),s.push(a))}return{newShape:e,keptDims:s}}function Cn(n,t){return bt(n,t)}function bt(n,t){let e=null;if(n==null||n==="float32")e=new Float32Array(t);else if(n==="int32")e=new Int32Array(t);else if(n==="bool")e=new Uint8Array(t);else if(n==="string")e=new Array(t);else throw new Error(`Unknown data type ${n}`);return e}function _f(n,t){for(let e=0;e<n.length;e++){const s=n[e];if(isNaN(s)||!isFinite(s))throw Error(`A tensor of type ${t} being uploaded contains ${s}.`)}}function Tf(n){return n==="bool"||n==="complex64"||n==="float32"||n==="int32"||n==="string"}function io(n){if(n==="float32"||n==="int32")return 4;if(n==="complex64")return 8;if(n==="bool")return 1;throw new Error(`Unknown dtype ${n}`)}function Nf(n){if(n==null)return 0;let t=0;return n.forEach(e=>t+=e.length),t}function Ws(n){return typeof n=="string"||n instanceof String}function Df(n){return typeof n=="boolean"}function ao(n){return typeof n=="number"}function is(n){return Array.isArray(n)?is(n[0]):n instanceof Float32Array?"float32":n instanceof Int32Array||n instanceof Uint8Array||n instanceof Uint8ClampedArray?"int32":ao(n)?"float32":Ws(n)?"string":Df(n)?"bool":"float32"}function lo(n){return!!(n&&n.constructor&&n.call&&n.apply)}function jt(n){const t=n.length;if(t<2)return[];const e=new Array(t-1);e[t-2]=n[t-1];for(let s=t-3;s>=0;--s)e[s]=e[s+1]*n[s+1];return e}function wa(n,t,e,s=!1){const r=new Array;if(t.length===1){const o=t[0]*(s?2:1);for(let i=0;i<o;i++)r[i]=e[n+i]}else{const o=t[0],i=t.slice(1),a=i.reduce((l,u)=>l*u)*(s?2:1);for(let l=0;l<o;l++)r[l]=wa(n+l*a,i,e,s)}return r}function xa(n,t,e=!1){if(n.length===0)return t[0];const s=n.reduce((r,o)=>r*o)*(e?2:1);if(s===0)return[];if(s!==t.length)throw new Error(`[${n}] does not match the input size ${t.length}${e?" for a complex tensor":""}.`);return wa(0,n,t,e)}function uo(n,t){if(Array.isArray(n))return n;if(t==="float32")return n instanceof Float32Array?n:new Float32Array(n);if(t==="int32")return n instanceof Int32Array?n:new Int32Array(n);if(t==="bool"||t==="string")return Uint8Array.from(new Int32Array(n));throw new Error(`Unknown dtype ${t}`)}function Sa(n,t){const e=ze(n,t);for(let s=0;s<e.length;s++)e[s]=1;return e}function ze(n,t){if(t==null||t==="float32"||t==="complex64")return new Float32Array(n);if(t==="int32")return new Int32Array(n);if(t==="bool")return new Uint8Array(n);throw new Error(`Unknown data type ${t}`)}function Re(n){n.forEach(t=>{w(Number.isInteger(t)&&t>=0,()=>`Tensor must have a shape comprised of positive integers but got shape [${n}].`)})}function co(n,t,e){if(t===0)return 0;if(t===1)return n[0];let s=n[n.length-1];for(let r=0;r<n.length-1;++r)s+=e[r]*n[r];return s}function ho(n,t,e){if(t===0)return[];if(t===1)return[n];const s=new Array(t);for(let r=0;r<s.length-1;++r)s[r]=Math.floor(n/e[r]),n-=s[r]*e[r];return s[s.length-1]=n,s}function fo(n){return n&&n.then&&typeof n.then=="function"}/**
+ */function If(n){let t=n.length,e=0;for(;t>0;)e=Math.random()*t|0,t--,En(n,t,e)}function En(n,t,e){const s=n[t];n[t]=n[e],n[e]=s}function Af(n){let t=0;for(let e=0;e<n.length;e++)t+=n[e];return t}function w(n,t){if(!n)throw new Error(typeof t=="string"?t:t())}function Ef(n,t,e=""){w(Zt(n,t),()=>e+` Shapes ${n} and ${t} must match`)}function ya(n){w(n!=null,()=>"The input to the tensor constructor must be a non-null value.")}function z(n){if(n.length===0)return 1;let t=n[0];for(let e=1;e<n.length;e++)t*=n[e];return t}function Zt(n,t){if(n===t)return!0;if(n==null||t==null||n.length!==t.length)return!1;for(let e=0;e<n.length;e++)if(n[e]!==t[e])return!1;return!0}function oo(n){return n%1===0}function Ws(n,t){return t<=n.length?n:n+" ".repeat(t-n.length)}function Cf(n,t){let e=1,s=-1;for(let o=0;o<n.length;++o)if(n[o]>=0)e*=n[o];else if(n[o]===-1){if(s!==-1)throw Error(`Shapes can only have 1 implicit size. Found -1 at dim ${s} and dim ${o}`);s=o}else if(n[o]<0)throw Error(`Shapes can not be < 0. Found ${n[o]} at dim ${o}`);if(s===-1){if(t>0&&t!==e)throw Error(`Size(${t}) must match the product of shape ${n}`);return n}if(e===0)throw Error(`Cannot infer the missing size in [${n}] when there are 0 elements`);if(t%e!==0)throw Error(`The implicit shape can't be a fractional number. Got ${t} / ${e}`);const r=n.slice();return r[s]=t/e,r}function is(n,t){const e=t.length;return n=n==null?t.map((s,r)=>r):[].concat(n),w(n.every(s=>s>=-e&&s<e),()=>`All values in axis param must be in range [-${e}, ${e}) but got axis ${n}`),w(n.every(s=>oo(s)),()=>`All values in axis param must be integers but got axis ${n}`),n.map(s=>s<0?e+s:s)}function kf(n,t){const e=[],s=[],r=t!=null&&Array.isArray(t)&&t.length===0,o=t==null||r?null:is(t,n).sort();let i=0;for(let a=0;a<n.length;++a){if(o!=null){if(o[i]===a&&n[a]!==1)throw new Error(`Can't squeeze axis ${a} since its dim '${n[a]}' is not 1`);(o[i]==null||o[i]>a)&&n[a]===1&&(e.push(n[a]),s.push(a)),o[i]<=a&&i++}n[a]!==1&&(e.push(n[a]),s.push(a))}return{newShape:e,keptDims:s}}function Cn(n,t){return bt(n,t)}function bt(n,t){let e=null;if(n==null||n==="float32")e=new Float32Array(t);else if(n==="int32")e=new Int32Array(t);else if(n==="bool")e=new Uint8Array(t);else if(n==="string")e=new Array(t);else throw new Error(`Unknown data type ${n}`);return e}function _f(n,t){for(let e=0;e<n.length;e++){const s=n[e];if(isNaN(s)||!isFinite(s))throw Error(`A tensor of type ${t} being uploaded contains ${s}.`)}}function Tf(n){return n==="bool"||n==="complex64"||n==="float32"||n==="int32"||n==="string"}function io(n){if(n==="float32"||n==="int32")return 4;if(n==="complex64")return 8;if(n==="bool")return 1;throw new Error(`Unknown dtype ${n}`)}function Nf(n){if(n==null)return 0;let t=0;return n.forEach(e=>t+=e.length),t}function Gs(n){return typeof n=="string"||n instanceof String}function Df(n){return typeof n=="boolean"}function ao(n){return typeof n=="number"}function as(n){return Array.isArray(n)?as(n[0]):n instanceof Float32Array?"float32":n instanceof Int32Array||n instanceof Uint8Array||n instanceof Uint8ClampedArray?"int32":ao(n)?"float32":Gs(n)?"string":Df(n)?"bool":"float32"}function lo(n){return!!(n&&n.constructor&&n.call&&n.apply)}function Kt(n){const t=n.length;if(t<2)return[];const e=new Array(t-1);e[t-2]=n[t-1];for(let s=t-3;s>=0;--s)e[s]=e[s+1]*n[s+1];return e}function wa(n,t,e,s=!1){const r=new Array;if(t.length===1){const o=t[0]*(s?2:1);for(let i=0;i<o;i++)r[i]=e[n+i]}else{const o=t[0],i=t.slice(1),a=i.reduce((l,u)=>l*u)*(s?2:1);for(let l=0;l<o;l++)r[l]=wa(n+l*a,i,e,s)}return r}function xa(n,t,e=!1){if(n.length===0)return t[0];const s=n.reduce((r,o)=>r*o)*(e?2:1);if(s===0)return[];if(s!==t.length)throw new Error(`[${n}] does not match the input size ${t.length}${e?" for a complex tensor":""}.`);return wa(0,n,t,e)}function uo(n,t){if(Array.isArray(n))return n;if(t==="float32")return n instanceof Float32Array?n:new Float32Array(n);if(t==="int32")return n instanceof Int32Array?n:new Int32Array(n);if(t==="bool"||t==="string")return Uint8Array.from(new Int32Array(n));throw new Error(`Unknown dtype ${t}`)}function Sa(n,t){const e=ze(n,t);for(let s=0;s<e.length;s++)e[s]=1;return e}function ze(n,t){if(t==null||t==="float32"||t==="complex64")return new Float32Array(n);if(t==="int32")return new Int32Array(n);if(t==="bool")return new Uint8Array(n);throw new Error(`Unknown data type ${t}`)}function Re(n){n.forEach(t=>{w(Number.isInteger(t)&&t>=0,()=>`Tensor must have a shape comprised of positive integers but got shape [${n}].`)})}function co(n,t,e){if(t===0)return 0;if(t===1)return n[0];let s=n[n.length-1];for(let r=0;r<n.length-1;++r)s+=e[r]*n[r];return s}function ho(n,t,e){if(t===0)return[];if(t===1)return[n];const s=new Array(t);for(let r=0;r<s.length-1;++r)s[r]=Math.floor(n/e[r]),n-=s[r]*e[r];return s[s.length-1]=n,s}function fo(n){return n&&n.then&&typeof n.then=="function"}/**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -88,7 +88,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const Gs=mo("kernelRegistry",()=>new Map),ap=mo("gradRegistry",()=>new Map);function Pa(n,t){const e=Oa(n,t);return Gs.get(e)}function La(n){return ap.get(n)}function Ma(n){const t=Gs.entries(),e=[];for(;;){const{done:s,value:r}=t.next();if(s)break;const[o,i]=r,[a]=o.split("_");a===n&&e.push(i)}return e}function lp(n){const{kernelName:t,backendName:e}=n,s=Oa(t,e);Gs.has(s)&&kn(`The kernel '${t}' for backend '${e}' is already registered`),Gs.set(s,n)}function Oa(n,t){return`${t}_${n}`}/**
+ */const Vs=mo("kernelRegistry",()=>new Map),ap=mo("gradRegistry",()=>new Map);function Pa(n,t){const e=Oa(n,t);return Vs.get(e)}function La(n){return ap.get(n)}function Ma(n){const t=Vs.entries(),e=[];for(;;){const{done:s,value:r}=t.next();if(s)break;const[o,i]=r,[a]=o.split("_");a===n&&e.push(i)}return e}function lp(n){const{kernelName:t,backendName:e}=n,s=Oa(t,e);Vs.has(s)&&kn(`The kernel '${t}' for backend '${e}' is already registered`),Vs.set(s,n)}function Oa(n,t){return`${t}_${n}`}/**
  * @license
  * Copyright 2023 Google LLC.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -103,7 +103,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Ba(n){return n instanceof Float32Array||n instanceof Int32Array||n instanceof Uint8Array||n instanceof Uint8ClampedArray}var Je=typeof globalThis<"u"?globalThis:typeof window<"u"?window:typeof global<"u"?global:typeof self<"u"?self:{};function up(n){return n&&n.__esModule&&Object.prototype.hasOwnProperty.call(n,"default")?n.default:n}function cp(n){if(n.__esModule)return n;var t=n.default;if(typeof t=="function"){var e=function s(){return this instanceof s?Reflect.construct(t,arguments,this.constructor):t.apply(this,arguments)};e.prototype=t.prototype}else e={};return Object.defineProperty(e,"__esModule",{value:!0}),Object.keys(n).forEach(function(s){var r=Object.getOwnPropertyDescriptor(n,s);Object.defineProperty(e,s,r.get?r:{enumerable:!0,get:function(){return n[s]}})}),e}var Fa=it,Qt=null;try{Qt=new WebAssembly.Instance(new WebAssembly.Module(new Uint8Array([0,97,115,109,1,0,0,0,1,13,2,96,0,1,127,96,4,127,127,127,127,1,127,3,7,6,0,1,1,1,1,1,6,6,1,127,1,65,0,11,7,50,6,3,109,117,108,0,1,5,100,105,118,95,115,0,2,5,100,105,118,95,117,0,3,5,114,101,109,95,115,0,4,5,114,101,109,95,117,0,5,8,103,101,116,95,104,105,103,104,0,0,10,191,1,6,4,0,35,0,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,126,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,127,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,128,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,129,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,130,34,4,66,32,135,167,36,0,32,4,167,11])),{}).exports}catch{}function it(n,t,e){this.low=n|0,this.high=t|0,this.unsigned=!!e}it.prototype.__isLong__,Object.defineProperty(it.prototype,"__isLong__",{value:!0});function Bt(n){return(n&&n.__isLong__)===!0}it.isLong=Bt;var za={},Ua={};function Ze(n,t){var e,s,r;return t?(n>>>=0,(r=0<=n&&n<256)&&(s=Ua[n],s)?s:(e=at(n,(n|0)<0?-1:0,!0),r&&(Ua[n]=e),e)):(n|=0,(r=-128<=n&&n<128)&&(s=za[n],s)?s:(e=at(n,n<0?-1:0,!1),r&&(za[n]=e),e))}it.fromInt=Ze;function te(n,t){if(isNaN(n))return t?Qe:ee;if(t){if(n<0)return Qe;if(n>=Ga)return Ka}else{if(n<=-Va)return Ft;if(n+1>=Va)return Ha}return n<0?te(-n,t).neg():at(n%_n|0,n/_n|0,t)}it.fromNumber=te;function at(n,t,e){return new it(n,t,e)}it.fromBits=at;var Vs=Math.pow;function wo(n,t,e){if(n.length===0)throw Error("empty string");if(n==="NaN"||n==="Infinity"||n==="+Infinity"||n==="-Infinity")return ee;if(typeof t=="number"?(e=t,t=!1):t=!!t,e=e||10,e<2||36<e)throw RangeError("radix");var s;if((s=n.indexOf("-"))>0)throw Error("interior hyphen");if(s===0)return wo(n.substring(1),t,e).neg();for(var r=te(Vs(e,8)),o=ee,i=0;i<n.length;i+=8){var a=Math.min(8,n.length-i),l=parseInt(n.substring(i,i+a),e);if(a<8){var u=te(Vs(e,a));o=o.mul(u).add(te(l))}else o=o.mul(r),o=o.add(te(l))}return o.unsigned=t,o}it.fromString=wo;function ue(n,t){return typeof n=="number"?te(n,t):typeof n=="string"?wo(n,t):at(n.low,n.high,typeof t=="boolean"?t:n.unsigned)}it.fromValue=ue;var Wa=65536,hp=1<<24,_n=Wa*Wa,Ga=_n*_n,Va=Ga/2,qa=Ze(hp),ee=Ze(0);it.ZERO=ee;var Qe=Ze(0,!0);it.UZERO=Qe;var Tn=Ze(1);it.ONE=Tn;var ja=Ze(1,!0);it.UONE=ja;var xo=Ze(-1);it.NEG_ONE=xo;var Ha=at(-1,2147483647,!1);it.MAX_VALUE=Ha;var Ka=at(-1,-1,!0);it.MAX_UNSIGNED_VALUE=Ka;var Ft=at(0,-2147483648,!1);it.MIN_VALUE=Ft;var R=it.prototype;R.toInt=function(){return this.unsigned?this.low>>>0:this.low},R.toNumber=function(){return this.unsigned?(this.high>>>0)*_n+(this.low>>>0):this.high*_n+(this.low>>>0)},R.toString=function(t){if(t=t||10,t<2||36<t)throw RangeError("radix");if(this.isZero())return"0";if(this.isNegative())if(this.eq(Ft)){var e=te(t),s=this.div(e),r=s.mul(e).sub(this);return s.toString(t)+r.toInt().toString(t)}else return"-"+this.neg().toString(t);for(var o=te(Vs(t,6),this.unsigned),i=this,a="";;){var l=i.div(o),u=i.sub(l.mul(o)).toInt()>>>0,c=u.toString(t);if(i=l,i.isZero())return c+a;for(;c.length<6;)c="0"+c;a=""+c+a}},R.getHighBits=function(){return this.high},R.getHighBitsUnsigned=function(){return this.high>>>0},R.getLowBits=function(){return this.low},R.getLowBitsUnsigned=function(){return this.low>>>0},R.getNumBitsAbs=function(){if(this.isNegative())return this.eq(Ft)?64:this.neg().getNumBitsAbs();for(var t=this.high!=0?this.high:this.low,e=31;e>0&&!(t&1<<e);e--);return this.high!=0?e+33:e+1},R.isZero=function(){return this.high===0&&this.low===0},R.eqz=R.isZero,R.isNegative=function(){return!this.unsigned&&this.high<0},R.isPositive=function(){return this.unsigned||this.high>=0},R.isOdd=function(){return(this.low&1)===1},R.isEven=function(){return(this.low&1)===0},R.equals=function(t){return Bt(t)||(t=ue(t)),this.unsigned!==t.unsigned&&this.high>>>31===1&&t.high>>>31===1?!1:this.high===t.high&&this.low===t.low},R.eq=R.equals,R.notEquals=function(t){return!this.eq(t)},R.neq=R.notEquals,R.ne=R.notEquals,R.lessThan=function(t){return this.comp(t)<0},R.lt=R.lessThan,R.lessThanOrEqual=function(t){return this.comp(t)<=0},R.lte=R.lessThanOrEqual,R.le=R.lessThanOrEqual,R.greaterThan=function(t){return this.comp(t)>0},R.gt=R.greaterThan,R.greaterThanOrEqual=function(t){return this.comp(t)>=0},R.gte=R.greaterThanOrEqual,R.ge=R.greaterThanOrEqual,R.compare=function(t){if(Bt(t)||(t=ue(t)),this.eq(t))return 0;var e=this.isNegative(),s=t.isNegative();return e&&!s?-1:!e&&s?1:this.unsigned?t.high>>>0>this.high>>>0||t.high===this.high&&t.low>>>0>this.low>>>0?-1:1:this.sub(t).isNegative()?-1:1},R.comp=R.compare,R.negate=function(){return!this.unsigned&&this.eq(Ft)?Ft:this.not().add(Tn)},R.neg=R.negate,R.add=function(t){Bt(t)||(t=ue(t));var e=this.high>>>16,s=this.high&65535,r=this.low>>>16,o=this.low&65535,i=t.high>>>16,a=t.high&65535,l=t.low>>>16,u=t.low&65535,c=0,h=0,f=0,d=0;return d+=o+u,f+=d>>>16,d&=65535,f+=r+l,h+=f>>>16,f&=65535,h+=s+a,c+=h>>>16,h&=65535,c+=e+i,c&=65535,at(f<<16|d,c<<16|h,this.unsigned)},R.subtract=function(t){return Bt(t)||(t=ue(t)),this.add(t.neg())},R.sub=R.subtract,R.multiply=function(t){if(this.isZero())return ee;if(Bt(t)||(t=ue(t)),Qt){var e=Qt.mul(this.low,this.high,t.low,t.high);return at(e,Qt.get_high(),this.unsigned)}if(t.isZero())return ee;if(this.eq(Ft))return t.isOdd()?Ft:ee;if(t.eq(Ft))return this.isOdd()?Ft:ee;if(this.isNegative())return t.isNegative()?this.neg().mul(t.neg()):this.neg().mul(t).neg();if(t.isNegative())return this.mul(t.neg()).neg();if(this.lt(qa)&&t.lt(qa))return te(this.toNumber()*t.toNumber(),this.unsigned);var s=this.high>>>16,r=this.high&65535,o=this.low>>>16,i=this.low&65535,a=t.high>>>16,l=t.high&65535,u=t.low>>>16,c=t.low&65535,h=0,f=0,d=0,p=0;return p+=i*c,d+=p>>>16,p&=65535,d+=o*c,f+=d>>>16,d&=65535,d+=i*u,f+=d>>>16,d&=65535,f+=r*c,h+=f>>>16,f&=65535,f+=o*u,h+=f>>>16,f&=65535,f+=i*l,h+=f>>>16,f&=65535,h+=s*c+r*u+o*l+i*a,h&=65535,at(d<<16|p,h<<16|f,this.unsigned)},R.mul=R.multiply,R.divide=function(t){if(Bt(t)||(t=ue(t)),t.isZero())throw Error("division by zero");if(Qt){if(!this.unsigned&&this.high===-2147483648&&t.low===-1&&t.high===-1)return this;var e=(this.unsigned?Qt.div_u:Qt.div_s)(this.low,this.high,t.low,t.high);return at(e,Qt.get_high(),this.unsigned)}if(this.isZero())return this.unsigned?Qe:ee;var s,r,o;if(this.unsigned){if(t.unsigned||(t=t.toUnsigned()),t.gt(this))return Qe;if(t.gt(this.shru(1)))return ja;o=Qe}else{if(this.eq(Ft)){if(t.eq(Tn)||t.eq(xo))return Ft;if(t.eq(Ft))return Tn;var i=this.shr(1);return s=i.div(t).shl(1),s.eq(ee)?t.isNegative()?Tn:xo:(r=this.sub(t.mul(s)),o=s.add(r.div(t)),o)}else if(t.eq(Ft))return this.unsigned?Qe:ee;if(this.isNegative())return t.isNegative()?this.neg().div(t.neg()):this.neg().div(t).neg();if(t.isNegative())return this.div(t.neg()).neg();o=ee}for(r=this;r.gte(t);){s=Math.max(1,Math.floor(r.toNumber()/t.toNumber()));for(var a=Math.ceil(Math.log(s)/Math.LN2),l=a<=48?1:Vs(2,a-48),u=te(s),c=u.mul(t);c.isNegative()||c.gt(r);)s-=l,u=te(s,this.unsigned),c=u.mul(t);u.isZero()&&(u=Tn),o=o.add(u),r=r.sub(c)}return o},R.div=R.divide,R.modulo=function(t){if(Bt(t)||(t=ue(t)),Qt){var e=(this.unsigned?Qt.rem_u:Qt.rem_s)(this.low,this.high,t.low,t.high);return at(e,Qt.get_high(),this.unsigned)}return this.sub(this.div(t).mul(t))},R.mod=R.modulo,R.rem=R.modulo,R.not=function(){return at(~this.low,~this.high,this.unsigned)},R.and=function(t){return Bt(t)||(t=ue(t)),at(this.low&t.low,this.high&t.high,this.unsigned)},R.or=function(t){return Bt(t)||(t=ue(t)),at(this.low|t.low,this.high|t.high,this.unsigned)},R.xor=function(t){return Bt(t)||(t=ue(t)),at(this.low^t.low,this.high^t.high,this.unsigned)},R.shiftLeft=function(t){return Bt(t)&&(t=t.toInt()),(t&=63)===0?this:t<32?at(this.low<<t,this.high<<t|this.low>>>32-t,this.unsigned):at(0,this.low<<t-32,this.unsigned)},R.shl=R.shiftLeft,R.shiftRight=function(t){return Bt(t)&&(t=t.toInt()),(t&=63)===0?this:t<32?at(this.low>>>t|this.high<<32-t,this.high>>t,this.unsigned):at(this.high>>t-32,this.high>=0?0:-1,this.unsigned)},R.shr=R.shiftRight,R.shiftRightUnsigned=function(t){if(Bt(t)&&(t=t.toInt()),t&=63,t===0)return this;var e=this.high;if(t<32){var s=this.low;return at(s>>>t|e<<32-t,e>>>t,this.unsigned)}else return t===32?at(e,0,this.unsigned):at(e>>>t-32,0,this.unsigned)},R.shru=R.shiftRightUnsigned,R.shr_u=R.shiftRightUnsigned,R.toSigned=function(){return this.unsigned?at(this.low,this.high,!1):this},R.toUnsigned=function(){return this.unsigned?this:at(this.low,this.high,!0)},R.toBytes=function(t){return t?this.toBytesLE():this.toBytesBE()},R.toBytesLE=function(){var t=this.high,e=this.low;return[e&255,e>>>8&255,e>>>16&255,e>>>24,t&255,t>>>8&255,t>>>16&255,t>>>24]},R.toBytesBE=function(){var t=this.high,e=this.low;return[t>>>24,t>>>16&255,t>>>8&255,t&255,e>>>24,e>>>16&255,e>>>8&255,e&255]},it.fromBytes=function(t,e,s){return s?it.fromBytesLE(t,e):it.fromBytesBE(t,e)},it.fromBytesLE=function(t,e){return new it(t[0]|t[1]<<8|t[2]<<16|t[3]<<24,t[4]|t[5]<<8|t[6]<<16|t[7]<<24,e)},it.fromBytesBE=function(t,e){return new it(t[4]<<24|t[5]<<16|t[6]<<8|t[7],t[0]<<24|t[1]<<16|t[2]<<8|t[3],e)};const Ya=up(Fa),fp=qt({__proto__:null,default:Ya},[Fa]);/**
+ */function Ba(n){return n instanceof Float32Array||n instanceof Int32Array||n instanceof Uint8Array||n instanceof Uint8ClampedArray}var Ze=typeof globalThis<"u"?globalThis:typeof window<"u"?window:typeof global<"u"?global:typeof self<"u"?self:{};function up(n){return n&&n.__esModule&&Object.prototype.hasOwnProperty.call(n,"default")?n.default:n}function cp(n){if(n.__esModule)return n;var t=n.default;if(typeof t=="function"){var e=function s(){return this instanceof s?Reflect.construct(t,arguments,this.constructor):t.apply(this,arguments)};e.prototype=t.prototype}else e={};return Object.defineProperty(e,"__esModule",{value:!0}),Object.keys(n).forEach(function(s){var r=Object.getOwnPropertyDescriptor(n,s);Object.defineProperty(e,s,r.get?r:{enumerable:!0,get:function(){return n[s]}})}),e}var Fa=it,Qt=null;try{Qt=new WebAssembly.Instance(new WebAssembly.Module(new Uint8Array([0,97,115,109,1,0,0,0,1,13,2,96,0,1,127,96,4,127,127,127,127,1,127,3,7,6,0,1,1,1,1,1,6,6,1,127,1,65,0,11,7,50,6,3,109,117,108,0,1,5,100,105,118,95,115,0,2,5,100,105,118,95,117,0,3,5,114,101,109,95,115,0,4,5,114,101,109,95,117,0,5,8,103,101,116,95,104,105,103,104,0,0,10,191,1,6,4,0,35,0,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,126,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,127,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,128,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,129,34,4,66,32,135,167,36,0,32,4,167,11,36,1,1,126,32,0,173,32,1,173,66,32,134,132,32,2,173,32,3,173,66,32,134,132,130,34,4,66,32,135,167,36,0,32,4,167,11])),{}).exports}catch{}function it(n,t,e){this.low=n|0,this.high=t|0,this.unsigned=!!e}it.prototype.__isLong__,Object.defineProperty(it.prototype,"__isLong__",{value:!0});function Bt(n){return(n&&n.__isLong__)===!0}it.isLong=Bt;var za={},Ua={};function Qe(n,t){var e,s,r;return t?(n>>>=0,(r=0<=n&&n<256)&&(s=Ua[n],s)?s:(e=at(n,(n|0)<0?-1:0,!0),r&&(Ua[n]=e),e)):(n|=0,(r=-128<=n&&n<128)&&(s=za[n],s)?s:(e=at(n,n<0?-1:0,!1),r&&(za[n]=e),e))}it.fromInt=Qe;function te(n,t){if(isNaN(n))return t?tn:ee;if(t){if(n<0)return tn;if(n>=Ga)return Ka}else{if(n<=-Va)return Ft;if(n+1>=Va)return Ha}return n<0?te(-n,t).neg():at(n%_n|0,n/_n|0,t)}it.fromNumber=te;function at(n,t,e){return new it(n,t,e)}it.fromBits=at;var qs=Math.pow;function wo(n,t,e){if(n.length===0)throw Error("empty string");if(n==="NaN"||n==="Infinity"||n==="+Infinity"||n==="-Infinity")return ee;if(typeof t=="number"?(e=t,t=!1):t=!!t,e=e||10,e<2||36<e)throw RangeError("radix");var s;if((s=n.indexOf("-"))>0)throw Error("interior hyphen");if(s===0)return wo(n.substring(1),t,e).neg();for(var r=te(qs(e,8)),o=ee,i=0;i<n.length;i+=8){var a=Math.min(8,n.length-i),l=parseInt(n.substring(i,i+a),e);if(a<8){var u=te(qs(e,a));o=o.mul(u).add(te(l))}else o=o.mul(r),o=o.add(te(l))}return o.unsigned=t,o}it.fromString=wo;function ue(n,t){return typeof n=="number"?te(n,t):typeof n=="string"?wo(n,t):at(n.low,n.high,typeof t=="boolean"?t:n.unsigned)}it.fromValue=ue;var Wa=65536,hp=1<<24,_n=Wa*Wa,Ga=_n*_n,Va=Ga/2,qa=Qe(hp),ee=Qe(0);it.ZERO=ee;var tn=Qe(0,!0);it.UZERO=tn;var Tn=Qe(1);it.ONE=Tn;var ja=Qe(1,!0);it.UONE=ja;var xo=Qe(-1);it.NEG_ONE=xo;var Ha=at(-1,2147483647,!1);it.MAX_VALUE=Ha;var Ka=at(-1,-1,!0);it.MAX_UNSIGNED_VALUE=Ka;var Ft=at(0,-2147483648,!1);it.MIN_VALUE=Ft;var R=it.prototype;R.toInt=function(){return this.unsigned?this.low>>>0:this.low},R.toNumber=function(){return this.unsigned?(this.high>>>0)*_n+(this.low>>>0):this.high*_n+(this.low>>>0)},R.toString=function(t){if(t=t||10,t<2||36<t)throw RangeError("radix");if(this.isZero())return"0";if(this.isNegative())if(this.eq(Ft)){var e=te(t),s=this.div(e),r=s.mul(e).sub(this);return s.toString(t)+r.toInt().toString(t)}else return"-"+this.neg().toString(t);for(var o=te(qs(t,6),this.unsigned),i=this,a="";;){var l=i.div(o),u=i.sub(l.mul(o)).toInt()>>>0,c=u.toString(t);if(i=l,i.isZero())return c+a;for(;c.length<6;)c="0"+c;a=""+c+a}},R.getHighBits=function(){return this.high},R.getHighBitsUnsigned=function(){return this.high>>>0},R.getLowBits=function(){return this.low},R.getLowBitsUnsigned=function(){return this.low>>>0},R.getNumBitsAbs=function(){if(this.isNegative())return this.eq(Ft)?64:this.neg().getNumBitsAbs();for(var t=this.high!=0?this.high:this.low,e=31;e>0&&!(t&1<<e);e--);return this.high!=0?e+33:e+1},R.isZero=function(){return this.high===0&&this.low===0},R.eqz=R.isZero,R.isNegative=function(){return!this.unsigned&&this.high<0},R.isPositive=function(){return this.unsigned||this.high>=0},R.isOdd=function(){return(this.low&1)===1},R.isEven=function(){return(this.low&1)===0},R.equals=function(t){return Bt(t)||(t=ue(t)),this.unsigned!==t.unsigned&&this.high>>>31===1&&t.high>>>31===1?!1:this.high===t.high&&this.low===t.low},R.eq=R.equals,R.notEquals=function(t){return!this.eq(t)},R.neq=R.notEquals,R.ne=R.notEquals,R.lessThan=function(t){return this.comp(t)<0},R.lt=R.lessThan,R.lessThanOrEqual=function(t){return this.comp(t)<=0},R.lte=R.lessThanOrEqual,R.le=R.lessThanOrEqual,R.greaterThan=function(t){return this.comp(t)>0},R.gt=R.greaterThan,R.greaterThanOrEqual=function(t){return this.comp(t)>=0},R.gte=R.greaterThanOrEqual,R.ge=R.greaterThanOrEqual,R.compare=function(t){if(Bt(t)||(t=ue(t)),this.eq(t))return 0;var e=this.isNegative(),s=t.isNegative();return e&&!s?-1:!e&&s?1:this.unsigned?t.high>>>0>this.high>>>0||t.high===this.high&&t.low>>>0>this.low>>>0?-1:1:this.sub(t).isNegative()?-1:1},R.comp=R.compare,R.negate=function(){return!this.unsigned&&this.eq(Ft)?Ft:this.not().add(Tn)},R.neg=R.negate,R.add=function(t){Bt(t)||(t=ue(t));var e=this.high>>>16,s=this.high&65535,r=this.low>>>16,o=this.low&65535,i=t.high>>>16,a=t.high&65535,l=t.low>>>16,u=t.low&65535,c=0,h=0,f=0,d=0;return d+=o+u,f+=d>>>16,d&=65535,f+=r+l,h+=f>>>16,f&=65535,h+=s+a,c+=h>>>16,h&=65535,c+=e+i,c&=65535,at(f<<16|d,c<<16|h,this.unsigned)},R.subtract=function(t){return Bt(t)||(t=ue(t)),this.add(t.neg())},R.sub=R.subtract,R.multiply=function(t){if(this.isZero())return ee;if(Bt(t)||(t=ue(t)),Qt){var e=Qt.mul(this.low,this.high,t.low,t.high);return at(e,Qt.get_high(),this.unsigned)}if(t.isZero())return ee;if(this.eq(Ft))return t.isOdd()?Ft:ee;if(t.eq(Ft))return this.isOdd()?Ft:ee;if(this.isNegative())return t.isNegative()?this.neg().mul(t.neg()):this.neg().mul(t).neg();if(t.isNegative())return this.mul(t.neg()).neg();if(this.lt(qa)&&t.lt(qa))return te(this.toNumber()*t.toNumber(),this.unsigned);var s=this.high>>>16,r=this.high&65535,o=this.low>>>16,i=this.low&65535,a=t.high>>>16,l=t.high&65535,u=t.low>>>16,c=t.low&65535,h=0,f=0,d=0,p=0;return p+=i*c,d+=p>>>16,p&=65535,d+=o*c,f+=d>>>16,d&=65535,d+=i*u,f+=d>>>16,d&=65535,f+=r*c,h+=f>>>16,f&=65535,f+=o*u,h+=f>>>16,f&=65535,f+=i*l,h+=f>>>16,f&=65535,h+=s*c+r*u+o*l+i*a,h&=65535,at(d<<16|p,h<<16|f,this.unsigned)},R.mul=R.multiply,R.divide=function(t){if(Bt(t)||(t=ue(t)),t.isZero())throw Error("division by zero");if(Qt){if(!this.unsigned&&this.high===-2147483648&&t.low===-1&&t.high===-1)return this;var e=(this.unsigned?Qt.div_u:Qt.div_s)(this.low,this.high,t.low,t.high);return at(e,Qt.get_high(),this.unsigned)}if(this.isZero())return this.unsigned?tn:ee;var s,r,o;if(this.unsigned){if(t.unsigned||(t=t.toUnsigned()),t.gt(this))return tn;if(t.gt(this.shru(1)))return ja;o=tn}else{if(this.eq(Ft)){if(t.eq(Tn)||t.eq(xo))return Ft;if(t.eq(Ft))return Tn;var i=this.shr(1);return s=i.div(t).shl(1),s.eq(ee)?t.isNegative()?Tn:xo:(r=this.sub(t.mul(s)),o=s.add(r.div(t)),o)}else if(t.eq(Ft))return this.unsigned?tn:ee;if(this.isNegative())return t.isNegative()?this.neg().div(t.neg()):this.neg().div(t).neg();if(t.isNegative())return this.div(t.neg()).neg();o=ee}for(r=this;r.gte(t);){s=Math.max(1,Math.floor(r.toNumber()/t.toNumber()));for(var a=Math.ceil(Math.log(s)/Math.LN2),l=a<=48?1:qs(2,a-48),u=te(s),c=u.mul(t);c.isNegative()||c.gt(r);)s-=l,u=te(s,this.unsigned),c=u.mul(t);u.isZero()&&(u=Tn),o=o.add(u),r=r.sub(c)}return o},R.div=R.divide,R.modulo=function(t){if(Bt(t)||(t=ue(t)),Qt){var e=(this.unsigned?Qt.rem_u:Qt.rem_s)(this.low,this.high,t.low,t.high);return at(e,Qt.get_high(),this.unsigned)}return this.sub(this.div(t).mul(t))},R.mod=R.modulo,R.rem=R.modulo,R.not=function(){return at(~this.low,~this.high,this.unsigned)},R.and=function(t){return Bt(t)||(t=ue(t)),at(this.low&t.low,this.high&t.high,this.unsigned)},R.or=function(t){return Bt(t)||(t=ue(t)),at(this.low|t.low,this.high|t.high,this.unsigned)},R.xor=function(t){return Bt(t)||(t=ue(t)),at(this.low^t.low,this.high^t.high,this.unsigned)},R.shiftLeft=function(t){return Bt(t)&&(t=t.toInt()),(t&=63)===0?this:t<32?at(this.low<<t,this.high<<t|this.low>>>32-t,this.unsigned):at(0,this.low<<t-32,this.unsigned)},R.shl=R.shiftLeft,R.shiftRight=function(t){return Bt(t)&&(t=t.toInt()),(t&=63)===0?this:t<32?at(this.low>>>t|this.high<<32-t,this.high>>t,this.unsigned):at(this.high>>t-32,this.high>=0?0:-1,this.unsigned)},R.shr=R.shiftRight,R.shiftRightUnsigned=function(t){if(Bt(t)&&(t=t.toInt()),t&=63,t===0)return this;var e=this.high;if(t<32){var s=this.low;return at(s>>>t|e<<32-t,e>>>t,this.unsigned)}else return t===32?at(e,0,this.unsigned):at(e>>>t-32,0,this.unsigned)},R.shru=R.shiftRightUnsigned,R.shr_u=R.shiftRightUnsigned,R.toSigned=function(){return this.unsigned?at(this.low,this.high,!1):this},R.toUnsigned=function(){return this.unsigned?this:at(this.low,this.high,!0)},R.toBytes=function(t){return t?this.toBytesLE():this.toBytesBE()},R.toBytesLE=function(){var t=this.high,e=this.low;return[e&255,e>>>8&255,e>>>16&255,e>>>24,t&255,t>>>8&255,t>>>16&255,t>>>24]},R.toBytesBE=function(){var t=this.high,e=this.low;return[t>>>24,t>>>16&255,t>>>8&255,t&255,e>>>24,e>>>16&255,e>>>8&255,e&255]},it.fromBytes=function(t,e,s){return s?it.fromBytesLE(t,e):it.fromBytesBE(t,e)},it.fromBytesLE=function(t,e){return new it(t[0]|t[1]<<8|t[2]<<16|t[3]<<24,t[4]|t[5]<<8|t[6]<<16|t[7]<<24,e)},it.fromBytesBE=function(t,e){return new it(t[4]<<24|t[5]<<16|t[6]<<8|t[7],t[0]<<24|t[1]<<16|t[2]<<8|t[3],e)};const Ya=up(Fa),fp=Ht({__proto__:null,default:Ya},[Fa]);/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -118,7 +118,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const tn=Ya||fp;function qs(n){return tn.fromString(n,!0,16)}const Xa=qs("c3a5c85c97cb3127"),en=qs("b492b66fbe98f273"),kt=qs("9ae16a3b2f90404f");function So(n){return n.xor(n.shru(47))}function Ja(n,t,e){const s=n.slice(t,t+e);return tn.fromBytes(Array.from(s),!0,!0)}function rt(n,t){return Ja(n,t,8)}function Za(n,t){return Ja(n,t,4)}function yt(n,t){return t===0?n:n.shru(t).or(n.shl(64-t))}function Ue(n,t,e=qs("9ddfea08eb382d69")){let s=n.xor(t).mul(e);s=s.xor(s.shru(47));let r=t.xor(s).mul(e);return r=r.xor(r.shru(47)),r=r.mul(e),r}function dp(n,t,e,s,r,o){r=r.add(n),o=yt(o.add(r).add(s),21);const i=r;return r=r.add(t),r=r.add(e),o=o.add(yt(r,44)),[r.add(s),o.add(i)]}function js(n,t,e,s){return dp(rt(n,t),rt(n,t+8),rt(n,t+16),rt(n,t+24),e,s)}function pp(n,t=n.length){if(t>=8){const e=kt.add(t*2),s=rt(n,0).add(kt),r=rt(n,t-8),o=yt(r,37).mul(e).add(s),i=yt(s,25).add(r).mul(e);return Ue(o,i,e)}if(t>=4){const e=kt.add(t*2),s=Za(n,0);return Ue(s.shl(3).add(t),Za(n,t-4),e)}if(t>0){const e=n[0],s=n[t>>1],r=n[t-1],o=e+(s<<8),i=t+(r<<2);return So(kt.mul(o).xor(Xa.mul(i))).mul(kt)}return kt}function mp(n,t=n.length){const e=kt.add(t*2),s=rt(n,0).mul(en),r=rt(n,8),o=rt(n,t-8).mul(e),i=rt(n,t-16).mul(kt);return Ue(yt(s.add(r),43).add(yt(o,30)).add(i),s.add(yt(r.add(kt),18)).add(o),e)}function gp(n,t=n.length){const e=kt.add(t*2),s=rt(n,0).mul(kt),r=rt(n,8),o=rt(n,t-8).mul(e),i=rt(n,t-16).mul(kt),a=yt(s.add(r),43).add(yt(o,30)).add(i),l=Ue(a,s.add(yt(r.add(kt),18)).add(o),e),u=rt(n,16).mul(e),c=rt(n,24),h=a.add(rt(n,t-32)).mul(e),f=l.add(rt(n,t-24)).mul(e);return Ue(yt(u.add(c),43).add(yt(h,30)).add(f),u.add(yt(c.add(s),18)).add(h),e)}function bp(n,t=n.length){const e=tn.fromNumber(81,!0);if(t<=32)return t<=16?pp(n,t):mp(n,t);if(t<=64)return gp(n,t);let s=e,r=e.mul(en).add(113),o=So(r.mul(kt).add(113)).mul(kt),i=[tn.UZERO,tn.UZERO],a=[tn.UZERO,tn.UZERO];s=s.mul(kt).add(rt(n,0));let l=0;const u=(t-1>>6)*64,c=u+(t-1&63)-63;do s=yt(s.add(r).add(i[0]).add(rt(n,l+8)),37).mul(en),r=yt(r.add(i[1]).add(rt(n,l+48)),42).mul(en),s=s.xor(a[1]),r=r.add(i[0]).add(rt(n,l+40)),o=yt(o.add(a[0]),33).mul(en),i=js(n,l,i[1].mul(en),s.add(a[0])),a=js(n,l+32,o.add(a[1]),r.add(rt(n,l+16))),[o,s]=[s,o],l+=64;while(l!==u);const h=en.add(o.and(255).shl(1));return l=c,a[0]=a[0].add(t-1&63),i[0]=i[0].add(a[0]),a[0]=a[0].add(i[0]),s=yt(s.add(r).add(i[0]).add(rt(n,l+8)),37).mul(h),r=yt(r.add(i[1]).add(rt(n,l+48)),42).mul(h),s=s.xor(a[1].mul(9)),r=r.add(i[0].mul(9).add(rt(n,l+40))),o=yt(o.add(a[0]),33).mul(h),i=js(n,l,i[1].mul(h),s.add(a[0])),a=js(n,l+32,o.add(a[1]),r.add(rt(n,l+16))),[o,s]=[s,o],Ue(Ue(i[0],a[0],h).add(So(r).mul(Xa)).add(o),Ue(i[1],a[1],h).add(s),h)}/**
+ */const en=Ya||fp;function js(n){return en.fromString(n,!0,16)}const Xa=js("c3a5c85c97cb3127"),nn=js("b492b66fbe98f273"),kt=js("9ae16a3b2f90404f");function So(n){return n.xor(n.shru(47))}function Ja(n,t,e){const s=n.slice(t,t+e);return en.fromBytes(Array.from(s),!0,!0)}function rt(n,t){return Ja(n,t,8)}function Za(n,t){return Ja(n,t,4)}function yt(n,t){return t===0?n:n.shru(t).or(n.shl(64-t))}function Ue(n,t,e=js("9ddfea08eb382d69")){let s=n.xor(t).mul(e);s=s.xor(s.shru(47));let r=t.xor(s).mul(e);return r=r.xor(r.shru(47)),r=r.mul(e),r}function dp(n,t,e,s,r,o){r=r.add(n),o=yt(o.add(r).add(s),21);const i=r;return r=r.add(t),r=r.add(e),o=o.add(yt(r,44)),[r.add(s),o.add(i)]}function Hs(n,t,e,s){return dp(rt(n,t),rt(n,t+8),rt(n,t+16),rt(n,t+24),e,s)}function pp(n,t=n.length){if(t>=8){const e=kt.add(t*2),s=rt(n,0).add(kt),r=rt(n,t-8),o=yt(r,37).mul(e).add(s),i=yt(s,25).add(r).mul(e);return Ue(o,i,e)}if(t>=4){const e=kt.add(t*2),s=Za(n,0);return Ue(s.shl(3).add(t),Za(n,t-4),e)}if(t>0){const e=n[0],s=n[t>>1],r=n[t-1],o=e+(s<<8),i=t+(r<<2);return So(kt.mul(o).xor(Xa.mul(i))).mul(kt)}return kt}function mp(n,t=n.length){const e=kt.add(t*2),s=rt(n,0).mul(nn),r=rt(n,8),o=rt(n,t-8).mul(e),i=rt(n,t-16).mul(kt);return Ue(yt(s.add(r),43).add(yt(o,30)).add(i),s.add(yt(r.add(kt),18)).add(o),e)}function gp(n,t=n.length){const e=kt.add(t*2),s=rt(n,0).mul(kt),r=rt(n,8),o=rt(n,t-8).mul(e),i=rt(n,t-16).mul(kt),a=yt(s.add(r),43).add(yt(o,30)).add(i),l=Ue(a,s.add(yt(r.add(kt),18)).add(o),e),u=rt(n,16).mul(e),c=rt(n,24),h=a.add(rt(n,t-32)).mul(e),f=l.add(rt(n,t-24)).mul(e);return Ue(yt(u.add(c),43).add(yt(h,30)).add(f),u.add(yt(c.add(s),18)).add(h),e)}function bp(n,t=n.length){const e=en.fromNumber(81,!0);if(t<=32)return t<=16?pp(n,t):mp(n,t);if(t<=64)return gp(n,t);let s=e,r=e.mul(nn).add(113),o=So(r.mul(kt).add(113)).mul(kt),i=[en.UZERO,en.UZERO],a=[en.UZERO,en.UZERO];s=s.mul(kt).add(rt(n,0));let l=0;const u=(t-1>>6)*64,c=u+(t-1&63)-63;do s=yt(s.add(r).add(i[0]).add(rt(n,l+8)),37).mul(nn),r=yt(r.add(i[1]).add(rt(n,l+48)),42).mul(nn),s=s.xor(a[1]),r=r.add(i[0]).add(rt(n,l+40)),o=yt(o.add(a[0]),33).mul(nn),i=Hs(n,l,i[1].mul(nn),s.add(a[0])),a=Hs(n,l+32,o.add(a[1]),r.add(rt(n,l+16))),[o,s]=[s,o],l+=64;while(l!==u);const h=nn.add(o.and(255).shl(1));return l=c,a[0]=a[0].add(t-1&63),i[0]=i[0].add(a[0]),a[0]=a[0].add(i[0]),s=yt(s.add(r).add(i[0]).add(rt(n,l+8)),37).mul(h),r=yt(r.add(i[1]).add(rt(n,l+48)),42).mul(h),s=s.xor(a[1].mul(9)),r=r.add(i[0].mul(9).add(rt(n,l+40))),o=yt(o.add(a[0]),33).mul(h),i=Hs(n,l,i[1].mul(h),s.add(a[0])),a=Hs(n,l+32,o.add(a[1]),r.add(rt(n,l+16))),[o,s]=[s,o],Ue(Ue(i[0],a[0],h).add(So(r).mul(Xa)).add(o),Ue(i[1],a[1],h).add(s),h)}/**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -133,7 +133,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function yp(n,t){return t==="string"?nn(n):Hs([n],t)}function wp(n,t){return n instanceof Float32Array&&t==="float32"||n instanceof Int32Array&&t==="int32"||n instanceof Uint8Array&&t==="bool"}function Hs(n,t){if(t==="string")throw new Error("Cannot convert a string[] to a TypedArray");if(Array.isArray(n)&&(n=sn(n)),W().getBool("DEBUG")&&_f(n,t),wp(n,t))return n;if(t==null||t==="float32"||t==="complex64")return new Float32Array(n);if(t==="int32")return new Int32Array(n);if(t==="bool"){const e=new Uint8Array(n.length);for(let s=0;s<e.length;++s)Math.round(n[s])!==0&&(e[s]=1);return e}else throw new Error(`Unknown data type ${t}`)}function Nn(){return W().platform.now()}function nn(n,t="utf-8"){return t=t||"utf-8",W().platform.encode(n,t)}function Ks(n,t="utf-8"){return t=t||"utf-8",W().platform.decode(n,t)}function ne(n){return W().platform.isTypedArray!=null?W().platform.isTypedArray(n):Ba(n)}function sn(n,t=[],e=!1){if(t==null&&(t=[]),typeof n=="boolean"||typeof n=="number"||typeof n=="string"||fo(n)||n==null||ne(n)&&e)t.push(n);else if(Array.isArray(n)||ne(n))for(let s=0;s<n.length;++s)sn(n[s],t,e);else{let s=-1;for(const r of Object.keys(n))/^([1-9]+[0-9]*|0)$/.test(r)&&(s=Math.max(s,Number(r)));for(let r=0;r<=s;r++)sn(n[r],t,e)}return t}/**
+ */function yp(n,t){return t==="string"?sn(n):Ks([n],t)}function wp(n,t){return n instanceof Float32Array&&t==="float32"||n instanceof Int32Array&&t==="int32"||n instanceof Uint8Array&&t==="bool"}function Ks(n,t){if(t==="string")throw new Error("Cannot convert a string[] to a TypedArray");if(Array.isArray(n)&&(n=rn(n)),W().getBool("DEBUG")&&_f(n,t),wp(n,t))return n;if(t==null||t==="float32"||t==="complex64")return new Float32Array(n);if(t==="int32")return new Int32Array(n);if(t==="bool"){const e=new Uint8Array(n.length);for(let s=0;s<e.length;++s)Math.round(n[s])!==0&&(e[s]=1);return e}else throw new Error(`Unknown data type ${t}`)}function Nn(){return W().platform.now()}function sn(n,t="utf-8"){return t=t||"utf-8",W().platform.encode(n,t)}function Ys(n,t="utf-8"){return t=t||"utf-8",W().platform.decode(n,t)}function ne(n){return W().platform.isTypedArray!=null?W().platform.isTypedArray(n):Ba(n)}function rn(n,t=[],e=!1){if(t==null&&(t=[]),typeof n=="boolean"||typeof n=="number"||typeof n=="string"||fo(n)||n==null||ne(n)&&e)t.push(n);else if(Array.isArray(n)||ne(n))for(let s=0;s<n.length;++s)rn(n[s],t,e);else{let s=-1;for(const r of Object.keys(n))/^([1-9]+[0-9]*|0)$/.test(r)&&(s=Math.max(s,Number(r)));for(let r=0;r<=s;r++)rn(n[r],t,e)}return t}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -148,7 +148,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class xp{constructor(t,e){this.backendTimer=t,this.logger=e,e==null&&(this.logger=new $p)}profileKernel(t,e,s){let r;const o=()=>{r=s()};let i;const a=Nn();if(this.backendTimer.timerAvailable())i=this.backendTimer.time(o);else{o();for(const u of r)u.dataSync();i=Promise.resolve({kernelMs:Nn()-a})}if(W().getBool("CHECK_COMPUTATION_FOR_ERRORS"))for(let u=0;u<r.length;u++){const c=r[u];c.data().then(h=>{Sp(h,c.dtype,t)})}return{kernelName:t,outputs:r,inputs:e,timeMs:i.then(u=>u.kernelMs),extraInfo:i.then(u=>u.getExtraProfileInfo!=null?u.getExtraProfileInfo():"")}}logKernelProfile(t){const{kernelName:e,outputs:s,timeMs:r,inputs:o,extraInfo:i}=t;s.forEach(a=>{Promise.all([a.data(),r,i]).then(l=>{this.logger.logKernelProfile(e,a,l[0],l[1],o,l[2])})})}}function Sp(n,t,e){if(t!=="float32")return!1;for(let s=0;s<n.length;s++){const r=n[s];if(isNaN(r)||!isFinite(r))return console.warn(`Found ${r} in the result of '${e}'`),!0}return!1}class $p{logKernelProfile(t,e,s,r,o,i){const a=typeof r=="number"?Us(`${r}ms`,9):r.error,l=Us(t,25),u=e.rank,c=e.size,h=Us(e.shape.toString(),14);let f="";for(const d in o){const p=o[d];if(p!=null){const g=p.shape||e.shape,m=g.length;f+=`${d}: ${m}D ${m>0?g:""} `}}console.log(`%c${l}	%c${a}	%c${u}D ${h}	%c${c}	%c${f}	%c${i}`,"font-weight:bold","color:red","color:blue","color: orange","color: green","color: steelblue")}}/**
+ */class xp{constructor(t,e){this.backendTimer=t,this.logger=e,e==null&&(this.logger=new $p)}profileKernel(t,e,s){let r;const o=()=>{r=s()};let i;const a=Nn();if(this.backendTimer.timerAvailable())i=this.backendTimer.time(o);else{o();for(const u of r)u.dataSync();i=Promise.resolve({kernelMs:Nn()-a})}if(W().getBool("CHECK_COMPUTATION_FOR_ERRORS"))for(let u=0;u<r.length;u++){const c=r[u];c.data().then(h=>{Sp(h,c.dtype,t)})}return{kernelName:t,outputs:r,inputs:e,timeMs:i.then(u=>u.kernelMs),extraInfo:i.then(u=>u.getExtraProfileInfo!=null?u.getExtraProfileInfo():"")}}logKernelProfile(t){const{kernelName:e,outputs:s,timeMs:r,inputs:o,extraInfo:i}=t;s.forEach(a=>{Promise.all([a.data(),r,i]).then(l=>{this.logger.logKernelProfile(e,a,l[0],l[1],o,l[2])})})}}function Sp(n,t,e){if(t!=="float32")return!1;for(let s=0;s<n.length;s++){const r=n[s];if(isNaN(r)||!isFinite(r))return console.warn(`Found ${r} in the result of '${e}'`),!0}return!1}class $p{logKernelProfile(t,e,s,r,o,i){const a=typeof r=="number"?Ws(`${r}ms`,9):r.error,l=Ws(t,25),u=e.rank,c=e.size,h=Ws(e.shape.toString(),14);let f="";for(const d in o){const p=o[d];if(p!=null){const g=p.shape||e.shape,m=g.length;f+=`${d}: ${m}D ${m>0?g:""} `}}console.log(`%c${l}	%c${a}	%c${u}D ${h}	%c${c}	%c${f}	%c${i}`,"font-weight:bold","color:red","color:blue","color: orange","color: green","color: steelblue")}}/**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -178,11 +178,11 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const Qa=20,as=3,$o=7;function Ap(n,t,e,s){const r=jt(t),o=Ep(n,t,e,r),i=t.length,a=Ys(n,t,e,r,o),l=["Tensor"];return s&&(l.push(`  dtype: ${e}`),l.push(`  rank: ${i}`),l.push(`  shape: [${t}]`),l.push("  values:")),l.push(a.map(u=>"    "+u).join(`
+ */const Qa=20,ls=3,$o=7;function Ap(n,t,e,s){const r=Kt(t),o=Ep(n,t,e,r),i=t.length,a=Xs(n,t,e,r,o),l=["Tensor"];return s&&(l.push(`  dtype: ${e}`),l.push(`  rank: ${i}`),l.push(`  shape: [${t}]`),l.push("  values:")),l.push(a.map(u=>"    "+u).join(`
 `)),l.join(`
-`)}function Ep(n,t,e,s){const r=z(t),o=s[s.length-1],i=new Array(o).fill(0),a=t.length,l=e==="complex64"?us(n):n;if(a>1)for(let u=0;u<r/o;u++){const c=u*o;for(let h=0;h<o;h++)i[h]=Math.max(i[h],ls(l[c+h],0,e).length)}return i}function ls(n,t,e){let s;return Array.isArray(n)?s=`${parseFloat(n[0].toFixed($o))} + ${parseFloat(n[1].toFixed($o))}j`:Ws(n)?s=`'${n}'`:e==="bool"?s=tl(n):s=parseFloat(n.toFixed($o)).toString(),Us(s,t)}function tl(n){return n===0?"false":"true"}function Ys(n,t,e,s,r,o=!0){const i=e==="complex64"?2:1,a=t[0],l=t.length;if(l===0){if(e==="complex64"){const g=us(n);return[ls(g[0],0,e)]}return e==="bool"?[tl(n[0])]:[n[0].toString()]}if(l===1){if(a>Qa){const m=as*i;let b=Array.from(n.slice(0,m)),y=Array.from(n.slice((a-as)*i,a*i));return e==="complex64"&&(b=us(b),y=us(y)),["["+b.map((S,x)=>ls(S,r[x],e)).join(", ")+", ..., "+y.map((S,x)=>ls(S,r[a-as+x],e)).join(", ")+"]"]}return["["+(e==="complex64"?us(n):Array.from(n)).map((m,b)=>ls(m,r[b],e)).join(", ")+"]"]}const u=t.slice(1),c=s.slice(1),h=s[0]*i,f=[];if(a>Qa){for(let g=0;g<as;g++){const m=g*h,b=m+h;f.push(...Ys(n.slice(m,b),u,e,c,r,!1))}f.push("...");for(let g=a-as;g<a;g++){const m=g*h,b=m+h;f.push(...Ys(n.slice(m,b),u,e,c,r,g===a-1))}}else for(let g=0;g<a;g++){const m=g*h,b=m+h;f.push(...Ys(n.slice(m,b),u,e,c,r,g===a-1))}const d=l===2?",":"";f[0]="["+(a>0?f[0]+d:"");for(let g=1;g<f.length-1;g++)f[g]=" "+f[g]+d;let p=`,
+`)}function Ep(n,t,e,s){const r=z(t),o=s[s.length-1],i=new Array(o).fill(0),a=t.length,l=e==="complex64"?cs(n):n;if(a>1)for(let u=0;u<r/o;u++){const c=u*o;for(let h=0;h<o;h++)i[h]=Math.max(i[h],us(l[c+h],0,e).length)}return i}function us(n,t,e){let s;return Array.isArray(n)?s=`${parseFloat(n[0].toFixed($o))} + ${parseFloat(n[1].toFixed($o))}j`:Gs(n)?s=`'${n}'`:e==="bool"?s=tl(n):s=parseFloat(n.toFixed($o)).toString(),Ws(s,t)}function tl(n){return n===0?"false":"true"}function Xs(n,t,e,s,r,o=!0){const i=e==="complex64"?2:1,a=t[0],l=t.length;if(l===0){if(e==="complex64"){const g=cs(n);return[us(g[0],0,e)]}return e==="bool"?[tl(n[0])]:[n[0].toString()]}if(l===1){if(a>Qa){const m=ls*i;let b=Array.from(n.slice(0,m)),y=Array.from(n.slice((a-ls)*i,a*i));return e==="complex64"&&(b=cs(b),y=cs(y)),["["+b.map((S,x)=>us(S,r[x],e)).join(", ")+", ..., "+y.map((S,x)=>us(S,r[a-ls+x],e)).join(", ")+"]"]}return["["+(e==="complex64"?cs(n):Array.from(n)).map((m,b)=>us(m,r[b],e)).join(", ")+"]"]}const u=t.slice(1),c=s.slice(1),h=s[0]*i,f=[];if(a>Qa){for(let g=0;g<ls;g++){const m=g*h,b=m+h;f.push(...Xs(n.slice(m,b),u,e,c,r,!1))}f.push("...");for(let g=a-ls;g<a;g++){const m=g*h,b=m+h;f.push(...Xs(n.slice(m,b),u,e,c,r,g===a-1))}}else for(let g=0;g<a;g++){const m=g*h,b=m+h;f.push(...Xs(n.slice(m,b),u,e,c,r,g===a-1))}const d=l===2?",":"";f[0]="["+(a>0?f[0]+d:"");for(let g=1;g<f.length-1;g++)f[g]=" "+f[g]+d;let p=`,
 `;for(let g=2;g<l;g++)p+=`
-`;return f[f.length-1]=" "+f[f.length-1]+"]"+(o?"":p),f}function us(n){const t=[];for(let e=0;e<n.length;e+=2)t.push([n[e],n[e+1]]);return t}/**
+`;return f[f.length-1]=" "+f[f.length-1]+"]"+(o?"":p),f}function cs(n){const t=[];for(let e=0;e<n.length;e+=2)t.push([n[e],n[e+1]]);return t}/**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -197,7 +197,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class Xs{constructor(t,e,s){if(this.dtype=e,this.shape=t.slice(),this.size=z(t),s!=null){const r=s.length;w(r===this.size,()=>`Length of values '${r}' does not match the size inferred by the shape '${this.size}'.`)}if(e==="complex64")throw new Error("complex64 dtype TensorBuffers are not supported. Please create a TensorBuffer for the real and imaginary parts separately and call tf.complex(real, imag).");this.values=s||bt(e,this.size),this.strides=jt(t)}set(t,...e){e.length===0&&(e=[0]),w(e.length===this.rank,()=>`The number of provided coordinates (${e.length}) must match the rank (${this.rank})`);const s=this.locToIndex(e);this.values[s]=t}get(...t){t.length===0&&(t=[0]);let e=0;for(const r of t){if(r<0||r>=this.shape[e]){const o=`Requested out of range element at ${t}.   Buffer shape=${this.shape}`;throw new Error(o)}e++}let s=t[t.length-1];for(let r=0;r<t.length-1;++r)s+=this.strides[r]*t[r];return this.values[s]}locToIndex(t){if(this.rank===0)return 0;if(this.rank===1)return t[0];let e=t[t.length-1];for(let s=0;s<t.length-1;++s)e+=this.strides[s]*t[s];return e}indexToLoc(t){if(this.rank===0)return[];if(this.rank===1)return[t];const e=new Array(this.shape.length);for(let s=0;s<e.length-1;++s)e[s]=Math.floor(t/this.strides[s]),t-=e[s]*this.strides[s];return e[e.length-1]=t,e}get rank(){return this.shape.length}toTensor(){return ce().makeTensor(this.values,this.shape,this.dtype)}}let ce=null,Dn=null;function Cp(n){ce=n}function kp(n){Dn=n}class At{constructor(t,e,s,r){this.kept=!1,this.isDisposedInternal=!1,this.shape=t.slice(),this.dtype=e||"float32",this.size=z(t),this.strides=jt(t),this.dataId=s,this.id=r,this.rankType=this.rank<5?this.rank.toString():"higher"}get rank(){return this.shape.length}async buffer(){const t=await this.data();return Dn.buffer(this.shape,this.dtype,t)}bufferSync(){return Dn.buffer(this.shape,this.dtype,this.dataSync())}async array(){const t=await this.data();return xa(this.shape,t,this.dtype==="complex64")}arraySync(){return xa(this.shape,this.dataSync(),this.dtype==="complex64")}async data(){this.throwIfDisposed();const t=ce().read(this.dataId);if(this.dtype==="string"){const e=await t;try{return e.map(s=>Ks(s))}catch{throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}}return t}dataToGPU(t){return this.throwIfDisposed(),ce().readToGPU(this.dataId,t)}dataSync(){this.throwIfDisposed();const t=ce().readSync(this.dataId);if(this.dtype==="string")try{return t.map(e=>Ks(e))}catch{throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}return t}async bytes(){this.throwIfDisposed();const t=await ce().read(this.dataId);return this.dtype==="string"?t:new Uint8Array(t.buffer)}dispose(){this.isDisposed||(this.kerasMask&&this.kerasMask.dispose(),ce().disposeTensor(this),this.isDisposedInternal=!0)}get isDisposed(){return this.isDisposedInternal}throwIfDisposed(){if(this.isDisposed)throw new Error("Tensor is disposed.")}print(t=!1){return Dn.print(this,t)}clone(){return this.throwIfDisposed(),Dn.clone(this)}toString(t=!1){const e=this.dataSync();return Ap(e,this.shape,this.dtype,t)}cast(t){return this.throwIfDisposed(),Dn.cast(this,t)}variable(t=!0,e,s){return this.throwIfDisposed(),ce().makeVariable(this,t,e,s)}}Object.defineProperty(At,Symbol.hasInstance,{value:n=>!!n&&n.data!=null&&n.dataSync!=null&&n.throwIfDisposed!=null});function el(){return mo("Tensor",()=>At)}el();class Js extends At{constructor(t,e,s,r){super(t.shape,t.dtype,t.dataId,r),this.trainable=e,this.name=s}assign(t){if(t.dtype!==this.dtype)throw new Error(`dtype of the new value (${t.dtype}) and previous value (${this.dtype}) must match`);if(!Zt(t.shape,this.shape))throw new Error(`shape of the new value (${t.shape}) and previous value (${this.shape}) must match`);ce().disposeTensor(this),this.dataId=t.dataId,ce().incRef(this,null)}dispose(){ce().disposeVariable(this),this.isDisposedInternal=!0}}Object.defineProperty(Js,Symbol.hasInstance,{value:n=>n instanceof At&&n.assign!=null&&n.assign instanceof Function});/**
+ */class Js{constructor(t,e,s){if(this.dtype=e,this.shape=t.slice(),this.size=z(t),s!=null){const r=s.length;w(r===this.size,()=>`Length of values '${r}' does not match the size inferred by the shape '${this.size}'.`)}if(e==="complex64")throw new Error("complex64 dtype TensorBuffers are not supported. Please create a TensorBuffer for the real and imaginary parts separately and call tf.complex(real, imag).");this.values=s||bt(e,this.size),this.strides=Kt(t)}set(t,...e){e.length===0&&(e=[0]),w(e.length===this.rank,()=>`The number of provided coordinates (${e.length}) must match the rank (${this.rank})`);const s=this.locToIndex(e);this.values[s]=t}get(...t){t.length===0&&(t=[0]);let e=0;for(const r of t){if(r<0||r>=this.shape[e]){const o=`Requested out of range element at ${t}.   Buffer shape=${this.shape}`;throw new Error(o)}e++}let s=t[t.length-1];for(let r=0;r<t.length-1;++r)s+=this.strides[r]*t[r];return this.values[s]}locToIndex(t){if(this.rank===0)return 0;if(this.rank===1)return t[0];let e=t[t.length-1];for(let s=0;s<t.length-1;++s)e+=this.strides[s]*t[s];return e}indexToLoc(t){if(this.rank===0)return[];if(this.rank===1)return[t];const e=new Array(this.shape.length);for(let s=0;s<e.length-1;++s)e[s]=Math.floor(t/this.strides[s]),t-=e[s]*this.strides[s];return e[e.length-1]=t,e}get rank(){return this.shape.length}toTensor(){return ce().makeTensor(this.values,this.shape,this.dtype)}}let ce=null,Dn=null;function Cp(n){ce=n}function kp(n){Dn=n}class At{constructor(t,e,s,r){this.kept=!1,this.isDisposedInternal=!1,this.shape=t.slice(),this.dtype=e||"float32",this.size=z(t),this.strides=Kt(t),this.dataId=s,this.id=r,this.rankType=this.rank<5?this.rank.toString():"higher"}get rank(){return this.shape.length}async buffer(){const t=await this.data();return Dn.buffer(this.shape,this.dtype,t)}bufferSync(){return Dn.buffer(this.shape,this.dtype,this.dataSync())}async array(){const t=await this.data();return xa(this.shape,t,this.dtype==="complex64")}arraySync(){return xa(this.shape,this.dataSync(),this.dtype==="complex64")}async data(){this.throwIfDisposed();const t=ce().read(this.dataId);if(this.dtype==="string"){const e=await t;try{return e.map(s=>Ys(s))}catch{throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}}return t}dataToGPU(t){return this.throwIfDisposed(),ce().readToGPU(this.dataId,t)}dataSync(){this.throwIfDisposed();const t=ce().readSync(this.dataId);if(this.dtype==="string")try{return t.map(e=>Ys(e))}catch{throw new Error("Failed to decode the string bytes into utf-8. To get the original bytes, call tensor.bytes().")}return t}async bytes(){this.throwIfDisposed();const t=await ce().read(this.dataId);return this.dtype==="string"?t:new Uint8Array(t.buffer)}dispose(){this.isDisposed||(this.kerasMask&&this.kerasMask.dispose(),ce().disposeTensor(this),this.isDisposedInternal=!0)}get isDisposed(){return this.isDisposedInternal}throwIfDisposed(){if(this.isDisposed)throw new Error("Tensor is disposed.")}print(t=!1){return Dn.print(this,t)}clone(){return this.throwIfDisposed(),Dn.clone(this)}toString(t=!1){const e=this.dataSync();return Ap(e,this.shape,this.dtype,t)}cast(t){return this.throwIfDisposed(),Dn.cast(this,t)}variable(t=!0,e,s){return this.throwIfDisposed(),ce().makeVariable(this,t,e,s)}}Object.defineProperty(At,Symbol.hasInstance,{value:n=>!!n&&n.data!=null&&n.dataSync!=null&&n.throwIfDisposed!=null});function el(){return mo("Tensor",()=>At)}el();class Zs extends At{constructor(t,e,s,r){super(t.shape,t.dtype,t.dataId,r),this.trainable=e,this.name=s}assign(t){if(t.dtype!==this.dtype)throw new Error(`dtype of the new value (${t.dtype}) and previous value (${this.dtype}) must match`);if(!Zt(t.shape,this.shape))throw new Error(`shape of the new value (${t.shape}) and previous value (${this.shape}) must match`);ce().disposeTensor(this),this.dataId=t.dataId,ce().incRef(this,null)}dispose(){ce().disposeVariable(this),this.isDisposedInternal=!0}}Object.defineProperty(Zs,Symbol.hasInstance,{value:n=>n instanceof At&&n.assign!=null&&n.assign instanceof Function});/**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -242,7 +242,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function ko(n){return n.kernelName!=null}class al{constructor(){this.registeredVariables={},this.nextTapeNodeId=0,this.numBytes=0,this.numTensors=0,this.numStringTensors=0,this.numDataBuffers=0,this.gradientDepth=0,this.kernelDepth=0,this.scopeStack=[],this.numDataMovesStack=[],this.nextScopeId=0,this.tensorInfo=new WeakMap,this.profiling=!1,this.activeProfile={newBytes:0,newTensors:0,peakBytes:0,kernels:[],result:null,get kernelNames(){return Array.from(new Set(this.kernels.map(t=>t.name)))}}}dispose(){for(const t in this.registeredVariables)this.registeredVariables[t].dispose()}}class Rn{constructor(t){this.ENV=t,this.registry={},this.registryFactory={},this.pendingBackendInitId=0,this.state=new al}async ready(){if(this.pendingBackendInit!=null)return this.pendingBackendInit.then(()=>{});if(this.backendInstance!=null)return;const t=this.getSortedBackends();for(let e=0;e<t.length;e++){const s=t[e];if(await this.initializeBackend(s).success){await this.setBackend(s);return}}throw new Error("Could not initialize any backends, all backend initializations failed.")}get backend(){if(this.pendingBackendInit!=null)throw new Error(`Backend '${this.backendName}' has not yet been initialized. Make sure to await tf.ready() or await tf.setBackend() before calling other methods`);if(this.backendInstance==null){const{name:t,asyncInit:e}=this.initializeBackendsAndReturnBest();if(e)throw new Error(`The highest priority backend '${t}' has not yet been initialized. Make sure to await tf.ready() or await tf.setBackend() before calling other methods`);this.setBackend(t)}return this.backendInstance}backendNames(){return Object.keys(this.registryFactory)}findBackend(t){if(!(t in this.registry))if(t in this.registryFactory){const{asyncInit:e}=this.initializeBackend(t);if(e)return null}else return null;return this.registry[t]}findBackendFactory(t){return t in this.registryFactory?this.registryFactory[t].factory:null}registerBackend(t,e,s=1){return t in this.registryFactory?(kn(`${t} backend was already registered. Reusing existing backend factory.`),!1):(this.registryFactory[t]={factory:e,priority:s},!0)}async setBackend(t){if(this.registryFactory[t]==null)throw new Error(`Backend name '${t}' not found in registry`);if(this.backendName=t,this.registry[t]==null){this.backendInstance=null;const{success:e,asyncInit:s}=this.initializeBackend(t);if(!(s?await e:e))return!1}return this.backendInstance=this.registry[t],this.setupRegisteredKernels(),this.profiler=new xp(this.backendInstance),!0}setupRegisteredKernels(){Ma(this.backendName).forEach(e=>{e.setupFunc!=null&&e.setupFunc(this.backendInstance)})}disposeRegisteredKernels(t){Ma(t).forEach(s=>{s.disposeFunc!=null&&s.disposeFunc(this.registry[t])})}initializeBackend(t){const e=this.registryFactory[t];if(e==null)throw new Error(`Cannot initialize backend ${t}, no registration found.`);try{const s=e.factory();if(s&&!(s instanceof ba)&&typeof s.then=="function"){const r=++this.pendingBackendInitId,o=s.then(i=>r<this.pendingBackendInitId?!1:(this.registry[t]=i,this.pendingBackendInit=null,!0)).catch(i=>(r<this.pendingBackendInitId||(this.pendingBackendInit=null,kn(`Initialization of backend ${t} failed`),kn(i.stack||i.message)),!1));return this.pendingBackendInit=o,{success:o,asyncInit:!0}}else return this.registry[t]=s,{success:!0,asyncInit:!1}}catch(s){return kn(`Initialization of backend ${t} failed`),kn(s.stack||s.message),{success:!1,asyncInit:!1}}}removeBackend(t){if(!(t in this.registryFactory))throw new Error(`${t} backend not found in registry`);this.backendName===t&&this.pendingBackendInit!=null&&this.pendingBackendInitId++,t in this.registry&&(this.disposeRegisteredKernels(t),this.registry[t].dispose(),delete this.registry[t]),delete this.registryFactory[t],this.backendName===t&&(this.pendingBackendInit=null,this.backendName=null,this.backendInstance=null)}getSortedBackends(){if(Object.keys(this.registryFactory).length===0)throw new Error("No backend found in registry.");return Object.keys(this.registryFactory).sort((t,e)=>this.registryFactory[e].priority-this.registryFactory[t].priority)}initializeBackendsAndReturnBest(){const t=this.getSortedBackends();for(let e=0;e<t.length;e++){const s=t[e],{success:r,asyncInit:o}=this.initializeBackend(s);if(o||r)return{name:s,asyncInit:o}}throw new Error("Could not initialize any backends, all backend initializations failed.")}moveData(t,e){const s=this.state.tensorInfo.get(e),r=s.backend,o=this.readSync(e),i=r.refCount(e);r.disposeData(e,!0),s.backend=t,t.move(e,o,s.shape,s.dtype,i),this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack[this.state.numDataMovesStack.length-1]++}tidy(t,e){let s=null;if(e==null){if(typeof t!="function")throw new Error("Please provide a function to tidy()");e=t}else{if(typeof t!="string"&&!(t instanceof String))throw new Error("When calling with two arguments, the first argument to tidy() must be a string");if(typeof e!="function")throw new Error("When calling with two arguments, the 2nd argument to tidy() must be a function");s=t}let r;return this.scopedRun(()=>this.startScope(s),()=>this.endScope(r),()=>(r=e(),r instanceof Promise&&console.error("Cannot return a Promise inside of tidy."),r))}scopedRun(t,e,s){t();try{const r=s();return e(),r}catch(r){throw e(),r}}nextTensorId(){return Rn.nextTensorId++}nextVariableId(){return Rn.nextVariableId++}clone(t){const e=A.runKernel(go,{x:t}),s={x:t},r=i=>({x:()=>{const a="float32",l={x:i},u={dtype:a};return A.runKernel(Ea,l,u)}}),o=[];return this.addTapeNode(this.state.activeScope.name,s,[e],r,o,{}),e}runKernel(t,e,s){if(this.backendName==null&&this.backend,!(Pa(t,this.backendName)!=null))throw new Error(`Kernel '${t}' not registered for backend '${this.backendName}'`);return this.runKernelFunc({kernelName:t,inputs:e,attrs:s})}shouldCheckForMemLeaks(){return this.ENV.getBool("IS_TEST")}checkKernelForMemLeak(t,e,s){const r=this.backend.numDataIds();let o=0;s.forEach(l=>{o+=l.dtype==="complex64"?3:1});const i=this.state.numDataMovesStack[this.state.numDataMovesStack.length-1],a=r-e-o-i;if(a>0)throw new Error(`Backend '${this.backendName}' has an internal memory leak (${a} data ids) after running '${t}'`)}runKernelFunc(t){let e,s=[];const r=this.isTapeOn(),o=this.state.numBytes,i=this.state.numTensors;this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack.push(0);let a;this.backendName==null&&this.backend;let l;const u=ko(t)?t.kernelName:this.state.activeScope!=null?this.state.activeScope.name:"";if(ko(t)){const{kernelName:p,inputs:g,attrs:m}=t;this.backendName==null&&this.backend;const b=Pa(p,this.backendName);w(b!=null,()=>`Cannot find registered kernel '${p}' for backend '${this.backendName}'`),a=()=>{const y=this.backend.numDataIds();l=b.kernelFunc({inputs:g,attrs:m,backend:this.backend});const S=Array.isArray(l)?l:[l];this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(p,y,S);const x=S.map($=>$.rank!=null?$:this.makeTensorFromTensorInfo($));if(r){const $=this.getTensorsForGradient(p,g,x);s=this.saveTensorsForBackwardMode($)}return x}}else{const{forwardFunc:p}=t,g=m=>{r&&(s=m.map(b=>this.keep(this.clone(b))))};a=()=>{const m=this.backend.numDataIds();l=this.tidy(()=>p(this.backend,g));const b=Array.isArray(l)?l:[l];return this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(u,m,b),b}}const{inputs:c,attrs:h}=t,f=ko(t)?null:t.backwardsFunc;let d;return this.scopedRun(()=>this.state.kernelDepth++,()=>this.state.kernelDepth--,()=>{!this.ENV.getBool("DEBUG")&&!this.state.profiling?e=a():(d=this.profiler.profileKernel(u,c,()=>a()),this.ENV.getBool("DEBUG")&&this.profiler.logKernelProfile(d),e=d.outputs)}),r&&this.addTapeNode(u,c,e,f,s,h),this.state.profiling&&this.state.activeProfile.kernels.push({name:u,bytesAdded:this.state.numBytes-o,totalBytesSnapshot:this.state.numBytes,tensorsAdded:this.state.numTensors-i,totalTensorsSnapshot:this.state.numTensors,inputShapes:Object.keys(c).map(p=>c[p]!=null?c[p].shape:null),outputShapes:e.map(p=>p.shape),kernelTimeMs:d.timeMs,extraInfo:d.extraInfo}),Array.isArray(l)?e:e[0]}saveTensorsForBackwardMode(t){return t.map(s=>this.keep(this.clone(s)))}getTensorsForGradient(t,e,s){const r=La(t);if(r!=null){const o=r.inputsToSave||[],i=r.outputsToSave||[];let a;r.saveAllInputs?(w(Array.isArray(e),()=>"saveAllInputs is true, expected inputs to be an array."),a=Object.keys(e).map(u=>e[u])):a=o.map(u=>e[u]);const l=s.filter((u,c)=>i[c]);return a.concat(l)}return[]}makeTensor(t,e,s,r){if(t==null)throw new Error("Values passed to engine.makeTensor() are null");s=s||"float32",r=r||this.backend;let o=t;s==="string"&&Ws(t[0])&&(o=t.map(l=>nn(l)));const i=r.write(o,e,s),a=new At(e,s,i,this.nextTensorId());if(this.trackTensor(a,r),s==="string"){const l=this.state.tensorInfo.get(i),u=Nf(o);this.state.numBytes+=u-l.bytes,l.bytes=u}return a}makeTensorFromDataId(t,e,s,r){s=s||"float32";const o={dataId:t,shape:e,dtype:s};return this.makeTensorFromTensorInfo(o,r)}makeTensorFromTensorInfo(t,e){const{dataId:s,shape:r,dtype:o}=t,i=new At(r,o,s,this.nextTensorId());return this.trackTensor(i,e),i}makeVariable(t,e=!0,s,r){s=s||this.nextVariableId().toString(),r!=null&&r!==t.dtype&&(t=t.cast(r));const o=new Js(t,e,s,this.nextTensorId());if(this.state.registeredVariables[o.name]!=null)throw new Error(`Variable with name ${o.name} was already registered`);return this.state.registeredVariables[o.name]=o,this.incRef(o,this.backend),o}trackTensor(t,e){this.state.numTensors++,t.dtype==="string"&&this.state.numStringTensors++;let s=0;t.dtype!=="complex64"&&t.dtype!=="string"&&(s=t.size*io(t.dtype)),this.state.numBytes+=s,this.state.tensorInfo.has(t.dataId)||(this.state.numDataBuffers++,this.state.tensorInfo.set(t.dataId,{backend:e||this.backend,dtype:t.dtype,shape:t.shape,bytes:s})),t instanceof Js||this.track(t)}incRef(t,e){this.trackTensor(t,e),this.backend.incRef(t.dataId)}removeDataId(t,e){this.state.tensorInfo.has(t)&&this.state.tensorInfo.get(t).backend===e&&(this.state.tensorInfo.delete(t),this.state.numDataBuffers--)}disposeTensor(t){if(!this.state.tensorInfo.has(t.dataId))return;const e=this.state.tensorInfo.get(t.dataId);if(this.state.numTensors--,t.dtype==="string"&&(this.state.numStringTensors--,this.state.numBytes-=e.bytes),t.dtype!=="complex64"&&t.dtype!=="string"){const s=t.size*io(t.dtype);this.state.numBytes-=s}e.backend.disposeData(t.dataId)&&this.removeDataId(t.dataId,e.backend)}disposeVariables(){for(const t in this.state.registeredVariables){const e=this.state.registeredVariables[t];this.disposeVariable(e)}}disposeVariable(t){this.disposeTensor(t),this.state.registeredVariables[t.name]!=null&&delete this.state.registeredVariables[t.name]}memory(){const t=this.backend.memory();return t.numTensors=this.state.numTensors,t.numDataBuffers=this.state.numDataBuffers,t.numBytes=this.state.numBytes,this.state.numStringTensors>0&&(t.unreliable=!0,t.reasons==null&&(t.reasons=[]),t.reasons.push("Memory usage by string tensors is approximate (2 bytes per character)")),t}async profile(t){this.state.profiling=!0;const e=this.state.numBytes,s=this.state.numTensors;this.state.activeProfile.kernels=[],this.state.activeProfile.result=await t(),this.state.profiling=!1,this.state.activeProfile.peakBytes=Math.max(...this.state.activeProfile.kernels.map(r=>r.totalBytesSnapshot)),this.state.activeProfile.newBytes=this.state.numBytes-e,this.state.activeProfile.newTensors=this.state.numTensors-s;for(const r of this.state.activeProfile.kernels)r.kernelTimeMs=await r.kernelTimeMs,r.extraInfo=await r.extraInfo;return this.state.activeProfile}isTapeOn(){return this.state.gradientDepth>0&&this.state.kernelDepth===0}addTapeNode(t,e,s,r,o,i){const a={id:this.state.nextTapeNodeId++,kernelName:t,inputs:e,outputs:s,saved:o},l=La(t);l!=null&&(r=l.gradFunc),r!=null&&(a.gradient=u=>(u=u.map((c,h)=>{if(c==null){const f=s[h],d=ze(f.size,f.dtype);return this.makeTensor(d,f.shape,f.dtype)}return c}),r(u.length>1?u:u[0],o,i))),this.state.activeTape.push(a)}keep(t){return t.kept=!0,t}startTape(){this.state.gradientDepth===0&&(this.state.activeTape=[]),this.state.gradientDepth++}endTape(){this.state.gradientDepth--}startScope(t){const e={track:[],name:"unnamed scope",id:this.state.nextScopeId++};t&&(e.name=t),this.state.scopeStack.push(e),this.state.activeScope=e}endScope(t){const e=ol(t),s=new Set(e.map(o=>o.id));for(let o=0;o<this.state.activeScope.track.length;o++){const i=this.state.activeScope.track[o];!i.kept&&!s.has(i.id)&&i.dispose()}const r=this.state.scopeStack.pop();this.state.activeScope=this.state.scopeStack.length===0?null:this.state.scopeStack[this.state.scopeStack.length-1],e.forEach(o=>{!o.kept&&o.scopeId===r.id&&this.track(o)})}gradients(t,e,s,r=!1){if(w(e.length>0,()=>"gradients() received an empty list of xs."),s!=null&&s.dtype!=="float32")throw new Error(`dy must have 'float32' dtype, but has '${s.dtype}'`);const o=this.scopedRun(()=>this.startTape(),()=>this.endTape(),()=>this.tidy("forward",t));w(o instanceof At,()=>"The result y returned by f() must be a tensor.");const i=vp(this.state.activeTape,e,o);if(!r&&i.length===0&&e.length>0)throw new Error("Cannot compute gradient of y=f(x) with respect to x. Make sure that the f you passed encloses all operations that lead from x to y.");return this.tidy("backward",()=>{const a={};a[o.id]=s??Dp(o.shape),Ip(a,i,u=>this.tidy(u),Rp);const l=e.map(u=>a[u.id]);return this.state.gradientDepth===0&&(this.state.activeTape.forEach(u=>{for(const c of u.saved)c.dispose()}),this.state.activeTape=null),{value:o,grads:l}})}customGrad(t){return w(lo(t),()=>"The f passed in customGrad(f) must be a function."),(...e)=>{w(e.every(a=>a instanceof At),()=>"The args passed in customGrad(f)(x1, x2,...) must all be tensors");let s;const r={};e.forEach((a,l)=>{r[l]=a});const o=(a,l)=>(s=t(...e,l),w(s.value instanceof At,()=>"The function f passed in customGrad(f) must return an object where `obj.value` is a tensor"),w(lo(s.gradFunc),()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function."),s.value),i=(a,l)=>{const u=s.gradFunc(a,l),c=Array.isArray(u)?u:[u];w(c.length===e.length,()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function that returns the same number of tensors as inputs passed to f(...)."),w(c.every(f=>f instanceof At),()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function that returns a list of only tensors.");const h={};return c.forEach((f,d)=>{h[d]=()=>f}),h};return this.runKernelFunc({forwardFunc:o,backwardsFunc:i,inputs:r})}}readSync(t){return this.state.tensorInfo.get(t).backend.readSync(t)}read(t){return this.state.tensorInfo.get(t).backend.read(t)}readToGPU(t,e){return this.state.tensorInfo.get(t).backend.readToGPU(t,e)}async time(t){const e=Nn(),s=await this.backend.time(t);return s.wallMs=Nn()-e,s}track(t){return this.state.activeScope!=null&&(t.scopeId=this.state.activeScope.id,this.state.activeScope.track.push(t)),t}get registeredVariables(){return this.state.registeredVariables}reset(){this.pendingBackendInitId++,this.state.dispose(),this.ENV.reset(),this.state=new al;for(const t in this.registry)this.disposeRegisteredKernels(t),this.registry[t].dispose(),delete this.registry[t];this.backendName=null,this.backendInstance=null,this.pendingBackendInit=null}}Rn.nextTensorId=0,Rn.nextVariableId=0;function Dp(n){const t=Sa(z(n),"float32");return A.makeTensor(t,n,"float32")}function ll(){const n=Ia();if(n._tfengine==null){const t=new Rf(n);n._tfengine=new Rn(t)}return Of(n._tfengine.ENV),Cp(()=>n._tfengine),n._tfengine}const A=ll();function Rp(n,t){const e={a:n,b:t};return A.runKernel(Aa,e)}/**
+ */function ko(n){return n.kernelName!=null}class al{constructor(){this.registeredVariables={},this.nextTapeNodeId=0,this.numBytes=0,this.numTensors=0,this.numStringTensors=0,this.numDataBuffers=0,this.gradientDepth=0,this.kernelDepth=0,this.scopeStack=[],this.numDataMovesStack=[],this.nextScopeId=0,this.tensorInfo=new WeakMap,this.profiling=!1,this.activeProfile={newBytes:0,newTensors:0,peakBytes:0,kernels:[],result:null,get kernelNames(){return Array.from(new Set(this.kernels.map(t=>t.name)))}}}dispose(){for(const t in this.registeredVariables)this.registeredVariables[t].dispose()}}class Rn{constructor(t){this.ENV=t,this.registry={},this.registryFactory={},this.pendingBackendInitId=0,this.state=new al}async ready(){if(this.pendingBackendInit!=null)return this.pendingBackendInit.then(()=>{});if(this.backendInstance!=null)return;const t=this.getSortedBackends();for(let e=0;e<t.length;e++){const s=t[e];if(await this.initializeBackend(s).success){await this.setBackend(s);return}}throw new Error("Could not initialize any backends, all backend initializations failed.")}get backend(){if(this.pendingBackendInit!=null)throw new Error(`Backend '${this.backendName}' has not yet been initialized. Make sure to await tf.ready() or await tf.setBackend() before calling other methods`);if(this.backendInstance==null){const{name:t,asyncInit:e}=this.initializeBackendsAndReturnBest();if(e)throw new Error(`The highest priority backend '${t}' has not yet been initialized. Make sure to await tf.ready() or await tf.setBackend() before calling other methods`);this.setBackend(t)}return this.backendInstance}backendNames(){return Object.keys(this.registryFactory)}findBackend(t){if(!(t in this.registry))if(t in this.registryFactory){const{asyncInit:e}=this.initializeBackend(t);if(e)return null}else return null;return this.registry[t]}findBackendFactory(t){return t in this.registryFactory?this.registryFactory[t].factory:null}registerBackend(t,e,s=1){return t in this.registryFactory?(kn(`${t} backend was already registered. Reusing existing backend factory.`),!1):(this.registryFactory[t]={factory:e,priority:s},!0)}async setBackend(t){if(this.registryFactory[t]==null)throw new Error(`Backend name '${t}' not found in registry`);if(this.backendName=t,this.registry[t]==null){this.backendInstance=null;const{success:e,asyncInit:s}=this.initializeBackend(t);if(!(s?await e:e))return!1}return this.backendInstance=this.registry[t],this.setupRegisteredKernels(),this.profiler=new xp(this.backendInstance),!0}setupRegisteredKernels(){Ma(this.backendName).forEach(e=>{e.setupFunc!=null&&e.setupFunc(this.backendInstance)})}disposeRegisteredKernels(t){Ma(t).forEach(s=>{s.disposeFunc!=null&&s.disposeFunc(this.registry[t])})}initializeBackend(t){const e=this.registryFactory[t];if(e==null)throw new Error(`Cannot initialize backend ${t}, no registration found.`);try{const s=e.factory();if(s&&!(s instanceof ba)&&typeof s.then=="function"){const r=++this.pendingBackendInitId,o=s.then(i=>r<this.pendingBackendInitId?!1:(this.registry[t]=i,this.pendingBackendInit=null,!0)).catch(i=>(r<this.pendingBackendInitId||(this.pendingBackendInit=null,kn(`Initialization of backend ${t} failed`),kn(i.stack||i.message)),!1));return this.pendingBackendInit=o,{success:o,asyncInit:!0}}else return this.registry[t]=s,{success:!0,asyncInit:!1}}catch(s){return kn(`Initialization of backend ${t} failed`),kn(s.stack||s.message),{success:!1,asyncInit:!1}}}removeBackend(t){if(!(t in this.registryFactory))throw new Error(`${t} backend not found in registry`);this.backendName===t&&this.pendingBackendInit!=null&&this.pendingBackendInitId++,t in this.registry&&(this.disposeRegisteredKernels(t),this.registry[t].dispose(),delete this.registry[t]),delete this.registryFactory[t],this.backendName===t&&(this.pendingBackendInit=null,this.backendName=null,this.backendInstance=null)}getSortedBackends(){if(Object.keys(this.registryFactory).length===0)throw new Error("No backend found in registry.");return Object.keys(this.registryFactory).sort((t,e)=>this.registryFactory[e].priority-this.registryFactory[t].priority)}initializeBackendsAndReturnBest(){const t=this.getSortedBackends();for(let e=0;e<t.length;e++){const s=t[e],{success:r,asyncInit:o}=this.initializeBackend(s);if(o||r)return{name:s,asyncInit:o}}throw new Error("Could not initialize any backends, all backend initializations failed.")}moveData(t,e){const s=this.state.tensorInfo.get(e),r=s.backend,o=this.readSync(e),i=r.refCount(e);r.disposeData(e,!0),s.backend=t,t.move(e,o,s.shape,s.dtype,i),this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack[this.state.numDataMovesStack.length-1]++}tidy(t,e){let s=null;if(e==null){if(typeof t!="function")throw new Error("Please provide a function to tidy()");e=t}else{if(typeof t!="string"&&!(t instanceof String))throw new Error("When calling with two arguments, the first argument to tidy() must be a string");if(typeof e!="function")throw new Error("When calling with two arguments, the 2nd argument to tidy() must be a function");s=t}let r;return this.scopedRun(()=>this.startScope(s),()=>this.endScope(r),()=>(r=e(),r instanceof Promise&&console.error("Cannot return a Promise inside of tidy."),r))}scopedRun(t,e,s){t();try{const r=s();return e(),r}catch(r){throw e(),r}}nextTensorId(){return Rn.nextTensorId++}nextVariableId(){return Rn.nextVariableId++}clone(t){const e=A.runKernel(go,{x:t}),s={x:t},r=i=>({x:()=>{const a="float32",l={x:i},u={dtype:a};return A.runKernel(Ea,l,u)}}),o=[];return this.addTapeNode(this.state.activeScope.name,s,[e],r,o,{}),e}runKernel(t,e,s){if(this.backendName==null&&this.backend,!(Pa(t,this.backendName)!=null))throw new Error(`Kernel '${t}' not registered for backend '${this.backendName}'`);return this.runKernelFunc({kernelName:t,inputs:e,attrs:s})}shouldCheckForMemLeaks(){return this.ENV.getBool("IS_TEST")}checkKernelForMemLeak(t,e,s){const r=this.backend.numDataIds();let o=0;s.forEach(l=>{o+=l.dtype==="complex64"?3:1});const i=this.state.numDataMovesStack[this.state.numDataMovesStack.length-1],a=r-e-o-i;if(a>0)throw new Error(`Backend '${this.backendName}' has an internal memory leak (${a} data ids) after running '${t}'`)}runKernelFunc(t){let e,s=[];const r=this.isTapeOn(),o=this.state.numBytes,i=this.state.numTensors;this.shouldCheckForMemLeaks()&&this.state.numDataMovesStack.push(0);let a;this.backendName==null&&this.backend;let l;const u=ko(t)?t.kernelName:this.state.activeScope!=null?this.state.activeScope.name:"";if(ko(t)){const{kernelName:p,inputs:g,attrs:m}=t;this.backendName==null&&this.backend;const b=Pa(p,this.backendName);w(b!=null,()=>`Cannot find registered kernel '${p}' for backend '${this.backendName}'`),a=()=>{const y=this.backend.numDataIds();l=b.kernelFunc({inputs:g,attrs:m,backend:this.backend});const S=Array.isArray(l)?l:[l];this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(p,y,S);const x=S.map($=>$.rank!=null?$:this.makeTensorFromTensorInfo($));if(r){const $=this.getTensorsForGradient(p,g,x);s=this.saveTensorsForBackwardMode($)}return x}}else{const{forwardFunc:p}=t,g=m=>{r&&(s=m.map(b=>this.keep(this.clone(b))))};a=()=>{const m=this.backend.numDataIds();l=this.tidy(()=>p(this.backend,g));const b=Array.isArray(l)?l:[l];return this.shouldCheckForMemLeaks()&&this.checkKernelForMemLeak(u,m,b),b}}const{inputs:c,attrs:h}=t,f=ko(t)?null:t.backwardsFunc;let d;return this.scopedRun(()=>this.state.kernelDepth++,()=>this.state.kernelDepth--,()=>{!this.ENV.getBool("DEBUG")&&!this.state.profiling?e=a():(d=this.profiler.profileKernel(u,c,()=>a()),this.ENV.getBool("DEBUG")&&this.profiler.logKernelProfile(d),e=d.outputs)}),r&&this.addTapeNode(u,c,e,f,s,h),this.state.profiling&&this.state.activeProfile.kernels.push({name:u,bytesAdded:this.state.numBytes-o,totalBytesSnapshot:this.state.numBytes,tensorsAdded:this.state.numTensors-i,totalTensorsSnapshot:this.state.numTensors,inputShapes:Object.keys(c).map(p=>c[p]!=null?c[p].shape:null),outputShapes:e.map(p=>p.shape),kernelTimeMs:d.timeMs,extraInfo:d.extraInfo}),Array.isArray(l)?e:e[0]}saveTensorsForBackwardMode(t){return t.map(s=>this.keep(this.clone(s)))}getTensorsForGradient(t,e,s){const r=La(t);if(r!=null){const o=r.inputsToSave||[],i=r.outputsToSave||[];let a;r.saveAllInputs?(w(Array.isArray(e),()=>"saveAllInputs is true, expected inputs to be an array."),a=Object.keys(e).map(u=>e[u])):a=o.map(u=>e[u]);const l=s.filter((u,c)=>i[c]);return a.concat(l)}return[]}makeTensor(t,e,s,r){if(t==null)throw new Error("Values passed to engine.makeTensor() are null");s=s||"float32",r=r||this.backend;let o=t;s==="string"&&Gs(t[0])&&(o=t.map(l=>sn(l)));const i=r.write(o,e,s),a=new At(e,s,i,this.nextTensorId());if(this.trackTensor(a,r),s==="string"){const l=this.state.tensorInfo.get(i),u=Nf(o);this.state.numBytes+=u-l.bytes,l.bytes=u}return a}makeTensorFromDataId(t,e,s,r){s=s||"float32";const o={dataId:t,shape:e,dtype:s};return this.makeTensorFromTensorInfo(o,r)}makeTensorFromTensorInfo(t,e){const{dataId:s,shape:r,dtype:o}=t,i=new At(r,o,s,this.nextTensorId());return this.trackTensor(i,e),i}makeVariable(t,e=!0,s,r){s=s||this.nextVariableId().toString(),r!=null&&r!==t.dtype&&(t=t.cast(r));const o=new Zs(t,e,s,this.nextTensorId());if(this.state.registeredVariables[o.name]!=null)throw new Error(`Variable with name ${o.name} was already registered`);return this.state.registeredVariables[o.name]=o,this.incRef(o,this.backend),o}trackTensor(t,e){this.state.numTensors++,t.dtype==="string"&&this.state.numStringTensors++;let s=0;t.dtype!=="complex64"&&t.dtype!=="string"&&(s=t.size*io(t.dtype)),this.state.numBytes+=s,this.state.tensorInfo.has(t.dataId)||(this.state.numDataBuffers++,this.state.tensorInfo.set(t.dataId,{backend:e||this.backend,dtype:t.dtype,shape:t.shape,bytes:s})),t instanceof Zs||this.track(t)}incRef(t,e){this.trackTensor(t,e),this.backend.incRef(t.dataId)}removeDataId(t,e){this.state.tensorInfo.has(t)&&this.state.tensorInfo.get(t).backend===e&&(this.state.tensorInfo.delete(t),this.state.numDataBuffers--)}disposeTensor(t){if(!this.state.tensorInfo.has(t.dataId))return;const e=this.state.tensorInfo.get(t.dataId);if(this.state.numTensors--,t.dtype==="string"&&(this.state.numStringTensors--,this.state.numBytes-=e.bytes),t.dtype!=="complex64"&&t.dtype!=="string"){const s=t.size*io(t.dtype);this.state.numBytes-=s}e.backend.disposeData(t.dataId)&&this.removeDataId(t.dataId,e.backend)}disposeVariables(){for(const t in this.state.registeredVariables){const e=this.state.registeredVariables[t];this.disposeVariable(e)}}disposeVariable(t){this.disposeTensor(t),this.state.registeredVariables[t.name]!=null&&delete this.state.registeredVariables[t.name]}memory(){const t=this.backend.memory();return t.numTensors=this.state.numTensors,t.numDataBuffers=this.state.numDataBuffers,t.numBytes=this.state.numBytes,this.state.numStringTensors>0&&(t.unreliable=!0,t.reasons==null&&(t.reasons=[]),t.reasons.push("Memory usage by string tensors is approximate (2 bytes per character)")),t}async profile(t){this.state.profiling=!0;const e=this.state.numBytes,s=this.state.numTensors;this.state.activeProfile.kernels=[],this.state.activeProfile.result=await t(),this.state.profiling=!1,this.state.activeProfile.peakBytes=Math.max(...this.state.activeProfile.kernels.map(r=>r.totalBytesSnapshot)),this.state.activeProfile.newBytes=this.state.numBytes-e,this.state.activeProfile.newTensors=this.state.numTensors-s;for(const r of this.state.activeProfile.kernels)r.kernelTimeMs=await r.kernelTimeMs,r.extraInfo=await r.extraInfo;return this.state.activeProfile}isTapeOn(){return this.state.gradientDepth>0&&this.state.kernelDepth===0}addTapeNode(t,e,s,r,o,i){const a={id:this.state.nextTapeNodeId++,kernelName:t,inputs:e,outputs:s,saved:o},l=La(t);l!=null&&(r=l.gradFunc),r!=null&&(a.gradient=u=>(u=u.map((c,h)=>{if(c==null){const f=s[h],d=ze(f.size,f.dtype);return this.makeTensor(d,f.shape,f.dtype)}return c}),r(u.length>1?u:u[0],o,i))),this.state.activeTape.push(a)}keep(t){return t.kept=!0,t}startTape(){this.state.gradientDepth===0&&(this.state.activeTape=[]),this.state.gradientDepth++}endTape(){this.state.gradientDepth--}startScope(t){const e={track:[],name:"unnamed scope",id:this.state.nextScopeId++};t&&(e.name=t),this.state.scopeStack.push(e),this.state.activeScope=e}endScope(t){const e=ol(t),s=new Set(e.map(o=>o.id));for(let o=0;o<this.state.activeScope.track.length;o++){const i=this.state.activeScope.track[o];!i.kept&&!s.has(i.id)&&i.dispose()}const r=this.state.scopeStack.pop();this.state.activeScope=this.state.scopeStack.length===0?null:this.state.scopeStack[this.state.scopeStack.length-1],e.forEach(o=>{!o.kept&&o.scopeId===r.id&&this.track(o)})}gradients(t,e,s,r=!1){if(w(e.length>0,()=>"gradients() received an empty list of xs."),s!=null&&s.dtype!=="float32")throw new Error(`dy must have 'float32' dtype, but has '${s.dtype}'`);const o=this.scopedRun(()=>this.startTape(),()=>this.endTape(),()=>this.tidy("forward",t));w(o instanceof At,()=>"The result y returned by f() must be a tensor.");const i=vp(this.state.activeTape,e,o);if(!r&&i.length===0&&e.length>0)throw new Error("Cannot compute gradient of y=f(x) with respect to x. Make sure that the f you passed encloses all operations that lead from x to y.");return this.tidy("backward",()=>{const a={};a[o.id]=s??Dp(o.shape),Ip(a,i,u=>this.tidy(u),Rp);const l=e.map(u=>a[u.id]);return this.state.gradientDepth===0&&(this.state.activeTape.forEach(u=>{for(const c of u.saved)c.dispose()}),this.state.activeTape=null),{value:o,grads:l}})}customGrad(t){return w(lo(t),()=>"The f passed in customGrad(f) must be a function."),(...e)=>{w(e.every(a=>a instanceof At),()=>"The args passed in customGrad(f)(x1, x2,...) must all be tensors");let s;const r={};e.forEach((a,l)=>{r[l]=a});const o=(a,l)=>(s=t(...e,l),w(s.value instanceof At,()=>"The function f passed in customGrad(f) must return an object where `obj.value` is a tensor"),w(lo(s.gradFunc),()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function."),s.value),i=(a,l)=>{const u=s.gradFunc(a,l),c=Array.isArray(u)?u:[u];w(c.length===e.length,()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function that returns the same number of tensors as inputs passed to f(...)."),w(c.every(f=>f instanceof At),()=>"The function f passed in customGrad(f) must return an object where `obj.gradFunc` is a function that returns a list of only tensors.");const h={};return c.forEach((f,d)=>{h[d]=()=>f}),h};return this.runKernelFunc({forwardFunc:o,backwardsFunc:i,inputs:r})}}readSync(t){return this.state.tensorInfo.get(t).backend.readSync(t)}read(t){return this.state.tensorInfo.get(t).backend.read(t)}readToGPU(t,e){return this.state.tensorInfo.get(t).backend.readToGPU(t,e)}async time(t){const e=Nn(),s=await this.backend.time(t);return s.wallMs=Nn()-e,s}track(t){return this.state.activeScope!=null&&(t.scopeId=this.state.activeScope.id,this.state.activeScope.track.push(t)),t}get registeredVariables(){return this.state.registeredVariables}reset(){this.pendingBackendInitId++,this.state.dispose(),this.ENV.reset(),this.state=new al;for(const t in this.registry)this.disposeRegisteredKernels(t),this.registry[t].dispose(),delete this.registry[t];this.backendName=null,this.backendInstance=null,this.pendingBackendInit=null}}Rn.nextTensorId=0,Rn.nextVariableId=0;function Dp(n){const t=Sa(z(n),"float32");return A.makeTensor(t,n,"float32")}function ll(){const n=Ia();if(n._tfengine==null){const t=new Rf(n);n._tfengine=new Rn(t)}return Of(n._tfengine.ENV),Cp(()=>n._tfengine),n._tfengine}const A=ll();function Rp(n,t){const e={a:n,b:t};return A.runKernel(Aa,e)}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -257,7 +257,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Zs(n,t){let e=n;if(ne(n))return t==="string"?[]:[n.length];if(sl(n)){const r=n.channels||"RGBA";return[n.height,n.width*r.length]}else if(rl(n))return[n.buffer.size/(t==null?4:io(t))];if(!Array.isArray(n))return[];const s=[];for(;Array.isArray(e)||ne(e)&&t!=="string";)s.push(e.length),e=e[0];return Array.isArray(n)&&W().getBool("TENSORLIKE_CHECK_SHAPE_CONSISTENCY")&&ul(n,s,[]),s}function ul(n,t,e){if(e=e||[],!Array.isArray(n)&&!ne(n)){w(t.length===0,()=>`Element arr[${e.join("][")}] is a primitive, but should be an array/TypedArray of ${t[0]} elements`);return}w(t.length>0,()=>`Element arr[${e.join("][")}] should be a primitive, but is an array of ${n.length} elements`),w(n.length===t[0],()=>`Element arr[${e.join("][")}] should have ${t[0]} elements, but has ${n.length} elements`);const s=t.slice(1);for(let r=0;r<n.length;++r)ul(n[r],s,e.concat(r))}function cl(n,t,e,s){if(n!=="string_or_numeric"){if(n==null)throw new Error("Expected dtype cannot be null.");if(n!=="numeric"&&n!==t||n==="numeric"&&t==="string")throw new Error(`Argument '${e}' passed to '${s}' must be ${n} tensor, but got ${t} tensor`)}}function v(n,t,e,s="numeric"){if(n instanceof el())return cl(s,n.dtype,t,e),n;let r=is(n);if(r!=="string"&&["bool","int32","float32"].indexOf(s)>=0&&(r=s),cl(s,r,t,e),n==null||!ne(n)&&!Array.isArray(n)&&typeof n!="number"&&typeof n!="boolean"&&typeof n!="string"){const l=n==null?"null":n.constructor.name;throw new Error(`Argument '${t}' passed to '${e}' must be a Tensor or TensorLike, but got '${l}'`)}const o=Zs(n,r);!ne(n)&&!Array.isArray(n)&&(n=[n]);const a=r!=="string"?Hs(n,r):sn(n,[],!0);return A.makeTensor(a,o,r)}function hl(n,t,e,s="numeric"){if(!Array.isArray(n))throw new Error(`Argument ${t} passed to ${e} must be a \`Tensor[]\` or \`TensorLike[]\``);return n.map((o,i)=>v(o,`${t}[${i}]`,e,s))}/**
+ */function Qs(n,t){let e=n;if(ne(n))return t==="string"?[]:[n.length];if(sl(n)){const r=n.channels||"RGBA";return[n.height,n.width*r.length]}else if(rl(n))return[n.buffer.size/(t==null?4:io(t))];if(!Array.isArray(n))return[];const s=[];for(;Array.isArray(e)||ne(e)&&t!=="string";)s.push(e.length),e=e[0];return Array.isArray(n)&&W().getBool("TENSORLIKE_CHECK_SHAPE_CONSISTENCY")&&ul(n,s,[]),s}function ul(n,t,e){if(e=e||[],!Array.isArray(n)&&!ne(n)){w(t.length===0,()=>`Element arr[${e.join("][")}] is a primitive, but should be an array/TypedArray of ${t[0]} elements`);return}w(t.length>0,()=>`Element arr[${e.join("][")}] should be a primitive, but is an array of ${n.length} elements`),w(n.length===t[0],()=>`Element arr[${e.join("][")}] should have ${t[0]} elements, but has ${n.length} elements`);const s=t.slice(1);for(let r=0;r<n.length;++r)ul(n[r],s,e.concat(r))}function cl(n,t,e,s){if(n!=="string_or_numeric"){if(n==null)throw new Error("Expected dtype cannot be null.");if(n!=="numeric"&&n!==t||n==="numeric"&&t==="string")throw new Error(`Argument '${e}' passed to '${s}' must be ${n} tensor, but got ${t} tensor`)}}function v(n,t,e,s="numeric"){if(n instanceof el())return cl(s,n.dtype,t,e),n;let r=as(n);if(r!=="string"&&["bool","int32","float32"].indexOf(s)>=0&&(r=s),cl(s,r,t,e),n==null||!ne(n)&&!Array.isArray(n)&&typeof n!="number"&&typeof n!="boolean"&&typeof n!="string"){const l=n==null?"null":n.constructor.name;throw new Error(`Argument '${t}' passed to '${e}' must be a Tensor or TensorLike, but got '${l}'`)}const o=Qs(n,r);!ne(n)&&!Array.isArray(n)&&(n=[n]);const a=r!=="string"?Ks(n,r):rn(n,[],!0);return A.makeTensor(a,o,r)}function hl(n,t,e,s="numeric"){if(!Array.isArray(n))throw new Error(`Argument ${t} passed to ${e} must be a \`Tensor[]\` or \`TensorLike[]\``);return n.map((o,i)=>v(o,`${t}[${i}]`,e,s))}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -272,7 +272,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Qs(n,t,e,s){if(s==null)s=is(n);else if(s==="complex64")throw new Error("Cannot construct a complex64 tensor directly. Please use tf.complex(real, imag).");if(rl(n)||sl(n)){if(s!=="float32"&&s!=="int32")throw new Error(`Creating tensor from GPU data only supports 'float32'|'int32' dtype, while the dtype is ${s}.`);return A.backend.createTensorFromGPUData(n,t||e,s)}if(!ne(n)&&!Array.isArray(n)&&typeof n!="number"&&typeof n!="boolean"&&typeof n!="string")throw new Error("values passed to tensor(values) must be a number/boolean/string or an array of numbers/booleans/strings, or a TypedArray");if(t!=null){Re(t);const r=z(t),o=z(e);w(r===o,()=>`Based on the provided shape, [${t}], the tensor should have ${r} values but has ${o}`);for(let i=0;i<e.length;++i){const a=e[i],l=i===e.length-1?a!==z(t.slice(i)):!0;w(e[i]===t[i]||!l,()=>`Error creating a new Tensor. Inferred shape (${e}) does not match the provided shape (${t}). `)}}return!ne(n)&&!Array.isArray(n)&&(n=[n]),t=t||e,n=s!=="string"?Hs(n,s):sn(n,[],!0),A.makeTensor(n,t,s)}/**
+ */function tr(n,t,e,s){if(s==null)s=as(n);else if(s==="complex64")throw new Error("Cannot construct a complex64 tensor directly. Please use tf.complex(real, imag).");if(rl(n)||sl(n)){if(s!=="float32"&&s!=="int32")throw new Error(`Creating tensor from GPU data only supports 'float32'|'int32' dtype, while the dtype is ${s}.`);return A.backend.createTensorFromGPUData(n,t||e,s)}if(!ne(n)&&!Array.isArray(n)&&typeof n!="number"&&typeof n!="boolean"&&typeof n!="string")throw new Error("values passed to tensor(values) must be a number/boolean/string or an array of numbers/booleans/strings, or a TypedArray");if(t!=null){Re(t);const r=z(t),o=z(e);w(r===o,()=>`Based on the provided shape, [${t}], the tensor should have ${r} values but has ${o}`);for(let i=0;i<e.length;++i){const a=e[i],l=i===e.length-1?a!==z(t.slice(i)):!0;w(e[i]===t[i]||!l,()=>`Error creating a new Tensor. Inferred shape (${e}) does not match the provided shape (${t}). `)}}return!ne(n)&&!Array.isArray(n)&&(n=[n]),t=t||e,n=s!=="string"?Ks(n,s):rn(n,[],!0),A.makeTensor(n,t,s)}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -287,7 +287,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function tr(n,t,e){const s=Zs(n,e);return Qs(n,t,s,e)}/**
+ */function er(n,t,e){const s=Qs(n,e);return tr(n,t,s,e)}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -302,7 +302,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Rt(n,t){ya(n);const e=Zs(n,t);if(e.length!==1)throw new Error("tensor1d() requires values to be a flat/TypedArray");return Qs(n,null,e,t)}/**
+ */function Rt(n,t){ya(n);const e=Qs(n,t);if(e.length!==1)throw new Error("tensor1d() requires values to be a flat/TypedArray");return tr(n,null,e,t)}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -377,7 +377,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Wp(n,t,e){const s=v(n,"x","slice4d");return w(s.rank===4,()=>`slice4d expects a rank-4 tensor, but got a rank-${s.rank} tensor`),Et(s,t,e)}const cs=C({slice4d_:Wp});/**
+ */function Wp(n,t,e){const s=v(n,"x","slice4d");return w(s.rank===4,()=>`slice4d expects a rank-4 tensor, but got a rank-${s.rank} tensor`),Et(s,t,e)}const hs=C({slice4d_:Wp});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -392,7 +392,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Gp(n){const e={x:v(n,"x","clone","string_or_numeric")};return A.runKernel(go,e)}const rn=C({clone_:Gp});/**
+ */function Gp(n){const e={x:v(n,"x","clone","string_or_numeric")};return A.runKernel(go,e)}const on=C({clone_:Gp});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -408,7 +408,7 @@
  * limitations under the License.
  * =============================================================================
  */function Vp(n,t=0){w(n.length>=1,()=>"Pass at least one tensor to concat");const e=hl(n,"tensors","concat","string_or_numeric");if(e[0].dtype==="complex64"&&e.forEach(o=>{if(o.dtype!=="complex64")throw new Error(`Cannot concatenate complex64 tensors with a tensor
-          with dtype ${o.dtype}. `)}),e.length===1)return rn(e[0]);const s=e,r={axis:t};return A.runKernel(Ca,s,r)}const on=C({concat_:Vp});function qp(n,t){return on(n,t)}const jp=C({concat4d_:qp});/**
+          with dtype ${o.dtype}. `)}),e.length===1)return on(e[0]);const s=e,r={axis:t};return A.runKernel(Ca,s,r)}const an=C({concat_:Vp});function qp(n,t){return an(n,t)}const jp=C({concat4d_:qp});/**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -513,7 +513,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const Do="tensorflowjs",Ro=1,an="models_store",We="model_info_store";function bl(){if(!W().getBool("IS_BROWSER"))throw new Error("Failed to obtain IndexedDB factory because the current environmentis not a web browser.");const n=typeof window>"u"?self:window,t=n.indexedDB||n.mozIndexedDB||n.webkitIndexedDB||n.msIndexedDB||n.shimIndexedDB;if(t==null)throw new Error("The current browser does not appear to support IndexedDB.");return t}function Po(n){const t=n.result;t.createObjectStore(an,{keyPath:"modelPath"}),t.createObjectStore(We,{keyPath:"modelPath"})}class ln{constructor(t){if(this.indexedDB=bl(),t==null||!t)throw new Error("For IndexedDB, modelPath must not be null, undefined or empty.");this.modelPath=t}async save(t){if(t.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");return this.databaseAction(this.modelPath,t)}async load(){return this.databaseAction(this.modelPath)}databaseAction(t,e){return new Promise((s,r)=>{const o=this.indexedDB.open(Do,Ro);o.onupgradeneeded=()=>Po(o),o.onsuccess=()=>{const i=o.result;if(e==null){const a=i.transaction(an,"readonly"),u=a.objectStore(an).get(this.modelPath);u.onsuccess=()=>{if(u.result==null)return i.close(),r(new Error(`Cannot find model with path '${this.modelPath}' in IndexedDB.`));s(u.result.modelArtifacts)},u.onerror=c=>(i.close(),r(u.error)),a.oncomplete=()=>i.close()}else{e.weightData=Pn.join(e.weightData);const a=gl(e),l=i.transaction(We,"readwrite");let u=l.objectStore(We),c;try{c=u.put({modelPath:this.modelPath,modelArtifactsInfo:a})}catch(f){return r(f)}let h;c.onsuccess=()=>{h=i.transaction(an,"readwrite");const f=h.objectStore(an);let d;try{d=f.put({modelPath:this.modelPath,modelArtifacts:e,modelArtifactsInfo:a})}catch(p){return r(p)}d.onsuccess=()=>s({modelArtifactsInfo:a}),d.onerror=p=>{u=l.objectStore(We);const g=u.delete(this.modelPath);g.onsuccess=()=>(i.close(),r(d.error)),g.onerror=m=>(i.close(),r(d.error))}},c.onerror=f=>(i.close(),r(c.error)),l.oncomplete=()=>{h==null?i.close():h.oncomplete=()=>i.close()}}},o.onerror=i=>r(o.error)})}}ln.URL_SCHEME="indexeddb://";const yl=n=>W().getBool("IS_BROWSER")&&!Array.isArray(n)&&n.startsWith(ln.URL_SCHEME)?sm(n.slice(ln.URL_SCHEME.length)):null;_t.registerSaveRouter(yl),_t.registerLoadRouter(yl);function sm(n){return new ln(n)}function rm(n){return n.startsWith(ln.URL_SCHEME)?n.slice(ln.URL_SCHEME.length):n}class om{constructor(){this.indexedDB=bl()}async listModels(){return new Promise((t,e)=>{const s=this.indexedDB.open(Do,Ro);s.onupgradeneeded=()=>Po(s),s.onsuccess=()=>{const r=s.result,o=r.transaction(We,"readonly"),a=o.objectStore(We).getAll();a.onsuccess=()=>{const l={};for(const u of a.result)l[u.modelPath]=u.modelArtifactsInfo;t(l)},a.onerror=l=>(r.close(),e(a.error)),o.oncomplete=()=>r.close()},s.onerror=r=>e(s.error)})}async removeModel(t){return t=rm(t),new Promise((e,s)=>{const r=this.indexedDB.open(Do,Ro);r.onupgradeneeded=()=>Po(r),r.onsuccess=()=>{const o=r.result,i=o.transaction(We,"readwrite"),a=i.objectStore(We),l=a.get(t);let u;l.onsuccess=()=>{if(l.result==null)return o.close(),s(new Error(`Cannot find model with path '${t}' in IndexedDB.`));{const c=a.delete(t),h=()=>{u=o.transaction(an,"readwrite");const d=u.objectStore(an).delete(t);d.onsuccess=()=>e(l.result.modelArtifactsInfo),d.onerror=p=>s(l.error)};c.onsuccess=h,c.onerror=f=>(h(),o.close(),s(l.error))}},l.onerror=c=>(o.close(),s(l.error)),i.oncomplete=()=>{u==null?o.close():u.oncomplete=()=>o.close()}},r.onerror=o=>s(r.error)})}}/**
+ */const Do="tensorflowjs",Ro=1,ln="models_store",We="model_info_store";function bl(){if(!W().getBool("IS_BROWSER"))throw new Error("Failed to obtain IndexedDB factory because the current environmentis not a web browser.");const n=typeof window>"u"?self:window,t=n.indexedDB||n.mozIndexedDB||n.webkitIndexedDB||n.msIndexedDB||n.shimIndexedDB;if(t==null)throw new Error("The current browser does not appear to support IndexedDB.");return t}function Po(n){const t=n.result;t.createObjectStore(ln,{keyPath:"modelPath"}),t.createObjectStore(We,{keyPath:"modelPath"})}class un{constructor(t){if(this.indexedDB=bl(),t==null||!t)throw new Error("For IndexedDB, modelPath must not be null, undefined or empty.");this.modelPath=t}async save(t){if(t.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");return this.databaseAction(this.modelPath,t)}async load(){return this.databaseAction(this.modelPath)}databaseAction(t,e){return new Promise((s,r)=>{const o=this.indexedDB.open(Do,Ro);o.onupgradeneeded=()=>Po(o),o.onsuccess=()=>{const i=o.result;if(e==null){const a=i.transaction(ln,"readonly"),u=a.objectStore(ln).get(this.modelPath);u.onsuccess=()=>{if(u.result==null)return i.close(),r(new Error(`Cannot find model with path '${this.modelPath}' in IndexedDB.`));s(u.result.modelArtifacts)},u.onerror=c=>(i.close(),r(u.error)),a.oncomplete=()=>i.close()}else{e.weightData=Pn.join(e.weightData);const a=gl(e),l=i.transaction(We,"readwrite");let u=l.objectStore(We),c;try{c=u.put({modelPath:this.modelPath,modelArtifactsInfo:a})}catch(f){return r(f)}let h;c.onsuccess=()=>{h=i.transaction(ln,"readwrite");const f=h.objectStore(ln);let d;try{d=f.put({modelPath:this.modelPath,modelArtifacts:e,modelArtifactsInfo:a})}catch(p){return r(p)}d.onsuccess=()=>s({modelArtifactsInfo:a}),d.onerror=p=>{u=l.objectStore(We);const g=u.delete(this.modelPath);g.onsuccess=()=>(i.close(),r(d.error)),g.onerror=m=>(i.close(),r(d.error))}},c.onerror=f=>(i.close(),r(c.error)),l.oncomplete=()=>{h==null?i.close():h.oncomplete=()=>i.close()}}},o.onerror=i=>r(o.error)})}}un.URL_SCHEME="indexeddb://";const yl=n=>W().getBool("IS_BROWSER")&&!Array.isArray(n)&&n.startsWith(un.URL_SCHEME)?sm(n.slice(un.URL_SCHEME.length)):null;_t.registerSaveRouter(yl),_t.registerLoadRouter(yl);function sm(n){return new un(n)}function rm(n){return n.startsWith(un.URL_SCHEME)?n.slice(un.URL_SCHEME.length):n}class om{constructor(){this.indexedDB=bl()}async listModels(){return new Promise((t,e)=>{const s=this.indexedDB.open(Do,Ro);s.onupgradeneeded=()=>Po(s),s.onsuccess=()=>{const r=s.result,o=r.transaction(We,"readonly"),a=o.objectStore(We).getAll();a.onsuccess=()=>{const l={};for(const u of a.result)l[u.modelPath]=u.modelArtifactsInfo;t(l)},a.onerror=l=>(r.close(),e(a.error)),o.oncomplete=()=>r.close()},s.onerror=r=>e(s.error)})}async removeModel(t){return t=rm(t),new Promise((e,s)=>{const r=this.indexedDB.open(Do,Ro);r.onupgradeneeded=()=>Po(r),r.onsuccess=()=>{const o=r.result,i=o.transaction(We,"readwrite"),a=i.objectStore(We),l=a.get(t);let u;l.onsuccess=()=>{if(l.result==null)return o.close(),s(new Error(`Cannot find model with path '${t}' in IndexedDB.`));{const c=a.delete(t),h=()=>{u=o.transaction(ln,"readwrite");const d=u.objectStore(ln).delete(t);d.onsuccess=()=>e(l.result.modelArtifactsInfo),d.onerror=p=>s(l.error)};c.onsuccess=h,c.onerror=f=>(h(),o.close(),s(l.error))}},l.onerror=c=>(o.close(),s(l.error)),i.oncomplete=()=>{u==null?o.close():u.oncomplete=()=>o.close()}},r.onerror=o=>s(r.error)})}}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -528,7 +528,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const Pe="/",Mn="tensorflowjs_models",wl="info",im="model_topology",am="weight_specs",lm="weight_data",um="model_metadata";function xl(n){return{info:[Mn,n,wl].join(Pe),topology:[Mn,n,im].join(Pe),weightSpecs:[Mn,n,am].join(Pe),weightData:[Mn,n,lm].join(Pe),modelMetadata:[Mn,n,um].join(Pe)}}function Sl(n){for(const t of Object.values(n))window.localStorage.removeItem(t)}function cm(n){const t=n.split(Pe);if(t.length<3)throw new Error(`Invalid key format: ${n}`);return t.slice(1,t.length-1).join(Pe)}function hm(n){return n.startsWith(un.URL_SCHEME)?n.slice(un.URL_SCHEME.length):n}class un{constructor(t){if(!W().getBool("IS_BROWSER")||typeof window>"u"||typeof window.localStorage>"u")throw new Error("The current environment does not support local storage.");if(this.LS=window.localStorage,t==null||!t)throw new Error("For local storage, modelPath must not be null, undefined or empty.");this.modelPath=t,this.keys=xl(this.modelPath)}async save(t){if(t.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");{const e=JSON.stringify(t.modelTopology),s=JSON.stringify(t.weightSpecs),r=gl(t),o=Pn.join(t.weightData);try{this.LS.setItem(this.keys.info,JSON.stringify(r)),this.LS.setItem(this.keys.topology,e),this.LS.setItem(this.keys.weightSpecs,s),this.LS.setItem(this.keys.weightData,Qp(o));const i={format:t.format,generatedBy:t.generatedBy,convertedBy:t.convertedBy,signature:t.signature!=null?t.signature:void 0,userDefinedMetadata:t.userDefinedMetadata!=null?t.userDefinedMetadata:void 0,modelInitializer:t.modelInitializer!=null?t.modelInitializer:void 0,initializerSignature:t.initializerSignature!=null?t.initializerSignature:void 0,trainingConfig:t.trainingConfig!=null?t.trainingConfig:void 0};return this.LS.setItem(this.keys.modelMetadata,JSON.stringify(i)),{modelArtifactsInfo:r}}catch{throw Sl(this.keys),new Error(`Failed to save model '${this.modelPath}' to local storage: size quota being exceeded is a possible cause of this failure: modelTopologyBytes=${r.modelTopologyBytes}, weightSpecsBytes=${r.weightSpecsBytes}, weightDataBytes=${r.weightDataBytes}.`)}}}async load(){const t=JSON.parse(this.LS.getItem(this.keys.info));if(t==null)throw new Error(`In local storage, there is no model with name '${this.modelPath}'`);if(t.modelTopologyType!=="JSON")throw new Error("BrowserLocalStorage does not support loading non-JSON model topology yet.");const e={},s=JSON.parse(this.LS.getItem(this.keys.topology));if(s==null)throw new Error(`In local storage, the topology of model '${this.modelPath}' is missing.`);e.modelTopology=s;const r=JSON.parse(this.LS.getItem(this.keys.weightSpecs));if(r==null)throw new Error(`In local storage, the weight specs of model '${this.modelPath}' are missing.`);e.weightSpecs=r;const o=this.LS.getItem(this.keys.modelMetadata);if(o!=null){const a=JSON.parse(o);e.format=a.format,e.generatedBy=a.generatedBy,e.convertedBy=a.convertedBy,a.signature!=null&&(e.signature=a.signature),a.userDefinedMetadata!=null&&(e.userDefinedMetadata=a.userDefinedMetadata),a.modelInitializer!=null&&(e.modelInitializer=a.modelInitializer),a.initializerSignature!=null&&(e.initializerSignature=a.initializerSignature),a.trainingConfig!=null&&(e.trainingConfig=a.trainingConfig)}const i=this.LS.getItem(this.keys.weightData);if(i==null)throw new Error(`In local storage, the binary weight values of model '${this.modelPath}' are missing.`);return e.weightData=tm(i),e}}un.URL_SCHEME="localstorage://";const $l=n=>W().getBool("IS_BROWSER")&&!Array.isArray(n)&&n.startsWith(un.URL_SCHEME)?fm(n.slice(un.URL_SCHEME.length)):null;_t.registerSaveRouter($l),_t.registerLoadRouter($l);function fm(n){return new un(n)}class dm{constructor(){w(W().getBool("IS_BROWSER"),()=>"Current environment is not a web browser"),w(typeof window>"u"||typeof window.localStorage<"u",()=>"Current browser does not appear to support localStorage"),this.LS=window.localStorage}async listModels(){const t={},e=Mn+Pe,s=Pe+wl;for(let r=0;r<this.LS.length;++r){const o=this.LS.key(r);if(o.startsWith(e)&&o.endsWith(s)){const i=cm(o);t[i]=JSON.parse(this.LS.getItem(o))}}return t}async removeModel(t){t=hm(t);const e=xl(t);if(this.LS.getItem(e.info)==null)throw new Error(`Cannot find model at path '${t}'`);const s=JSON.parse(this.LS.getItem(e.info));return Sl(e),s}}/**
+ */const Pe="/",Mn="tensorflowjs_models",wl="info",im="model_topology",am="weight_specs",lm="weight_data",um="model_metadata";function xl(n){return{info:[Mn,n,wl].join(Pe),topology:[Mn,n,im].join(Pe),weightSpecs:[Mn,n,am].join(Pe),weightData:[Mn,n,lm].join(Pe),modelMetadata:[Mn,n,um].join(Pe)}}function Sl(n){for(const t of Object.values(n))window.localStorage.removeItem(t)}function cm(n){const t=n.split(Pe);if(t.length<3)throw new Error(`Invalid key format: ${n}`);return t.slice(1,t.length-1).join(Pe)}function hm(n){return n.startsWith(cn.URL_SCHEME)?n.slice(cn.URL_SCHEME.length):n}class cn{constructor(t){if(!W().getBool("IS_BROWSER")||typeof window>"u"||typeof window.localStorage>"u")throw new Error("The current environment does not support local storage.");if(this.LS=window.localStorage,t==null||!t)throw new Error("For local storage, modelPath must not be null, undefined or empty.");this.modelPath=t,this.keys=xl(this.modelPath)}async save(t){if(t.modelTopology instanceof ArrayBuffer)throw new Error("BrowserLocalStorage.save() does not support saving model topology in binary formats yet.");{const e=JSON.stringify(t.modelTopology),s=JSON.stringify(t.weightSpecs),r=gl(t),o=Pn.join(t.weightData);try{this.LS.setItem(this.keys.info,JSON.stringify(r)),this.LS.setItem(this.keys.topology,e),this.LS.setItem(this.keys.weightSpecs,s),this.LS.setItem(this.keys.weightData,Qp(o));const i={format:t.format,generatedBy:t.generatedBy,convertedBy:t.convertedBy,signature:t.signature!=null?t.signature:void 0,userDefinedMetadata:t.userDefinedMetadata!=null?t.userDefinedMetadata:void 0,modelInitializer:t.modelInitializer!=null?t.modelInitializer:void 0,initializerSignature:t.initializerSignature!=null?t.initializerSignature:void 0,trainingConfig:t.trainingConfig!=null?t.trainingConfig:void 0};return this.LS.setItem(this.keys.modelMetadata,JSON.stringify(i)),{modelArtifactsInfo:r}}catch{throw Sl(this.keys),new Error(`Failed to save model '${this.modelPath}' to local storage: size quota being exceeded is a possible cause of this failure: modelTopologyBytes=${r.modelTopologyBytes}, weightSpecsBytes=${r.weightSpecsBytes}, weightDataBytes=${r.weightDataBytes}.`)}}}async load(){const t=JSON.parse(this.LS.getItem(this.keys.info));if(t==null)throw new Error(`In local storage, there is no model with name '${this.modelPath}'`);if(t.modelTopologyType!=="JSON")throw new Error("BrowserLocalStorage does not support loading non-JSON model topology yet.");const e={},s=JSON.parse(this.LS.getItem(this.keys.topology));if(s==null)throw new Error(`In local storage, the topology of model '${this.modelPath}' is missing.`);e.modelTopology=s;const r=JSON.parse(this.LS.getItem(this.keys.weightSpecs));if(r==null)throw new Error(`In local storage, the weight specs of model '${this.modelPath}' are missing.`);e.weightSpecs=r;const o=this.LS.getItem(this.keys.modelMetadata);if(o!=null){const a=JSON.parse(o);e.format=a.format,e.generatedBy=a.generatedBy,e.convertedBy=a.convertedBy,a.signature!=null&&(e.signature=a.signature),a.userDefinedMetadata!=null&&(e.userDefinedMetadata=a.userDefinedMetadata),a.modelInitializer!=null&&(e.modelInitializer=a.modelInitializer),a.initializerSignature!=null&&(e.initializerSignature=a.initializerSignature),a.trainingConfig!=null&&(e.trainingConfig=a.trainingConfig)}const i=this.LS.getItem(this.keys.weightData);if(i==null)throw new Error(`In local storage, the binary weight values of model '${this.modelPath}' are missing.`);return e.weightData=tm(i),e}}cn.URL_SCHEME="localstorage://";const $l=n=>W().getBool("IS_BROWSER")&&!Array.isArray(n)&&n.startsWith(cn.URL_SCHEME)?fm(n.slice(cn.URL_SCHEME.length)):null;_t.registerSaveRouter($l),_t.registerLoadRouter($l);function fm(n){return new cn(n)}class dm{constructor(){w(W().getBool("IS_BROWSER"),()=>"Current environment is not a web browser"),w(typeof window>"u"||typeof window.localStorage<"u",()=>"Current browser does not appear to support localStorage"),this.LS=window.localStorage}async listModels(){const t={},e=Mn+Pe,s=Pe+wl;for(let r=0;r<this.LS.length;++r){const o=this.LS.key(r);if(o.startsWith(e)&&o.endsWith(s)){const i=cm(o);t[i]=JSON.parse(this.LS.getItem(o))}}return t}async removeModel(t){t=hm(t);const e=xl(t);if(this.LS.getItem(e.info)==null)throw new Error(`Cannot find model at path '${t}'`);const s=JSON.parse(this.LS.getItem(e.info));return Sl(e),s}}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -558,7 +558,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class pm{constructor(){this.messageName="setTimeoutCustom",this.functionRefs=[],this.handledMessageCount=0,this.hasEventListener=!1}fetch(t,e){return fetch(t,e)}now(){return performance.now()}encode(t,e){if(e!=="utf-8"&&e!=="utf8")throw new Error(`Browser's encoder only supports utf-8, but got ${e}`);return this.textEncoder==null&&(this.textEncoder=new TextEncoder),this.textEncoder.encode(t)}decode(t,e){return new TextDecoder(e).decode(t)}setTimeoutCustom(t,e){if(typeof window>"u"||!W().getBool("USE_SETTIMEOUTCUSTOM")){setTimeout(t,e);return}this.functionRefs.push(t),setTimeout(()=>{window.postMessage({name:this.messageName,index:this.functionRefs.length-1},"*")},e),this.hasEventListener||(this.hasEventListener=!0,window.addEventListener("message",s=>{if(s.source===window&&s.data.name===this.messageName){s.stopPropagation();const r=this.functionRefs[s.data.index];r(),this.handledMessageCount++,this.handledMessageCount===this.functionRefs.length&&(this.functionRefs=[],this.handledMessageCount=0)}},!0))}isTypedArray(t){return Ba(t)}}if(W().get("IS_BROWSER")){W().setPlatform("browser",new pm);try{we.registerManager(un.URL_SCHEME,new dm)}catch{}try{we.registerManager(ln.URL_SCHEME,new om)}catch{}}/**
+ */class pm{constructor(){this.messageName="setTimeoutCustom",this.functionRefs=[],this.handledMessageCount=0,this.hasEventListener=!1}fetch(t,e){return fetch(t,e)}now(){return performance.now()}encode(t,e){if(e!=="utf-8"&&e!=="utf8")throw new Error(`Browser's encoder only supports utf-8, but got ${e}`);return this.textEncoder==null&&(this.textEncoder=new TextEncoder),this.textEncoder.encode(t)}decode(t,e){return new TextDecoder(e).decode(t)}setTimeoutCustom(t,e){if(typeof window>"u"||!W().getBool("USE_SETTIMEOUTCUSTOM")){setTimeout(t,e);return}this.functionRefs.push(t),setTimeout(()=>{window.postMessage({name:this.messageName,index:this.functionRefs.length-1},"*")},e),this.hasEventListener||(this.hasEventListener=!0,window.addEventListener("message",s=>{if(s.source===window&&s.data.name===this.messageName){s.stopPropagation();const r=this.functionRefs[s.data.index];r(),this.handledMessageCount++,this.handledMessageCount===this.functionRefs.length&&(this.functionRefs=[],this.handledMessageCount=0)}},!0))}isTypedArray(t){return Ba(t)}}if(W().get("IS_BROWSER")){W().setPlatform("browser",new pm);try{we.registerManager(cn.URL_SCHEME,new dm)}catch{}try{we.registerManager(un.URL_SCHEME,new om)}catch{}}/**
  * @license
  * Copyright 2019 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -588,7 +588,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function xt(n,t="float32",e){return t=t||"float32",Re(n),new Xs(n,t,e)}/**
+ */function xt(n,t="float32",e){return t=t||"float32",Re(n),new Js(n,t,e)}/**
  * @license
  * Copyright 2020 Google Inc. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -633,7 +633,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */ll(),kp({buffer:xt,cast:ot,clone:rn,print:ym});/**
+ */ll(),kp({buffer:xt,cast:ot,clone:on,print:ym});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -738,7 +738,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Cm(n,t=0){const s={x:v(n,"x","argMax")},r={axis:t};return A.runKernel(Uf,s,r)}const er=C({argMax_:Cm});/**
+ */function Cm(n,t=0){const s={x:v(n,"x","argMax")},r={axis:t};return A.runKernel(Uf,s,r)}const nr=C({argMax_:Cm});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -753,7 +753,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function km(n,t,e,s,r,o,i="channelsLast"){const[a,l]=hs(t);let u;if(i==="channelsLast")u=[a,l,n[3],n[3]];else if(i==="channelsFirst")u=[a,l,n[1],n[1]];else throw new Error(`Unknown dataFormat ${i}`);return Mo(n,u,e,s,r,o,!1,i)}function Mo(n,t,e,s,r,o,i=!1,a="channelsLast"){let[l,u,c,h]=[-1,-1,-1,-1];if(a==="channelsLast")[l,u,c,h]=n;else if(a==="channelsFirst")[l,h,u,c]=n;else throw new Error(`Unknown dataFormat ${a}`);const[f,d,,p]=t,[g,m]=hs(e),[b,y]=hs(s),S=Oo(f,b),x=Oo(d,y),{padInfo:$,outHeight:E,outWidth:D}=Nm(r,u,c,g,m,S,x,o,a),_=i?p*h:p;let T;return a==="channelsFirst"?T=[l,_,E,D]:a==="channelsLast"&&(T=[l,E,D,_]),{batchSize:l,dataFormat:a,inHeight:u,inWidth:c,inChannels:h,outHeight:E,outWidth:D,outChannels:_,padInfo:$,strideHeight:g,strideWidth:m,filterHeight:f,filterWidth:d,effectiveFilterHeight:S,effectiveFilterWidth:x,dilationHeight:b,dilationWidth:y,inShape:n,outShape:T,filterShape:t}}function _m(n,t,e,s,r){s==null&&(s=Tm(n,t,e));const o=n[0],i=n[1],a=nr((o-t+2*s)/e+1,r),l=nr((i-t+2*s)/e+1,r);return[a,l]}function Tm(n,t,e,s=1){const r=Oo(t,s);return Math.floor((n[0]*(e-1)-e+r)/2)}function hs(n){return typeof n=="number"?[n,n,n]:n.length===2?[n[0],n[1],1]:n}function Oo(n,t){return t<=1?n:n+(n-1)*(t-1)}function Nm(n,t,e,s,r,o,i,a,l){let u,c,h;if(typeof n=="number"){u={top:n,bottom:n,left:n,right:n,type:n===0?"VALID":"NUMBER"};const d=_m([t,e],o,s,n,a);c=d[0],h=d[1]}else if(n==="same"){c=Math.ceil(t/s),h=Math.ceil(e/r);const f=Math.max(0,(c-1)*s+o-t),d=Math.max(0,(h-1)*r+i-e),p=Math.floor(f/2),g=f-p,m=Math.floor(d/2),b=d-m;u={top:p,bottom:g,left:m,right:b,type:"SAME"}}else if(n==="valid")u={top:0,bottom:0,left:0,right:0,type:"VALID"},c=Math.ceil((t-o+1)/s),h=Math.ceil((e-i+1)/r);else if(typeof n=="object"){const f=l==="channelsLast"?n[1][0]:n[2][0],d=l==="channelsLast"?n[1][1]:n[2][1],p=l==="channelsLast"?n[2][0]:n[3][0],g=l==="channelsLast"?n[2][1]:n[3][1];u={top:f,bottom:d,left:p,right:g,type:f===0&&d===0&&p===0&&g===0?"VALID":"EXPLICIT"},c=nr((t-o+f+d)/s+1,a),h=nr((e-i+p+g)/r+1,a)}else throw Error(`Unknown padding parameter: ${n}`);return{padInfo:u,outHeight:c,outWidth:h}}function nr(n,t){if(!t)return Math.trunc(n);switch(t){case"round":return Math.round(n);case"ceil":return Math.ceil(n);case"floor":return Math.floor(n);default:throw new Error(`Unknown roundingMode ${t}`)}}function Bo(n){const[t,e,s]=hs(n);return t===1&&e===1&&s===1}function On(n,t){return Bo(n)||Bo(t)}function Bn(n){return hs(n).every(t=>t>0)}function Dm(n){if(n==="NHWC")return"channelsLast";if(n==="NCHW")return"channelsFirst";throw new Error(`Unknown dataFormat ${n}`)}function xe(n,t,e){if(e!=null){if(typeof t=="string")throw Error(`Error in ${n}: pad must be an integer when using dimRoundingMode ${e} but got pad ${t}.`);if(typeof t=="number")w(oo(t),()=>`Error in ${n}: pad must be an integer when using dimRoundingMode ${e} but got pad ${t}.`);else if(typeof t=="object")t.forEach(s=>{s.forEach(r=>{w(oo(r),()=>`Error in ${n}: pad must be an integer when using dimRoundingMode ${e} but got pad ${r}.`)})});else throw Error(`Error in ${n}: Unknown padding parameter: ${t}`)}}/**
+ */function km(n,t,e,s,r,o,i="channelsLast"){const[a,l]=fs(t);let u;if(i==="channelsLast")u=[a,l,n[3],n[3]];else if(i==="channelsFirst")u=[a,l,n[1],n[1]];else throw new Error(`Unknown dataFormat ${i}`);return Mo(n,u,e,s,r,o,!1,i)}function Mo(n,t,e,s,r,o,i=!1,a="channelsLast"){let[l,u,c,h]=[-1,-1,-1,-1];if(a==="channelsLast")[l,u,c,h]=n;else if(a==="channelsFirst")[l,h,u,c]=n;else throw new Error(`Unknown dataFormat ${a}`);const[f,d,,p]=t,[g,m]=fs(e),[b,y]=fs(s),S=Oo(f,b),x=Oo(d,y),{padInfo:$,outHeight:E,outWidth:D}=Nm(r,u,c,g,m,S,x,o,a),_=i?p*h:p;let T;return a==="channelsFirst"?T=[l,_,E,D]:a==="channelsLast"&&(T=[l,E,D,_]),{batchSize:l,dataFormat:a,inHeight:u,inWidth:c,inChannels:h,outHeight:E,outWidth:D,outChannels:_,padInfo:$,strideHeight:g,strideWidth:m,filterHeight:f,filterWidth:d,effectiveFilterHeight:S,effectiveFilterWidth:x,dilationHeight:b,dilationWidth:y,inShape:n,outShape:T,filterShape:t}}function _m(n,t,e,s,r){s==null&&(s=Tm(n,t,e));const o=n[0],i=n[1],a=sr((o-t+2*s)/e+1,r),l=sr((i-t+2*s)/e+1,r);return[a,l]}function Tm(n,t,e,s=1){const r=Oo(t,s);return Math.floor((n[0]*(e-1)-e+r)/2)}function fs(n){return typeof n=="number"?[n,n,n]:n.length===2?[n[0],n[1],1]:n}function Oo(n,t){return t<=1?n:n+(n-1)*(t-1)}function Nm(n,t,e,s,r,o,i,a,l){let u,c,h;if(typeof n=="number"){u={top:n,bottom:n,left:n,right:n,type:n===0?"VALID":"NUMBER"};const d=_m([t,e],o,s,n,a);c=d[0],h=d[1]}else if(n==="same"){c=Math.ceil(t/s),h=Math.ceil(e/r);const f=Math.max(0,(c-1)*s+o-t),d=Math.max(0,(h-1)*r+i-e),p=Math.floor(f/2),g=f-p,m=Math.floor(d/2),b=d-m;u={top:p,bottom:g,left:m,right:b,type:"SAME"}}else if(n==="valid")u={top:0,bottom:0,left:0,right:0,type:"VALID"},c=Math.ceil((t-o+1)/s),h=Math.ceil((e-i+1)/r);else if(typeof n=="object"){const f=l==="channelsLast"?n[1][0]:n[2][0],d=l==="channelsLast"?n[1][1]:n[2][1],p=l==="channelsLast"?n[2][0]:n[3][0],g=l==="channelsLast"?n[2][1]:n[3][1];u={top:f,bottom:d,left:p,right:g,type:f===0&&d===0&&p===0&&g===0?"VALID":"EXPLICIT"},c=sr((t-o+f+d)/s+1,a),h=sr((e-i+p+g)/r+1,a)}else throw Error(`Unknown padding parameter: ${n}`);return{padInfo:u,outHeight:c,outWidth:h}}function sr(n,t){if(!t)return Math.trunc(n);switch(t){case"round":return Math.round(n);case"ceil":return Math.ceil(n);case"floor":return Math.floor(n);default:throw new Error(`Unknown roundingMode ${t}`)}}function Bo(n){const[t,e,s]=fs(n);return t===1&&e===1&&s===1}function On(n,t){return Bo(n)||Bo(t)}function Bn(n){return fs(n).every(t=>t>0)}function Dm(n){if(n==="NHWC")return"channelsLast";if(n==="NCHW")return"channelsFirst";throw new Error(`Unknown dataFormat ${n}`)}function xe(n,t,e){if(e!=null){if(typeof t=="string")throw Error(`Error in ${n}: pad must be an integer when using dimRoundingMode ${e} but got pad ${t}.`);if(typeof t=="number")w(oo(t),()=>`Error in ${n}: pad must be an integer when using dimRoundingMode ${e} but got pad ${t}.`);else if(typeof t=="object")t.forEach(s=>{s.forEach(r=>{w(oo(r),()=>`Error in ${n}: pad must be an integer when using dimRoundingMode ${e} but got pad ${r}.`)})});else throw Error(`Error in ${n}: Unknown padding parameter: ${t}`)}}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -873,7 +873,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Gm(n,t){let e=v(n,"broadcastTo","x");const s=e.shape;if(Re(t),t.length<e.rank)throw new Error(`broadcastTo(): shape.length=${t.length} < input.rank=${e.rank}.`);if(t.length>e.rank){const u=e.shape.slice();for(;u.length<t.length;)u.unshift(1);e=L(e,u)}const r=e.shape,o=Array.from(t);for(let u=t.length-1;u>=0;u--)if(r[u]===t[u])o[u]=1;else if(e.shape[u]!==1)throw new Error(`broadcastTo(): [${s}] cannot be broadcast to [${t}].`);if(o.map((u,c)=>u>1?c:-1).filter(u=>u>=0).length===0)return rn(e);const a={x:e},l={reps:o};return A.runKernel(Ra,a,l)}const sr=C({broadcastTo_:Gm});/**
+ */function Gm(n,t){let e=v(n,"broadcastTo","x");const s=e.shape;if(Re(t),t.length<e.rank)throw new Error(`broadcastTo(): shape.length=${t.length} < input.rank=${e.rank}.`);if(t.length>e.rank){const u=e.shape.slice();for(;u.length<t.length;)u.unshift(1);e=L(e,u)}const r=e.shape,o=Array.from(t);for(let u=t.length-1;u>=0;u--)if(r[u]===t[u])o[u]=1;else if(e.shape[u]!==1)throw new Error(`broadcastTo(): [${s}] cannot be broadcast to [${t}].`);if(o.map((u,c)=>u>1?c:-1).filter(u=>u>=0).length===0)return on(e);const a={x:e},l={reps:o};return A.runKernel(Ra,a,l)}const rr=C({broadcastTo_:Gm});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -888,7 +888,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function rr(n,t,e){Re(n),e=e||is(t);const s={shape:n,value:t,dtype:e};return A.runKernel(ud,{},s)}/**
+ */function or(n,t,e){Re(n),e=e||as(t);const s={shape:n,value:t,dtype:e};return A.runKernel(ud,{},s)}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -903,7 +903,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Vm(n,t,e){const s=v(n,"x","clipByValue");if(w(t<=e,()=>`Error in clip: min (${t}) must be less than or equal to max (${e}).`),t===e)return rr(s.shape,t,s.dtype);const r={x:s},o={clipValueMin:t,clipValueMax:e};return A.runKernel(jf,r,o)}const he=C({clipByValue_:Vm});/**
+ */function Vm(n,t,e){const s=v(n,"x","clipByValue");if(w(t<=e,()=>`Error in clip: min (${t}) must be less than or equal to max (${e}).`),t===e)return or(s.shape,t,s.dtype);const r={x:s},o={clipValueMin:t,clipValueMax:e};return A.runKernel(jf,r,o)}const he=C({clipByValue_:Vm});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -993,7 +993,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function or(n,t){const e=n.length,s=[];for(let r=0;r<e;r++){const o=e-1-r,i=n[o]||1;(t[t.length-1-r]||1)>1&&i===1&&s.unshift(o)}return s}function og(n,t){const e=[];for(let s=0;s<t.length;s++){const r=n[n.length-s-1],o=t.length-s-1,i=t[o];(r==null||r===1&&i>1)&&e.unshift(o)}return e}function zt(n,t){const e=Math.max(n.length,t.length),s=new Array(e);for(let r=0;r<e;r++){let o=n[n.length-r-1];o==null&&(o=1);let i=t[t.length-r-1];if(i==null&&(i=1),o===1)s[e-r-1]=i;else if(i===1)s[e-r-1]=o;else if(o!==i){const a=`Operands could not be broadcast together with shapes ${n} and ${t}.`;throw Error(a)}else s[e-r-1]=o}return s}/**
+ */function ir(n,t){const e=n.length,s=[];for(let r=0;r<e;r++){const o=e-1-r,i=n[o]||1;(t[t.length-1-r]||1)>1&&i===1&&s.unshift(o)}return s}function og(n,t){const e=[];for(let s=0;s<t.length;s++){const r=n[n.length-s-1],o=t.length-s-1,i=t[o];(r==null||r===1&&i>1)&&e.unshift(o)}return e}function zt(n,t){const e=Math.max(n.length,t.length),s=new Array(e);for(let r=0;r<e;r++){let o=n[n.length-r-1];o==null&&(o=1);let i=t[t.length-r-1];if(i==null&&(i=1),o===1)s[e-r-1]=i;else if(i===1)s[e-r-1]=o;else if(o!==i){const a=`Operands could not be broadcast together with shapes ${n} and ${t}.`;throw Error(a)}else s[e-r-1]=o}return s}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1008,7 +1008,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function ig(n,t){let e=v(n,"a","equal","string_or_numeric"),s=v(t,"b","equal","string_or_numeric");[e,s]=Dt(e,s),zt(e.shape,s.shape);const r={a:e,b:s};return A.runKernel(id,r)}const cn=C({equal_:ig});/**
+ */function ig(n,t){let e=v(n,"a","equal","string_or_numeric"),s=v(t,"b","equal","string_or_numeric");[e,s]=Dt(e,s),zt(e.shape,s.shape);const r={a:e,b:s};return A.runKernel(id,r)}const hn=C({equal_:ig});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1023,7 +1023,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function ag(n,t,e){const s=v(t,"a","where"),r=v(e,"b","where"),o=v(n,"condition","where","bool"),i=zt(zt(o.shape,s.shape),r.shape),a=sr(o,i),l=sr(s,i),u=sr(r,i),c={condition:a,t:l,e:u};return A.runKernel(jd,c)}const hn=C({where_:ag});/**
+ */function ag(n,t,e){const s=v(t,"a","where"),r=v(e,"b","where"),o=v(n,"condition","where","bool"),i=zt(zt(o.shape,s.shape),r.shape),a=rr(o,i),l=rr(s,i),u=rr(r,i),c={condition:a,t:l,e:u};return A.runKernel(jd,c)}const fn=C({where_:ag});/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1053,7 +1053,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function ug(n,...t){const e=t.map((r,o)=>v(r,`tensors${o}`,"einsum")),s={equation:n};return A.runKernel(sd,e,s)}const fs=C({einsum_:ug});/**
+ */function ug(n,...t){const e=t.map((r,o)=>v(r,`tensors${o}`,"einsum")),s={equation:n};return A.runKernel(sd,e,s)}const ds=C({einsum_:ug});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1143,7 +1143,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function wg(n,t){let e=v(n,"base","pow"),s=v(t,"exp","pow");[e,s]=Dt(e,s);const r={a:e,b:s};return A.runKernel(Od,r)}const ir=C({pow_:wg});/**
+ */function wg(n,t){let e=v(n,"base","pow"),s=v(t,"exp","pow");[e,s]=Dt(e,s);const r={a:e,b:s};return A.runKernel(Od,r)}const ar=C({pow_:wg});/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1158,7 +1158,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Ht(n,t){if((ne(n)&&t!=="string"||Array.isArray(n))&&t!=="complex64")throw new Error("Error creating a new Scalar: value must be a primitive (number|boolean|string)");if(t==="string"&&ne(n)&&!(n instanceof Uint8Array))throw new Error("When making a scalar from encoded string, the value must be `Uint8Array`.");return Qs(n,[],[],t)}/**
+ */function Yt(n,t){if((ne(n)&&t!=="string"||Array.isArray(n))&&t!=="complex64")throw new Error("Error creating a new Scalar: value must be a primitive (number|boolean|string)");if(t==="string"&&ne(n)&&!(n instanceof Uint8Array))throw new Error("When making a scalar from encoded string, the value must be `Uint8Array`.");return tr(n,[],[],t)}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1218,7 +1218,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function vg(n,t="euclidean",e=null,s=!1){n=v(n,"x","norm");const r=_l(n,t,e);let o=r.shape;if(s){const i=os(e,n.shape);o=Cl(r.shape,i)}return L(r,o)}function _l(n,t,e=null){if(n.rank===0)return Lt(n);if(n.rank!==1&&e===null)return _l(L(n,[-1]),t,e);if(n.rank===1||typeof e=="number"||Array.isArray(e)&&e.length===1){if(t===1)return et(Lt(n),e);if(t===1/0)return Ge(Lt(n),e);if(t===-1/0)return kl(Lt(n),e);if(t==="euclidean"||t===2)return fe(et(ir(Lt(n),Ht(2,"int32")),e));throw new Error(`Error in norm: invalid ord value: ${t}`)}if(Array.isArray(e)&&e.length===2){if(t===1)return Ge(et(Lt(n),e[0]),e[1]-1);if(t===1/0)return Ge(et(Lt(n),e[1]),e[0]);if(t===-1/0)return kl(et(Lt(n),e[1]),e[0]);if(t==="fro"||t==="euclidean")return fe(et(Ve(n),e));throw new Error(`Error in norm: invalid ord value: ${t}`)}throw new Error(`Error in norm: invalid axis: ${e}`)}const Tl=C({norm_:vg});/**
+ */function vg(n,t="euclidean",e=null,s=!1){n=v(n,"x","norm");const r=_l(n,t,e);let o=r.shape;if(s){const i=is(e,n.shape);o=Cl(r.shape,i)}return L(r,o)}function _l(n,t,e=null){if(n.rank===0)return Lt(n);if(n.rank!==1&&e===null)return _l(L(n,[-1]),t,e);if(n.rank===1||typeof e=="number"||Array.isArray(e)&&e.length===1){if(t===1)return et(Lt(n),e);if(t===1/0)return Ge(Lt(n),e);if(t===-1/0)return kl(Lt(n),e);if(t==="euclidean"||t===2)return fe(et(ar(Lt(n),Yt(2,"int32")),e));throw new Error(`Error in norm: invalid ord value: ${t}`)}if(Array.isArray(e)&&e.length===2){if(t===1)return Ge(et(Lt(n),e[0]),e[1]-1);if(t===1/0)return Ge(et(Lt(n),e[1]),e[0]);if(t===-1/0)return kl(et(Lt(n),e[1]),e[0]);if(t==="fro"||t==="euclidean")return fe(et(Ve(n),e));throw new Error(`Error in norm: invalid ord value: ${t}`)}throw new Error(`Error in norm: invalid axis: ${e}`)}const Tl=C({norm_:vg});/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1263,7 +1263,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Eg(n,t){const e=v(n,"x","tile","string_or_numeric");w(e.rank===t.length,()=>`Error in transpose: rank of input ${e.rank} must match length of reps ${t}.`);const s={x:e},r={reps:t};return A.runKernel(Ra,s,r)}const ar=C({tile_:Eg});/**
+ */function Eg(n,t){const e=v(n,"x","tile","string_or_numeric");w(e.rank===t.length,()=>`Error in transpose: rank of input ${e.rank} must match length of reps ${t}.`);const s={x:e},r={reps:t};return A.runKernel(Ra,s,r)}const lr=C({tile_:Eg});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1278,7 +1278,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Cg(n,t,e,s="float32"){t==null&&(t=n);const r=xt([n,t],s),o=n<=t?n:t;for(let a=0;a<o;++a)r.set(1,a,a);const i=L(r.toTensor(),[n,t]);if(e==null)return i;if(e.length===1)return ar(ve(i,0),[e[0],1,1]);if(e.length===2)return ar(ve(ve(i,0),0),[e[0],e[1],1,1]);if(e.length===3)return ar(ve(ve(ve(i,0),0),0),[e[0],e[1],e[2],1,1]);throw new Error(`eye() currently supports only 1D and 2D batchShapes, but received ${e.length}D.`)}const Nl=C({eye_:Cg});/**
+ */function Cg(n,t,e,s="float32"){t==null&&(t=n);const r=xt([n,t],s),o=n<=t?n:t;for(let a=0;a<o;++a)r.set(1,a,a);const i=L(r.toTensor(),[n,t]);if(e==null)return i;if(e.length===1)return lr(ve(i,0),[e[0],1,1]);if(e.length===2)return lr(ve(ve(i,0),0),[e[0],e[1],1,1]);if(e.length===3)return lr(ve(ve(ve(i,0),0),0),[e[0],e[1],e[2],1,1]);throw new Error(`eye() currently supports only 1D and 2D batchShapes, but received ${e.length}D.`)}const Nl=C({eye_:Cg});/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1323,7 +1323,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Dg(n,t){let e=v(n,"a","greater","string_or_numeric"),s=v(t,"b","greater","string_or_numeric");[e,s]=Dt(e,s),zt(e.shape,s.shape);const r={a:e,b:s};return A.runKernel(pd,r)}const ds=C({greater_:Dg});/**
+ */function Dg(n,t){let e=v(n,"a","greater","string_or_numeric"),s=v(t,"b","greater","string_or_numeric");[e,s]=Dt(e,s),zt(e.shape,s.shape);const r={a:e,b:s};return A.runKernel(pd,r)}const ps=C({greater_:Dg});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1413,7 +1413,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Ug(n){const e={x:v(n,"x","log","float32")};return A.runKernel(xd,e)}const fn=C({log_:Ug});/**
+ */function Ug(n){const e={x:v(n,"x","log","float32")};return A.runKernel(xd,e)}const dn=C({log_:Ug});/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1443,7 +1443,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Vg(n,t){w(lo(n),()=>"The f passed in variableGrads(f) must be a function"),w(t==null||Array.isArray(t)&&t.every(u=>u instanceof Js),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");const e=t!=null;if(!e){t=[];for(const u in A.registeredVariables)t.push(A.registeredVariables[u])}const s=e?t.filter(u=>!u.trainable):null,r=t.length;t=t.filter(u=>u.trainable),w(t.length>0,()=>`variableGrads() expects at least one of the input variables to be trainable, but none of the ${r} variables is trainable.`);const o=!0,{value:i,grads:a}=A.gradients(n,t,null,o);w(a.some(u=>u!=null),()=>"Cannot find a connection between any variable and the result of the loss function y=f(x). Please make sure the operations that use variables are inside the function f passed to minimize()."),w(i.rank===0,()=>`The f passed in variableGrads(f) must return a scalar, but it returned a rank-${i.rank} tensor`);const l={};return t.forEach((u,c)=>{a[c]!=null&&(l[u.name]=a[c])}),s!=null&&s.forEach(u=>l[u.name]=null),{value:i,grads:l}}function Vo(n){return A.customGrad(n)}/**
+ */function Vg(n,t){w(lo(n),()=>"The f passed in variableGrads(f) must be a function"),w(t==null||Array.isArray(t)&&t.every(u=>u instanceof Zs),()=>"The varList passed in variableGrads(f, varList) must be an array of variables");const e=t!=null;if(!e){t=[];for(const u in A.registeredVariables)t.push(A.registeredVariables[u])}const s=e?t.filter(u=>!u.trainable):null,r=t.length;t=t.filter(u=>u.trainable),w(t.length>0,()=>`variableGrads() expects at least one of the input variables to be trainable, but none of the ${r} variables is trainable.`);const o=!0,{value:i,grads:a}=A.gradients(n,t,null,o);w(a.some(u=>u!=null),()=>"Cannot find a connection between any variable and the result of the loss function y=f(x). Please make sure the operations that use variables are inside the function f passed to minimize()."),w(i.rank===0,()=>`The f passed in variableGrads(f) must return a scalar, but it returned a rank-${i.rank} tensor`);const l={};return t.forEach((u,c)=>{a[c]!=null&&(l[u.name]=a[c])}),s!=null&&s.forEach(u=>l[u.name]=null),{value:i,grads:l}}function Vo(n){return A.customGrad(n)}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1503,7 +1503,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Kg(n,t=-1){const e=v(n,"logits","logSoftmax");if(t===-1&&(t=e.rank-1),t!==e.rank-1)throw Error(`Log Softmax along a non-last dimension is not yet supported. Logits was rank ${e.rank} and axis was ${t}`);return Vo((r,o)=>{const a=Ge(r,t,!0),l=J(r,a),u=J(ot(l,"float32"),fn(et(Go(l),t,!0)));return o([u]),{value:u,gradFunc:(h,f)=>{const[d]=f,p=!0,g=Go(d);return J(h,N(et(h,t,p),g))}}})(e)}const Yg=C({logSoftmax_:Kg});/**
+ */function Kg(n,t=-1){const e=v(n,"logits","logSoftmax");if(t===-1&&(t=e.rank-1),t!==e.rank-1)throw Error(`Log Softmax along a non-last dimension is not yet supported. Logits was rank ${e.rank} and axis was ${t}`);return Vo((r,o)=>{const a=Ge(r,t,!0),l=J(r,a),u=J(ot(l,"float32"),dn(et(Go(l),t,!0)));return o([u]),{value:u,gradFunc:(h,f)=>{const[d]=f,p=!0,g=Go(d);return J(h,N(et(h,t,p),g))}}})(e)}const Yg=C({logSoftmax_:Kg});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1518,7 +1518,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Xg(n,t){const e=v(n,"a","logicalAnd","bool"),s=v(t,"b","logicalAnd","bool");zt(e.shape,s.shape);const r={a:e,b:s};return A.runKernel($d,r)}const lr=C({logicalAnd_:Xg});/**
+ */function Xg(n,t){const e=v(n,"a","logicalAnd","bool"),s=v(t,"b","logicalAnd","bool");zt(e.shape,s.shape);const r={a:e,b:s};return A.runKernel($d,r)}const ur=C({logicalAnd_:Xg});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1623,7 +1623,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function s0(n,t){let e=v(n,"a","minimum"),s=v(t,"b","minimum");[e,s]=Dt(e,s),e.dtype==="bool"&&(e=ot(e,"int32"),s=ot(s,"int32")),zt(e.shape,s.shape);const r={a:e,b:s};return A.runKernel(kd,r)}const ur=C({minimum_:s0});/**
+ */function s0(n,t){let e=v(n,"a","minimum"),s=v(t,"b","minimum");[e,s]=Dt(e,s),e.dtype==="bool"&&(e=ot(e,"int32"),s=ot(s,"int32")),zt(e.shape,s.shape);const r={a:e,b:s};return A.runKernel(kd,r)}const cr=C({minimum_:s0});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1668,7 +1668,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function a0(n,t){const e=v(n,"x","prelu"),s=v(t,"alpha","prelu"),r={x:e,alpha:s};return A.runKernel(Bd,r)}const l0=C({prelu_:a0});var Ho={exports:{}};Ho.exports,function(n){(function(t,e,s){function r(l){var u=this,c=a();u.next=function(){var h=2091639*u.s0+u.c*23283064365386963e-26;return u.s0=u.s1,u.s1=u.s2,u.s2=h-(u.c=h|0)},u.c=1,u.s0=c(" "),u.s1=c(" "),u.s2=c(" "),u.s0-=c(l),u.s0<0&&(u.s0+=1),u.s1-=c(l),u.s1<0&&(u.s1+=1),u.s2-=c(l),u.s2<0&&(u.s2+=1),c=null}function o(l,u){return u.c=l.c,u.s0=l.s0,u.s1=l.s1,u.s2=l.s2,u}function i(l,u){var c=new r(l),h=u&&u.state,f=c.next;return f.int32=function(){return c.next()*4294967296|0},f.double=function(){return f()+(f()*2097152|0)*11102230246251565e-32},f.quick=f,h&&(typeof h=="object"&&o(h,c),f.state=function(){return o(c,{})}),f}function a(){var l=4022871197,u=function(c){c=String(c);for(var h=0;h<c.length;h++){l+=c.charCodeAt(h);var f=.02519603282416938*l;l=f>>>0,f-=l,f*=l,l=f>>>0,f-=l,l+=f*4294967296}return(l>>>0)*23283064365386963e-26};return u}e&&e.exports?e.exports=i:this.alea=i})(Je,n)}(Ho);var u0=Ho.exports,Ko={exports:{}};Ko.exports,function(n){(function(t,e,s){function r(a){var l=this,u="";l.x=0,l.y=0,l.z=0,l.w=0,l.next=function(){var h=l.x^l.x<<11;return l.x=l.y,l.y=l.z,l.z=l.w,l.w^=l.w>>>19^h^h>>>8},a===(a|0)?l.x=a:u+=a;for(var c=0;c<u.length+64;c++)l.x^=u.charCodeAt(c)|0,l.next()}function o(a,l){return l.x=a.x,l.y=a.y,l.z=a.z,l.w=a.w,l}function i(a,l){var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(typeof c=="object"&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.xor128=i})(Je,n)}(Ko);var c0=Ko.exports,Yo={exports:{}};Yo.exports,function(n){(function(t,e,s){function r(a){var l=this,u="";l.next=function(){var h=l.x^l.x>>>2;return l.x=l.y,l.y=l.z,l.z=l.w,l.w=l.v,(l.d=l.d+362437|0)+(l.v=l.v^l.v<<4^(h^h<<1))|0},l.x=0,l.y=0,l.z=0,l.w=0,l.v=0,a===(a|0)?l.x=a:u+=a;for(var c=0;c<u.length+64;c++)l.x^=u.charCodeAt(c)|0,c==u.length&&(l.d=l.x<<10^l.x>>>4),l.next()}function o(a,l){return l.x=a.x,l.y=a.y,l.z=a.z,l.w=a.w,l.v=a.v,l.d=a.d,l}function i(a,l){var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(typeof c=="object"&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.xorwow=i})(Je,n)}(Yo);var h0=Yo.exports,Xo={exports:{}};Xo.exports,function(n){(function(t,e,s){function r(a){var l=this;l.next=function(){var c=l.x,h=l.i,f,d;return f=c[h],f^=f>>>7,d=f^f<<24,f=c[h+1&7],d^=f^f>>>10,f=c[h+3&7],d^=f^f>>>3,f=c[h+4&7],d^=f^f<<7,f=c[h+7&7],f=f^f<<13,d^=f^f<<9,c[h]=d,l.i=h+1&7,d};function u(c,h){var f,d=[];if(h===(h|0))d[0]=h;else for(h=""+h,f=0;f<h.length;++f)d[f&7]=d[f&7]<<15^h.charCodeAt(f)+d[f+1&7]<<13;for(;d.length<8;)d.push(0);for(f=0;f<8&&d[f]===0;++f);for(f==8?d[7]=-1:d[f],c.x=d,c.i=0,f=256;f>0;--f)c.next()}u(l,a)}function o(a,l){return l.x=a.x.slice(),l.i=a.i,l}function i(a,l){a==null&&(a=+new Date);var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(c.x&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.xorshift7=i})(Je,n)}(Xo);var f0=Xo.exports,Jo={exports:{}};Jo.exports,function(n){(function(t,e,s){function r(a){var l=this;l.next=function(){var c=l.w,h=l.X,f=l.i,d,p;return l.w=c=c+1640531527|0,p=h[f+34&127],d=h[f=f+1&127],p^=p<<13,d^=d<<17,p^=p>>>15,d^=d>>>12,p=h[f]=p^d,l.i=f,p+(c^c>>>16)|0};function u(c,h){var f,d,p,g,m,b=[],y=128;for(h===(h|0)?(d=h,h=null):(h=h+"\0",d=0,y=Math.max(y,h.length)),p=0,g=-32;g<y;++g)h&&(d^=h.charCodeAt((g+32)%h.length)),g===0&&(m=d),d^=d<<10,d^=d>>>15,d^=d<<4,d^=d>>>13,g>=0&&(m=m+1640531527|0,f=b[g&127]^=d+m,p=f==0?p+1:0);for(p>=128&&(b[(h&&h.length||0)&127]=-1),p=127,g=4*128;g>0;--g)d=b[p+34&127],f=b[p=p+1&127],d^=d<<13,f^=f<<17,d^=d>>>15,f^=f>>>12,b[p]=d^f;c.w=m,c.X=b,c.i=p}u(l,a)}function o(a,l){return l.i=a.i,l.w=a.w,l.X=a.X.slice(),l}function i(a,l){a==null&&(a=+new Date);var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(c.X&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.xor4096=i})(Je,n)}(Jo);var d0=Jo.exports,Zo={exports:{}};Zo.exports,function(n){(function(t,e,s){function r(a){var l=this,u="";l.next=function(){var h=l.b,f=l.c,d=l.d,p=l.a;return h=h<<25^h>>>7^f,f=f-d|0,d=d<<24^d>>>8^p,p=p-h|0,l.b=h=h<<20^h>>>12^f,l.c=f=f-d|0,l.d=d<<16^f>>>16^p,l.a=p-h|0},l.a=0,l.b=0,l.c=-1640531527,l.d=1367130551,a===Math.floor(a)?(l.a=a/4294967296|0,l.b=a|0):u+=a;for(var c=0;c<u.length+20;c++)l.b^=u.charCodeAt(c)|0,l.next()}function o(a,l){return l.a=a.a,l.b=a.b,l.c=a.c,l.d=a.d,l}function i(a,l){var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(typeof c=="object"&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.tychei=i})(Je,n)}(Zo);var p0=Zo.exports,Ll={exports:{}};const m0=cp(Object.freeze(Object.defineProperty({__proto__:null,default:{}},Symbol.toStringTag,{value:"Module"})));(function(n){(function(t,e,s){var r=256,o=6,i=52,a="random",l=s.pow(r,o),u=s.pow(2,i),c=u*2,h=r-1,f;function d(x,$,E){var D=[];$=$==!0?{entropy:!0}:$||{};var _=b(m($.entropy?[x,S(e)]:x??y(),3),D),T=new p(D),P=function(){for(var B=T.g(o),Y=l,j=0;B<u;)B=(B+j)*r,Y*=r,j=T.g(1);for(;B>=c;)B/=2,Y/=2,j>>>=1;return(B+j)/Y};return P.int32=function(){return T.g(4)|0},P.quick=function(){return T.g(4)/4294967296},P.double=P,b(S(T.S),e),($.pass||E||function(B,Y,j,F){return F&&(F.S&&g(F,T),B.state=function(){return g(T,{})}),j?(s[a]=B,Y):B})(P,_,"global"in $?$.global:this==s,$.state)}function p(x){var $,E=x.length,D=this,_=0,T=D.i=D.j=0,P=D.S=[];for(E||(x=[E++]);_<r;)P[_]=_++;for(_=0;_<r;_++)P[_]=P[T=h&T+x[_%E]+($=P[_])],P[T]=$;(D.g=function(B){for(var Y,j=0,F=D.i,G=D.j,q=D.S;B--;)Y=q[F=h&F+1],j=j*r+q[h&(q[F]=q[G=h&G+Y])+(q[G]=Y)];return D.i=F,D.j=G,j})(r)}function g(x,$){return $.i=x.i,$.j=x.j,$.S=x.S.slice(),$}function m(x,$){var E=[],D=typeof x,_;if($&&D=="object")for(_ in x)try{E.push(m(x[_],$-1))}catch{}return E.length?E:D=="string"?x:x+"\0"}function b(x,$){for(var E=x+"",D,_=0;_<E.length;)$[h&_]=h&(D^=$[h&_]*19)+E.charCodeAt(_++);return S($)}function y(){try{var x;return f&&(x=f.randomBytes)?x=x(r):(x=new Uint8Array(r),(t.crypto||t.msCrypto).getRandomValues(x)),S(x)}catch{var $=t.navigator,E=$&&$.plugins;return[+new Date,t,E,t.screen,S(e)]}}function S(x){return String.fromCharCode.apply(0,x)}if(b(s.random(),e),n.exports){n.exports=d;try{f=m0}catch{}}else s["seed"+a]=d})(typeof self<"u"?self:Je,[],Math)})(Ll);var g0=Ll.exports,b0=u0,y0=c0,w0=h0,x0=f0,S0=d0,$0=p0,dn=g0;dn.alea=b0,dn.xor128=y0,dn.xorwow=w0,dn.xorshift7=x0,dn.xor4096=S0,dn.tychei=$0;var Ml=dn;/**
+ */function a0(n,t){const e=v(n,"x","prelu"),s=v(t,"alpha","prelu"),r={x:e,alpha:s};return A.runKernel(Bd,r)}const l0=C({prelu_:a0});var Ho={exports:{}};Ho.exports,function(n){(function(t,e,s){function r(l){var u=this,c=a();u.next=function(){var h=2091639*u.s0+u.c*23283064365386963e-26;return u.s0=u.s1,u.s1=u.s2,u.s2=h-(u.c=h|0)},u.c=1,u.s0=c(" "),u.s1=c(" "),u.s2=c(" "),u.s0-=c(l),u.s0<0&&(u.s0+=1),u.s1-=c(l),u.s1<0&&(u.s1+=1),u.s2-=c(l),u.s2<0&&(u.s2+=1),c=null}function o(l,u){return u.c=l.c,u.s0=l.s0,u.s1=l.s1,u.s2=l.s2,u}function i(l,u){var c=new r(l),h=u&&u.state,f=c.next;return f.int32=function(){return c.next()*4294967296|0},f.double=function(){return f()+(f()*2097152|0)*11102230246251565e-32},f.quick=f,h&&(typeof h=="object"&&o(h,c),f.state=function(){return o(c,{})}),f}function a(){var l=4022871197,u=function(c){c=String(c);for(var h=0;h<c.length;h++){l+=c.charCodeAt(h);var f=.02519603282416938*l;l=f>>>0,f-=l,f*=l,l=f>>>0,f-=l,l+=f*4294967296}return(l>>>0)*23283064365386963e-26};return u}e&&e.exports?e.exports=i:this.alea=i})(Ze,n)}(Ho);var u0=Ho.exports,Ko={exports:{}};Ko.exports,function(n){(function(t,e,s){function r(a){var l=this,u="";l.x=0,l.y=0,l.z=0,l.w=0,l.next=function(){var h=l.x^l.x<<11;return l.x=l.y,l.y=l.z,l.z=l.w,l.w^=l.w>>>19^h^h>>>8},a===(a|0)?l.x=a:u+=a;for(var c=0;c<u.length+64;c++)l.x^=u.charCodeAt(c)|0,l.next()}function o(a,l){return l.x=a.x,l.y=a.y,l.z=a.z,l.w=a.w,l}function i(a,l){var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(typeof c=="object"&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.xor128=i})(Ze,n)}(Ko);var c0=Ko.exports,Yo={exports:{}};Yo.exports,function(n){(function(t,e,s){function r(a){var l=this,u="";l.next=function(){var h=l.x^l.x>>>2;return l.x=l.y,l.y=l.z,l.z=l.w,l.w=l.v,(l.d=l.d+362437|0)+(l.v=l.v^l.v<<4^(h^h<<1))|0},l.x=0,l.y=0,l.z=0,l.w=0,l.v=0,a===(a|0)?l.x=a:u+=a;for(var c=0;c<u.length+64;c++)l.x^=u.charCodeAt(c)|0,c==u.length&&(l.d=l.x<<10^l.x>>>4),l.next()}function o(a,l){return l.x=a.x,l.y=a.y,l.z=a.z,l.w=a.w,l.v=a.v,l.d=a.d,l}function i(a,l){var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(typeof c=="object"&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.xorwow=i})(Ze,n)}(Yo);var h0=Yo.exports,Xo={exports:{}};Xo.exports,function(n){(function(t,e,s){function r(a){var l=this;l.next=function(){var c=l.x,h=l.i,f,d;return f=c[h],f^=f>>>7,d=f^f<<24,f=c[h+1&7],d^=f^f>>>10,f=c[h+3&7],d^=f^f>>>3,f=c[h+4&7],d^=f^f<<7,f=c[h+7&7],f=f^f<<13,d^=f^f<<9,c[h]=d,l.i=h+1&7,d};function u(c,h){var f,d=[];if(h===(h|0))d[0]=h;else for(h=""+h,f=0;f<h.length;++f)d[f&7]=d[f&7]<<15^h.charCodeAt(f)+d[f+1&7]<<13;for(;d.length<8;)d.push(0);for(f=0;f<8&&d[f]===0;++f);for(f==8?d[7]=-1:d[f],c.x=d,c.i=0,f=256;f>0;--f)c.next()}u(l,a)}function o(a,l){return l.x=a.x.slice(),l.i=a.i,l}function i(a,l){a==null&&(a=+new Date);var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(c.x&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.xorshift7=i})(Ze,n)}(Xo);var f0=Xo.exports,Jo={exports:{}};Jo.exports,function(n){(function(t,e,s){function r(a){var l=this;l.next=function(){var c=l.w,h=l.X,f=l.i,d,p;return l.w=c=c+1640531527|0,p=h[f+34&127],d=h[f=f+1&127],p^=p<<13,d^=d<<17,p^=p>>>15,d^=d>>>12,p=h[f]=p^d,l.i=f,p+(c^c>>>16)|0};function u(c,h){var f,d,p,g,m,b=[],y=128;for(h===(h|0)?(d=h,h=null):(h=h+"\0",d=0,y=Math.max(y,h.length)),p=0,g=-32;g<y;++g)h&&(d^=h.charCodeAt((g+32)%h.length)),g===0&&(m=d),d^=d<<10,d^=d>>>15,d^=d<<4,d^=d>>>13,g>=0&&(m=m+1640531527|0,f=b[g&127]^=d+m,p=f==0?p+1:0);for(p>=128&&(b[(h&&h.length||0)&127]=-1),p=127,g=4*128;g>0;--g)d=b[p+34&127],f=b[p=p+1&127],d^=d<<13,f^=f<<17,d^=d>>>15,f^=f>>>12,b[p]=d^f;c.w=m,c.X=b,c.i=p}u(l,a)}function o(a,l){return l.i=a.i,l.w=a.w,l.X=a.X.slice(),l}function i(a,l){a==null&&(a=+new Date);var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(c.X&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.xor4096=i})(Ze,n)}(Jo);var d0=Jo.exports,Zo={exports:{}};Zo.exports,function(n){(function(t,e,s){function r(a){var l=this,u="";l.next=function(){var h=l.b,f=l.c,d=l.d,p=l.a;return h=h<<25^h>>>7^f,f=f-d|0,d=d<<24^d>>>8^p,p=p-h|0,l.b=h=h<<20^h>>>12^f,l.c=f=f-d|0,l.d=d<<16^f>>>16^p,l.a=p-h|0},l.a=0,l.b=0,l.c=-1640531527,l.d=1367130551,a===Math.floor(a)?(l.a=a/4294967296|0,l.b=a|0):u+=a;for(var c=0;c<u.length+20;c++)l.b^=u.charCodeAt(c)|0,l.next()}function o(a,l){return l.a=a.a,l.b=a.b,l.c=a.c,l.d=a.d,l}function i(a,l){var u=new r(a),c=l&&l.state,h=function(){return(u.next()>>>0)/4294967296};return h.double=function(){do var f=u.next()>>>11,d=(u.next()>>>0)/4294967296,p=(f+d)/(1<<21);while(p===0);return p},h.int32=u.next,h.quick=h,c&&(typeof c=="object"&&o(c,u),h.state=function(){return o(u,{})}),h}e&&e.exports?e.exports=i:this.tychei=i})(Ze,n)}(Zo);var p0=Zo.exports,Ll={exports:{}};const m0=cp(Object.freeze(Object.defineProperty({__proto__:null,default:{}},Symbol.toStringTag,{value:"Module"})));(function(n){(function(t,e,s){var r=256,o=6,i=52,a="random",l=s.pow(r,o),u=s.pow(2,i),c=u*2,h=r-1,f;function d(x,$,E){var D=[];$=$==!0?{entropy:!0}:$||{};var _=b(m($.entropy?[x,S(e)]:x??y(),3),D),T=new p(D),P=function(){for(var B=T.g(o),Y=l,j=0;B<u;)B=(B+j)*r,Y*=r,j=T.g(1);for(;B>=c;)B/=2,Y/=2,j>>>=1;return(B+j)/Y};return P.int32=function(){return T.g(4)|0},P.quick=function(){return T.g(4)/4294967296},P.double=P,b(S(T.S),e),($.pass||E||function(B,Y,j,F){return F&&(F.S&&g(F,T),B.state=function(){return g(T,{})}),j?(s[a]=B,Y):B})(P,_,"global"in $?$.global:this==s,$.state)}function p(x){var $,E=x.length,D=this,_=0,T=D.i=D.j=0,P=D.S=[];for(E||(x=[E++]);_<r;)P[_]=_++;for(_=0;_<r;_++)P[_]=P[T=h&T+x[_%E]+($=P[_])],P[T]=$;(D.g=function(B){for(var Y,j=0,F=D.i,G=D.j,q=D.S;B--;)Y=q[F=h&F+1],j=j*r+q[h&(q[F]=q[G=h&G+Y])+(q[G]=Y)];return D.i=F,D.j=G,j})(r)}function g(x,$){return $.i=x.i,$.j=x.j,$.S=x.S.slice(),$}function m(x,$){var E=[],D=typeof x,_;if($&&D=="object")for(_ in x)try{E.push(m(x[_],$-1))}catch{}return E.length?E:D=="string"?x:x+"\0"}function b(x,$){for(var E=x+"",D,_=0;_<E.length;)$[h&_]=h&(D^=$[h&_]*19)+E.charCodeAt(_++);return S($)}function y(){try{var x;return f&&(x=f.randomBytes)?x=x(r):(x=new Uint8Array(r),(t.crypto||t.msCrypto).getRandomValues(x)),S(x)}catch{var $=t.navigator,E=$&&$.plugins;return[+new Date,t,E,t.screen,S(e)]}}function S(x){return String.fromCharCode.apply(0,x)}if(b(s.random(),e),n.exports){n.exports=d;try{f=m0}catch{}}else s["seed"+a]=d})(typeof self<"u"?self:Ze,[],Math)})(Ll);var g0=Ll.exports,b0=u0,y0=c0,w0=h0,x0=f0,S0=d0,$0=p0,pn=g0;pn.alea=b0,pn.xor128=y0,pn.xorwow=w0,pn.xorshift7=x0,pn.xor4096=S0,pn.tychei=$0;var Ml=pn;/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1728,7 +1728,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function cr(n,t,e=1,s="float32"){if(e===0)throw new Error("Cannot have a step of zero");const r={start:n,stop:t,step:e,dtype:s};return A.runKernel(Fd,{},r)}/**
+ */function hr(n,t,e=1,s="float32"){if(e===0)throw new Error("Cannot have a step of zero");const r={start:n,stop:t,step:e,dtype:s};return A.runKernel(Fd,{},r)}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1758,7 +1758,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function _0(n){const e={x:v(n,"x","relu")};return A.runKernel(Ud,e)}const ps=C({relu_:_0});/**
+ */function _0(n){const e={x:v(n,"x","relu")};return A.runKernel(Ud,e)}const ms=C({relu_:_0});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1893,7 +1893,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function G0(n,t){const e=v(n,"x","squeeze","string_or_numeric");return L(e,kf(e.shape,t).newShape)}const hr=C({squeeze_:G0});/**
+ */function G0(n,t){const e=v(n,"x","squeeze","string_or_numeric");return L(e,kf(e.shape,t).newShape)}const fr=C({squeeze_:G0});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1908,7 +1908,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function V0(n,t=0){const e=hl(n,"tensors","stack","string_or_numeric");w(e.length>=1,()=>"Pass at least one tensor to tf.stack"),e.length>0&&w(t<=e[0].rank,()=>"Axis must be <= rank of the tensor");const s=e,r={axis:t};return A.runKernel(Md,s,r)}const fr=C({stack_:V0});/**
+ */function V0(n,t=0){const e=hl(n,"tensors","stack","string_or_numeric");w(e.length>=1,()=>"Pass at least one tensor to tf.stack"),e.length>0&&w(t<=e[0].rank,()=>"Axis must be <= rank of the tensor");const s=e,r={axis:t};return A.runKernel(Md,s,r)}const dr=C({stack_:V0});/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -1938,7 +1938,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function ei(n,t,e){if(ya(n),t!=null&&t.length!==2)throw new Error("tensor2d() requires shape to have two numbers");const s=Zs(n,e);if(s.length!==2&&s.length!==1)throw new Error("tensor2d() requires values to be number[][] or flat/TypedArray");if(s.length===1&&t==null)throw new Error("tensor2d() requires shape to be provided when `values` are a flat/TypedArray");return Qs(n,t,s,e)}/**
+ */function ei(n,t,e){if(ya(n),t!=null&&t.length!==2)throw new Error("tensor2d() requires shape to have two numbers");const s=Qs(n,e);if(s.length!==2&&s.length!==1)throw new Error("tensor2d() requires values to be number[][] or flat/TypedArray");if(s.length===1&&t==null)throw new Error("tensor2d() requires shape to be provided when `values` are a flat/TypedArray");return tr(n,t,s,e)}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2028,7 +2028,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Q0(n,t,e){if(e==null||e==="linear")return n;if(e==="relu")return N(n,j0(t));throw new Error(`Cannot compute gradient for fused activation ${e}.`)}function tb(n,t){let e=t;const s=og(n.shape,t.shape);return s.length>0&&(e=et(e,s)),L(e,n.shape)}function eb(n,t,e,s){if(t==="linear")return n;if(t==="relu")return ps(n);if(t==="elu")return Al(n);if(t==="relu6")return N0(n);if(t==="prelu")return l0(n,e);if(t==="leakyrelu")return Bg(n,s);if(t==="sigmoid")return Fo(n);throw new Error(`Unknown fused activation ${t}.`)}const nb=(n,t)=>!(n>0)||t==="linear";/**
+ */function Q0(n,t,e){if(e==null||e==="linear")return n;if(e==="relu")return N(n,j0(t));throw new Error(`Cannot compute gradient for fused activation ${e}.`)}function tb(n,t){let e=t;const s=og(n.shape,t.shape);return s.length>0&&(e=et(e,s)),L(e,n.shape)}function eb(n,t,e,s){if(t==="linear")return n;if(t==="relu")return ms(n);if(t==="elu")return Al(n);if(t==="relu6")return N0(n);if(t==="prelu")return l0(n,e);if(t==="leakyrelu")return Bg(n,s);if(t==="sigmoid")return Fo(n);throw new Error(`Unknown fused activation ${t}.`)}const nb=(n,t)=>!(n>0)||t==="linear";/**
  * @license
  * Copyright 2019 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2088,7 +2088,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function ub(n){const t=v(n,"image","grayscaleToRGB"),e=t.rank-1,s=t.shape[e];w(t.rank>=2,()=>`Error in grayscaleToRGB: images must be at least rank 2, but got rank ${t.rank}.`),w(s===1,()=>`Error in grayscaleToRGB: last dimension of a grayscale image should be size 1, but got size ${s}.`);const r=new Array(t.rank);return r.fill(1,0,e),r[e]=3,ar(t,r)}const cb=C({grayscaleToRGB_:ub});/**
+ */function ub(n){const t=v(n,"image","grayscaleToRGB"),e=t.rank-1,s=t.shape[e];w(t.rank>=2,()=>`Error in grayscaleToRGB: images must be at least rank 2, but got rank ${t.rank}.`),w(s===1,()=>`Error in grayscaleToRGB: last dimension of a grayscale image should be size 1, but got size ${s}.`);const r=new Array(t.rank);return r.fill(1,0,e),r[e]=3,lr(t,r)}const cb=C({grayscaleToRGB_:ub});/**
  * @license
  * Copyright 2023 Google LLC.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2103,7 +2103,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function hb(n){const t=v(n,"image","RGBToGrayscale"),e=t.rank-1,s=t.shape[e];w(t.rank>=2,()=>`Error in RGBToGrayscale: images must be at least rank 2, but got rank ${t.rank}.`),w(s===3,()=>`Error in RGBToGrayscale: last dimension of an RGB image should be size 3, but got size ${s}.`);const r=t.dtype,o=ot(t,"float32"),i=Rt([.2989,.587,.114]);let a;switch(t.rank){case 2:a=fs("ij,j->i",o,i);break;case 3:a=fs("ijk,k->ij",o,i);break;case 4:a=fs("ijkl,l->ijk",o,i);break;case 5:a=fs("ijklm,m->ijkl",o,i);break;case 6:a=fs("ijklmn,n->ijklm",o,i);break;default:throw new Error("Not a valid tensor rank.")}return a=ve(a,-1),ot(a,r)}const fb=C({rgbToGrayscale_:hb});/**
+ */function hb(n){const t=v(n,"image","RGBToGrayscale"),e=t.rank-1,s=t.shape[e];w(t.rank>=2,()=>`Error in RGBToGrayscale: images must be at least rank 2, but got rank ${t.rank}.`),w(s===3,()=>`Error in RGBToGrayscale: last dimension of an RGB image should be size 3, but got size ${s}.`);const r=t.dtype,o=ot(t,"float32"),i=Rt([.2989,.587,.114]);let a;switch(t.rank){case 2:a=ds("ij,j->i",o,i);break;case 3:a=ds("ijk,k->ij",o,i);break;case 4:a=ds("ijkl,l->ijk",o,i);break;case 5:a=ds("ijklm,m->ijkl",o,i);break;case 6:a=ds("ijklmn,n->ijklm",o,i);break;default:throw new Error("Not a valid tensor rank.")}return a=ve(a,-1),ot(a,r)}const fb=C({rgbToGrayscale_:hb});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2253,7 +2253,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */async function Pb(n,t,e,s=.5,r=Number.NEGATIVE_INFINITY,o=!1){const i=v(n,"boxes","nonMaxSuppressionAsync"),a=v(t,"scores","nonMaxSuppressionAsync"),l=Wn(i,a,e,s,r,null),u=l.maxOutputSize,c=l.iouThreshold,h=l.scoreThreshold,[f,d]=await Promise.all([i.data(),a.data()]),{selectedIndices:p,validOutputs:g}=$b(f,d,u,c,h,o);return i!==n&&i.dispose(),a!==t&&a.dispose(),{selectedIndices:Rt(p,"int32"),validOutputs:Ht(g,"int32")}}const Lb=Pb;/**
+ */async function Pb(n,t,e,s=.5,r=Number.NEGATIVE_INFINITY,o=!1){const i=v(n,"boxes","nonMaxSuppressionAsync"),a=v(t,"scores","nonMaxSuppressionAsync"),l=Wn(i,a,e,s,r,null),u=l.maxOutputSize,c=l.iouThreshold,h=l.scoreThreshold,[f,d]=await Promise.all([i.data(),a.data()]),{selectedIndices:p,validOutputs:g}=$b(f,d,u,c,h,o);return i!==n&&i.dispose(),a!==t&&a.dispose(),{selectedIndices:Rt(p,"int32"),validOutputs:Yt(g,"int32")}}const Lb=Pb;/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2298,7 +2298,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function zb(n,t="binary",e=!1,s=.5){const r=v(n,"image","threshold"),o=.2989,i=.587,a=.114,l=r.shape[0]*r.shape[1];let u=N(Rt([s]),255),c,h,f,d;if(w(r.rank===3,()=>`Error in threshold: image must be rank 3,but got rank ${r.rank}.`),w(r.shape[2]===3||r.shape[2]===1,()=>`Error in threshold: image color channel must be equal to 3 or 1but got ${r.shape[2]}.`),w(r.dtype==="int32"||r.dtype==="float32",()=>`Error in dtype: image dtype must be int32 or float32,but got dtype ${r.dtype}.`),w(t==="otsu"||t==="binary",()=>`Method must be binary or otsu, but was ${t}`),r.shape[2]===3){[c,h,f]=Ul(r,[1,1,1],-1);const m=N(c,o),b=N(h,i),y=N(f,a);d=O(O(m,b),y)}else d=n;if(t==="otsu"){const m=Wm(ot(R0(d),"int32"),tr([]),256);u=Ub(m,l)}const p=e?Rl(d,u):ds(d,u);return ot(N(p,255),"int32")}function Ub(n,t){let e=Rt([-1]),s=Rt([0]),r=Rt([0]),o,i,a,l,u,c;for(let h=0;h<n.size-1;h++){o=Et(n,0,h+1),i=Et(n,h+1),u=X(et(o),t),c=X(et(i),t);const f=et(N(o,cr(0,o.size)));a=X(f,et(o));const d=rr(i.shape,o.size),p=O(cr(0,i.size),d),g=N(i,p);l=X(et(g),et(i));const m=J(a,l),b=J(a,l),y=N(u,c);r=N(N(y,m),b);const S=ds(r,s);s=hn(S,r,s),e=hn(S,Rt([h]),e)}return e}const Wb=C({threshold_:zb});/**
+ */function zb(n,t="binary",e=!1,s=.5){const r=v(n,"image","threshold"),o=.2989,i=.587,a=.114,l=r.shape[0]*r.shape[1];let u=N(Rt([s]),255),c,h,f,d;if(w(r.rank===3,()=>`Error in threshold: image must be rank 3,but got rank ${r.rank}.`),w(r.shape[2]===3||r.shape[2]===1,()=>`Error in threshold: image color channel must be equal to 3 or 1but got ${r.shape[2]}.`),w(r.dtype==="int32"||r.dtype==="float32",()=>`Error in dtype: image dtype must be int32 or float32,but got dtype ${r.dtype}.`),w(t==="otsu"||t==="binary",()=>`Method must be binary or otsu, but was ${t}`),r.shape[2]===3){[c,h,f]=Ul(r,[1,1,1],-1);const m=N(c,o),b=N(h,i),y=N(f,a);d=O(O(m,b),y)}else d=n;if(t==="otsu"){const m=Wm(ot(R0(d),"int32"),er([]),256);u=Ub(m,l)}const p=e?Rl(d,u):ps(d,u);return ot(N(p,255),"int32")}function Ub(n,t){let e=Rt([-1]),s=Rt([0]),r=Rt([0]),o,i,a,l,u,c;for(let h=0;h<n.size-1;h++){o=Et(n,0,h+1),i=Et(n,h+1),u=X(et(o),t),c=X(et(i),t);const f=et(N(o,hr(0,o.size)));a=X(f,et(o));const d=or(i.shape,o.size),p=O(hr(0,i.size),d),g=N(i,p);l=X(et(g),et(i));const m=J(a,l),b=J(a,l),y=N(u,c);r=N(N(y,m),b);const S=ps(r,s);s=fn(S,r,s),e=fn(S,Rt([h]),e)}return e}const Wb=C({threshold_:zb});/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2328,7 +2328,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function qb(n,t,e){const s=v(n,"a","bandPart");w(s.rank>=2,()=>`bandPart(): Rank must be at least 2, got ${s.rank}.`);const r=s.shape,[o,i]=s.shape.slice(-2);let a,l;typeof t=="number"?(w(t%1===0,()=>`bandPart(): numLower must be an integer, got ${t}.`),w(t<=o,()=>`bandPart(): numLower (${t}) must not be greater than the number of rows (${o}).`),a=v(t<0?o:t,"numLower","bandPart")):(w(t.dtype==="int32",()=>"bandPart(): numLower's dtype must be an int32."),a=hn(Dl(t,0),o,ur(t,o))),typeof e=="number"?(w(e%1===0,()=>`bandPart(): numUpper must be an integer, got ${e}.`),w(e<=i,()=>`bandPart(): numUpper (${e}) must not be greater than the number of columns (${i}).`),l=v(e<0?i:e,"numUpper","bandPart")):(w(e.dtype==="int32",()=>"bandPart(): numUpper's dtype must be an int32."),l=hn(Dl(e,0),i,ur(e,i)));const u=L(cr(0,o,1,"int32"),[-1,1]),c=cr(0,i,1,"int32"),h=J(u,c),f=lr(Rl(h,a),Pg(h,Fn(l))),d=Un([o,i],s.dtype);return L(fr(Gl(L(s,[-1,o,i])).map(p=>hn(f,p,d))),r)}const jb=C({bandPart_:qb});/**
+ */function qb(n,t,e){const s=v(n,"a","bandPart");w(s.rank>=2,()=>`bandPart(): Rank must be at least 2, got ${s.rank}.`);const r=s.shape,[o,i]=s.shape.slice(-2);let a,l;typeof t=="number"?(w(t%1===0,()=>`bandPart(): numLower must be an integer, got ${t}.`),w(t<=o,()=>`bandPart(): numLower (${t}) must not be greater than the number of rows (${o}).`),a=v(t<0?o:t,"numLower","bandPart")):(w(t.dtype==="int32",()=>"bandPart(): numLower's dtype must be an int32."),a=fn(Dl(t,0),o,cr(t,o))),typeof e=="number"?(w(e%1===0,()=>`bandPart(): numUpper must be an integer, got ${e}.`),w(e<=i,()=>`bandPart(): numUpper (${e}) must not be greater than the number of columns (${i}).`),l=v(e<0?i:e,"numUpper","bandPart")):(w(e.dtype==="int32",()=>"bandPart(): numUpper's dtype must be an int32."),l=fn(Dl(e,0),i,cr(e,i)));const u=L(hr(0,o,1,"int32"),[-1,1]),c=hr(0,i,1,"int32"),h=J(u,c),f=ur(Rl(h,a),Pg(h,Fn(l))),d=Un([o,i],s.dtype);return L(dr(Gl(L(s,[-1,o,i])).map(p=>fn(f,p,d))),r)}const jb=C({bandPart_:qb});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2343,7 +2343,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Hb(n){let t;if(Array.isArray(n)){t=!1,w(n!=null&&n.length>0,()=>"Gram-Schmidt process: input must not be null, undefined, or empty");const r=n[0].shape[0];for(let o=1;o<n.length;++o)w(n[o].shape[0]===r,()=>`Gram-Schmidt: Non-unique lengths found in the input vectors: (${n[o].shape[0]} vs. ${r})`)}else t=!0,n=Ul(n,n.shape[0],0).map(r=>hr(r,[0]));w(n.length<=n[0].shape[0],()=>`Gram-Schmidt: Number of vectors (${n.length}) exceeds number of dimensions (${n[0].shape[0]}).`);const e=[],s=n;for(let r=0;r<n.length;++r)e.push(A.tidy(()=>{let o=s[r];if(r>0)for(let i=0;i<r;++i){const a=N(et(N(e[i],o)),e[i]);o=J(o,a)}return X(o,Tl(o,"euclidean"))}));return t?fr(e,0):e}const Kb=C({gramSchmidt_:Hb});/**
+ */function Hb(n){let t;if(Array.isArray(n)){t=!1,w(n!=null&&n.length>0,()=>"Gram-Schmidt process: input must not be null, undefined, or empty");const r=n[0].shape[0];for(let o=1;o<n.length;++o)w(n[o].shape[0]===r,()=>`Gram-Schmidt: Non-unique lengths found in the input vectors: (${n[o].shape[0]} vs. ${r})`)}else t=!0,n=Ul(n,n.shape[0],0).map(r=>fr(r,[0]));w(n.length<=n[0].shape[0],()=>`Gram-Schmidt: Number of vectors (${n.length}) exceeds number of dimensions (${n[0].shape[0]}).`);const e=[],s=n;for(let r=0;r<n.length;++r)e.push(A.tidy(()=>{let o=s[r];if(r>0)for(let i=0;i<r;++i){const a=N(et(N(e[i],o)),e[i]);o=J(o,a)}return X(o,Tl(o,"euclidean"))}));return t?dr(e,0):e}const Kb=C({gramSchmidt_:Hb});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2358,7 +2358,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Yb(n,t=!1){if(w(n.rank>=2,()=>`qr() requires input tensor to have a rank >= 2, but got rank ${n.rank}`),n.rank===2)return ql(n,t);{const e=n.shape.slice(0,n.shape.length-2).reduce((l,u)=>l*u),s=Gl(L(n,[e,n.shape[n.shape.length-2],n.shape[n.shape.length-1]]),0),r=[],o=[];s.forEach(l=>{const[u,c]=ql(l,t);r.push(u),o.push(c)});const i=L(fr(r,0),n.shape),a=L(fr(o,0),n.shape);return[i,a]}}function ql(n,t=!1){return A.tidy(()=>{w(n.shape.length===2,()=>`qr2d() requires a 2D Tensor, but got a ${n.shape.length}D Tensor.`);const e=n.shape[0],s=n.shape[1];let r=Nl(e),o=rn(n);const i=ei([[1]],[1,1]);let a=rn(i);const l=e>=s?s:e;for(let u=0;u<l;++u){const c=o,h=a,f=r;[a,o,r]=A.tidy(()=>{const d=Et(o,[u,u],[e-u,1]),p=Tl(d),g=Et(o,[u,u],[1,1]),m=hn(ds(g,0),ei([[-1]]),ei([[1]])),b=J(g,N(m,p)),y=X(d,b);y.shape[0]===1?a=rn(i):a=on([i,Et(y,[1,0],[y.shape[0]-1,y.shape[1]])],0);const S=Fn(X(Se(m,b),p)),x=Et(o,[u,0],[e-u,s]),$=N(S,a),E=pt(a);if(u===0)o=J(x,Se($,Se(E,x)));else{const T=J(x,Se($,Se(E,x)));o=on([Et(o,[0,0],[u,s]),T],0)}const D=pt($),_=Et(r,[0,u],[e,r.shape[1]-u]);if(u===0)r=J(_,Se(Se(_,a),D));else{const T=J(_,Se(Se(_,a),D));r=on([Et(r,[0,0],[e,u]),T],1)}return[a,o,r]}),ut([c,h,f])}return!t&&e>s&&(r=Et(r,[0,0],[e,s]),o=Et(o,[0,0],[s,s])),[r,o]})}const Xb=C({qr_:Yb});/**
+ */function Yb(n,t=!1){if(w(n.rank>=2,()=>`qr() requires input tensor to have a rank >= 2, but got rank ${n.rank}`),n.rank===2)return ql(n,t);{const e=n.shape.slice(0,n.shape.length-2).reduce((l,u)=>l*u),s=Gl(L(n,[e,n.shape[n.shape.length-2],n.shape[n.shape.length-1]]),0),r=[],o=[];s.forEach(l=>{const[u,c]=ql(l,t);r.push(u),o.push(c)});const i=L(dr(r,0),n.shape),a=L(dr(o,0),n.shape);return[i,a]}}function ql(n,t=!1){return A.tidy(()=>{w(n.shape.length===2,()=>`qr2d() requires a 2D Tensor, but got a ${n.shape.length}D Tensor.`);const e=n.shape[0],s=n.shape[1];let r=Nl(e),o=on(n);const i=ei([[1]],[1,1]);let a=on(i);const l=e>=s?s:e;for(let u=0;u<l;++u){const c=o,h=a,f=r;[a,o,r]=A.tidy(()=>{const d=Et(o,[u,u],[e-u,1]),p=Tl(d),g=Et(o,[u,u],[1,1]),m=fn(ps(g,0),ei([[-1]]),ei([[1]])),b=J(g,N(m,p)),y=X(d,b);y.shape[0]===1?a=on(i):a=an([i,Et(y,[1,0],[y.shape[0]-1,y.shape[1]])],0);const S=Fn(X(Se(m,b),p)),x=Et(o,[u,0],[e-u,s]),$=N(S,a),E=pt(a);if(u===0)o=J(x,Se($,Se(E,x)));else{const T=J(x,Se($,Se(E,x)));o=an([Et(o,[0,0],[u,s]),T],0)}const D=pt($),_=Et(r,[0,u],[e,r.shape[1]-u]);if(u===0)r=J(_,Se(Se(_,a),D));else{const T=J(_,Se(Se(_,a),D));r=an([Et(r,[0,0],[e,u]),T],1)}return[a,o,r]}),ut([c,h,f])}return!t&&e>s&&(r=Et(r,[0,0],[e,s]),o=Et(o,[0,0],[s,s])),[r,o]})}const Xb=C({qr_:Yb});/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2373,7 +2373,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const dr={flipLeftRight:lb,grayscaleToRGB:cb,resizeNearestNeighbor:Fb,resizeBilinear:Ob,rgbToGrayscale:fb,rotateWithOffset:pb,cropAndResize:ib,nonMaxSuppression:gb,nonMaxSuppressionAsync:Cb,nonMaxSuppressionWithScore:_b,nonMaxSuppressionWithScoreAsync:Nb,nonMaxSuppressionPadded:Rb,nonMaxSuppressionPaddedAsync:Lb,threshold:Wb,transform:Vb},Jb={bandPart:jb,gramSchmidt:Kb,qr:Xb};/**
+ */const pr={flipLeftRight:lb,grayscaleToRGB:cb,resizeNearestNeighbor:Fb,resizeBilinear:Ob,rgbToGrayscale:fb,rotateWithOffset:pb,cropAndResize:ib,nonMaxSuppression:gb,nonMaxSuppressionAsync:Cb,nonMaxSuppressionWithScore:_b,nonMaxSuppressionWithScoreAsync:Nb,nonMaxSuppressionPadded:Rb,nonMaxSuppressionPaddedAsync:Lb,threshold:Wb,transform:Vb},Jb={bandPart:jb,gramSchmidt:Kb,qr:Xb};/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2403,7 +2403,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class qe extends Gn{minimize(t,e=!1,s){const{value:r,grads:o}=this.computeGradients(t,s);if(s!=null){const i=s.map(a=>({name:a.name,tensor:o[a.name]}));this.applyGradients(i)}else this.applyGradients(o);return ut(o),e?r:(r.dispose(),null)}get iterations(){return this.iterations_==null&&(this.iterations_=0),this.iterations_}incrementIterations(){this.iterations_=this.iterations+1}computeGradients(t,e){return Vg(t,e)}dispose(){this.iterations_!=null&&ut(this.iterations_)}async saveIterations(){return this.iterations_==null&&(this.iterations_=0),{name:"iter",tensor:Ht(this.iterations_,"int32")}}async getWeights(){throw new Error("getWeights() is not implemented for this optimizer yet.")}async setWeights(t){throw new Error(`setWeights() is not implemented for this optimizer class ${this.getClassName()}`)}async extractIterations(t){return this.iterations_=(await t[0].tensor.data())[0],t.slice(1)}}Object.defineProperty(qe,Symbol.hasInstance,{value:n=>n.minimize!=null&&n.computeGradients!=null&&n.applyGradients!=null});/**
+ */class qe extends Gn{minimize(t,e=!1,s){const{value:r,grads:o}=this.computeGradients(t,s);if(s!=null){const i=s.map(a=>({name:a.name,tensor:o[a.name]}));this.applyGradients(i)}else this.applyGradients(o);return ut(o),e?r:(r.dispose(),null)}get iterations(){return this.iterations_==null&&(this.iterations_=0),this.iterations_}incrementIterations(){this.iterations_=this.iterations+1}computeGradients(t,e){return Vg(t,e)}dispose(){this.iterations_!=null&&ut(this.iterations_)}async saveIterations(){return this.iterations_==null&&(this.iterations_=0),{name:"iter",tensor:Yt(this.iterations_,"int32")}}async getWeights(){throw new Error("getWeights() is not implemented for this optimizer yet.")}async setWeights(t){throw new Error(`setWeights() is not implemented for this optimizer class ${this.getClassName()}`)}async extractIterations(t){return this.iterations_=(await t[0].tensor.data())[0],t.slice(1)}}Object.defineProperty(qe,Symbol.hasInstance,{value:n=>n.minimize!=null&&n.computeGradients!=null&&n.applyGradients!=null});/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2433,7 +2433,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class Hl extends qe{static get className(){return"Adagrad"}constructor(t,e=.1){super(),this.learningRate=t,this.initialAccumulatorValue=e,this.accumulatedGrads=[]}applyGradients(t){(Array.isArray(t)?t.map(s=>s.name):Object.keys(t)).forEach((s,r)=>{const o=A.registeredVariables[s];this.accumulatedGrads[r]==null&&(this.accumulatedGrads[r]={originalName:`${s}/accumulator`,variable:k(()=>rr(o.shape,this.initialAccumulatorValue).variable(!1))});const i=Array.isArray(t)?t[r].tensor:t[s];if(i==null)return;const a=this.accumulatedGrads[r].variable;k(()=>{const l=O(a,Ve(i));a.assign(l);const u=O(N(X(i,fe(O(l,A.backend.epsilon()))),-this.learningRate),o);o.assign(u)})}),this.incrementIterations()}dispose(){this.accumulatedGrads!=null&&ut(this.accumulatedGrads.map(t=>t.variable))}async getWeights(){return[await this.saveIterations()].concat(this.accumulatedGrads.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(t){t=await this.extractIterations(t);const e=!1;this.accumulatedGrads=t.map(s=>({originalName:s.name,variable:s.tensor.variable(e)}))}getConfig(){return{learningRate:this.learningRate,initialAccumulatorValue:this.initialAccumulatorValue}}static fromConfig(t,e){return new t(e.learningRate,e.initialAccumulatorValue)}}/**
+ */class Hl extends qe{static get className(){return"Adagrad"}constructor(t,e=.1){super(),this.learningRate=t,this.initialAccumulatorValue=e,this.accumulatedGrads=[]}applyGradients(t){(Array.isArray(t)?t.map(s=>s.name):Object.keys(t)).forEach((s,r)=>{const o=A.registeredVariables[s];this.accumulatedGrads[r]==null&&(this.accumulatedGrads[r]={originalName:`${s}/accumulator`,variable:k(()=>or(o.shape,this.initialAccumulatorValue).variable(!1))});const i=Array.isArray(t)?t[r].tensor:t[s];if(i==null)return;const a=this.accumulatedGrads[r].variable;k(()=>{const l=O(a,Ve(i));a.assign(l);const u=O(N(X(i,fe(O(l,A.backend.epsilon()))),-this.learningRate),o);o.assign(u)})}),this.incrementIterations()}dispose(){this.accumulatedGrads!=null&&ut(this.accumulatedGrads.map(t=>t.variable))}async getWeights(){return[await this.saveIterations()].concat(this.accumulatedGrads.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(t){t=await this.extractIterations(t);const e=!1;this.accumulatedGrads=t.map(s=>({originalName:s.name,variable:s.tensor.variable(e)}))}getConfig(){return{learningRate:this.learningRate,initialAccumulatorValue:this.initialAccumulatorValue}}static fromConfig(t,e){return new t(e.learningRate,e.initialAccumulatorValue)}}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2448,7 +2448,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class Kl extends qe{static get className(){return"Adam"}constructor(t,e,s,r=null){super(),this.learningRate=t,this.beta1=e,this.beta2=s,this.epsilon=r,this.accumulatedFirstMoment=[],this.accumulatedSecondMoment=[],k(()=>{this.accBeta1=Ht(e).variable(),this.accBeta2=Ht(s).variable()}),r==null&&(this.epsilon=A.backend.epsilon())}applyGradients(t){const e=Array.isArray(t)?t.map(s=>s.name):Object.keys(t);k(()=>{const s=J(1,this.accBeta1),r=J(1,this.accBeta2);e.forEach((o,i)=>{const a=A.registeredVariables[o],l=!1;this.accumulatedFirstMoment[i]==null&&(this.accumulatedFirstMoment[i]={originalName:`${o}/m`,variable:k(()=>$e(a).variable(l))}),this.accumulatedSecondMoment[i]==null&&(this.accumulatedSecondMoment[i]={originalName:`${o}/v`,variable:k(()=>$e(a).variable(l))});const u=Array.isArray(t)?t[i].tensor:t[o];if(u==null)return;const c=this.accumulatedFirstMoment[i].variable,h=this.accumulatedSecondMoment[i].variable,f=O(N(c,this.beta1),N(u,1-this.beta1)),d=O(N(h,this.beta2),N(Ve(u),1-this.beta2)),p=X(f,s),g=X(d,r);c.assign(f),h.assign(d);const m=O(N(X(p,O(fe(g),this.epsilon)),-this.learningRate),a);a.assign(m)}),this.accBeta1.assign(N(this.accBeta1,this.beta1)),this.accBeta2.assign(N(this.accBeta2,this.beta2))}),this.incrementIterations()}dispose(){this.accBeta1.dispose(),this.accBeta2.dispose(),this.accumulatedFirstMoment!=null&&ut(this.accumulatedFirstMoment.map(t=>t.variable)),this.accumulatedSecondMoment!=null&&ut(this.accumulatedSecondMoment.map(t=>t.variable))}async getWeights(){const t=[...this.accumulatedFirstMoment,...this.accumulatedSecondMoment];return[await this.saveIterations()].concat(t.map(e=>({name:e.originalName,tensor:e.variable})))}async setWeights(t){t=await this.extractIterations(t),k(()=>{this.accBeta1.assign(ir(this.beta1,this.iterations_+1)),this.accBeta2.assign(ir(this.beta2,this.iterations_+1))});const e=t.length/2,s=!1;this.accumulatedFirstMoment=t.slice(0,e).map(r=>({originalName:r.name,variable:r.tensor.variable(s)})),this.accumulatedSecondMoment=t.slice(e,e*2).map(r=>({originalName:r.name,variable:r.tensor.variable(s)}))}getConfig(){return{learningRate:this.learningRate,beta1:this.beta1,beta2:this.beta2,epsilon:this.epsilon}}static fromConfig(t,e){return new t(e.learningRate,e.beta1,e.beta2,e.epsilon)}}/**
+ */class Kl extends qe{static get className(){return"Adam"}constructor(t,e,s,r=null){super(),this.learningRate=t,this.beta1=e,this.beta2=s,this.epsilon=r,this.accumulatedFirstMoment=[],this.accumulatedSecondMoment=[],k(()=>{this.accBeta1=Yt(e).variable(),this.accBeta2=Yt(s).variable()}),r==null&&(this.epsilon=A.backend.epsilon())}applyGradients(t){const e=Array.isArray(t)?t.map(s=>s.name):Object.keys(t);k(()=>{const s=J(1,this.accBeta1),r=J(1,this.accBeta2);e.forEach((o,i)=>{const a=A.registeredVariables[o],l=!1;this.accumulatedFirstMoment[i]==null&&(this.accumulatedFirstMoment[i]={originalName:`${o}/m`,variable:k(()=>$e(a).variable(l))}),this.accumulatedSecondMoment[i]==null&&(this.accumulatedSecondMoment[i]={originalName:`${o}/v`,variable:k(()=>$e(a).variable(l))});const u=Array.isArray(t)?t[i].tensor:t[o];if(u==null)return;const c=this.accumulatedFirstMoment[i].variable,h=this.accumulatedSecondMoment[i].variable,f=O(N(c,this.beta1),N(u,1-this.beta1)),d=O(N(h,this.beta2),N(Ve(u),1-this.beta2)),p=X(f,s),g=X(d,r);c.assign(f),h.assign(d);const m=O(N(X(p,O(fe(g),this.epsilon)),-this.learningRate),a);a.assign(m)}),this.accBeta1.assign(N(this.accBeta1,this.beta1)),this.accBeta2.assign(N(this.accBeta2,this.beta2))}),this.incrementIterations()}dispose(){this.accBeta1.dispose(),this.accBeta2.dispose(),this.accumulatedFirstMoment!=null&&ut(this.accumulatedFirstMoment.map(t=>t.variable)),this.accumulatedSecondMoment!=null&&ut(this.accumulatedSecondMoment.map(t=>t.variable))}async getWeights(){const t=[...this.accumulatedFirstMoment,...this.accumulatedSecondMoment];return[await this.saveIterations()].concat(t.map(e=>({name:e.originalName,tensor:e.variable})))}async setWeights(t){t=await this.extractIterations(t),k(()=>{this.accBeta1.assign(ar(this.beta1,this.iterations_+1)),this.accBeta2.assign(ar(this.beta2,this.iterations_+1))});const e=t.length/2,s=!1;this.accumulatedFirstMoment=t.slice(0,e).map(r=>({originalName:r.name,variable:r.tensor.variable(s)})),this.accumulatedSecondMoment=t.slice(e,e*2).map(r=>({originalName:r.name,variable:r.tensor.variable(s)}))}getConfig(){return{learningRate:this.learningRate,beta1:this.beta1,beta2:this.beta2,epsilon:this.epsilon}}static fromConfig(t,e){return new t(e.learningRate,e.beta1,e.beta2,e.epsilon)}}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2463,7 +2463,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class Yl extends qe{static get className(){return"Adamax"}constructor(t,e,s,r=null,o=0){super(),this.learningRate=t,this.beta1=e,this.beta2=s,this.epsilon=r,this.decay=o,this.accumulatedFirstMoment=[],this.accumulatedWeightedInfNorm=[],k(()=>{this.iteration=Ht(0).variable(),this.accBeta1=Ht(e).variable()}),r==null&&(this.epsilon=A.backend.epsilon())}applyGradients(t){const e=Array.isArray(t)?t.map(s=>s.name):Object.keys(t);k(()=>{const s=J(1,this.accBeta1),r=X(-this.learningRate,O(N(this.iteration,this.decay),1));e.forEach((o,i)=>{const a=A.registeredVariables[o],l=!1;this.accumulatedFirstMoment[i]==null&&(this.accumulatedFirstMoment[i]={originalName:`${o}/m`,variable:$e(a).variable(l)}),this.accumulatedWeightedInfNorm[i]==null&&(this.accumulatedWeightedInfNorm[i]={originalName:`${o}/v`,variable:$e(a).variable(l)});const u=Array.isArray(t)?t[i].tensor:t[o];if(u==null)return;const c=this.accumulatedFirstMoment[i].variable,h=this.accumulatedWeightedInfNorm[i].variable,f=O(N(c,this.beta1),N(u,1-this.beta1)),d=N(h,this.beta2),p=Lt(u),g=zn(d,p);c.assign(f),h.assign(g);const m=O(N(X(r,s),X(f,O(g,this.epsilon))),a);a.assign(m)}),this.iteration.assign(O(this.iteration,1)),this.accBeta1.assign(N(this.accBeta1,this.beta1))}),this.incrementIterations()}dispose(){this.accBeta1.dispose(),this.iteration.dispose(),this.accumulatedFirstMoment!=null&&ut(this.accumulatedFirstMoment.map(t=>t.variable)),this.accumulatedWeightedInfNorm!=null&&ut(this.accumulatedWeightedInfNorm.map(t=>t.variable))}async getWeights(){throw new Error("getWeights() is not implemented for Adamax yet.")}async setWeights(t){throw new Error("setWeights() is not implemented for Adamax yet.")}getConfig(){return{learningRate:this.learningRate,beta1:this.beta1,beta2:this.beta2,epsilon:this.epsilon,decay:this.decay}}static fromConfig(t,e){return new t(e.learningRate,e.beta1,e.beta2,e.epsilon,e.decay)}}/**
+ */class Yl extends qe{static get className(){return"Adamax"}constructor(t,e,s,r=null,o=0){super(),this.learningRate=t,this.beta1=e,this.beta2=s,this.epsilon=r,this.decay=o,this.accumulatedFirstMoment=[],this.accumulatedWeightedInfNorm=[],k(()=>{this.iteration=Yt(0).variable(),this.accBeta1=Yt(e).variable()}),r==null&&(this.epsilon=A.backend.epsilon())}applyGradients(t){const e=Array.isArray(t)?t.map(s=>s.name):Object.keys(t);k(()=>{const s=J(1,this.accBeta1),r=X(-this.learningRate,O(N(this.iteration,this.decay),1));e.forEach((o,i)=>{const a=A.registeredVariables[o],l=!1;this.accumulatedFirstMoment[i]==null&&(this.accumulatedFirstMoment[i]={originalName:`${o}/m`,variable:$e(a).variable(l)}),this.accumulatedWeightedInfNorm[i]==null&&(this.accumulatedWeightedInfNorm[i]={originalName:`${o}/v`,variable:$e(a).variable(l)});const u=Array.isArray(t)?t[i].tensor:t[o];if(u==null)return;const c=this.accumulatedFirstMoment[i].variable,h=this.accumulatedWeightedInfNorm[i].variable,f=O(N(c,this.beta1),N(u,1-this.beta1)),d=N(h,this.beta2),p=Lt(u),g=zn(d,p);c.assign(f),h.assign(g);const m=O(N(X(r,s),X(f,O(g,this.epsilon))),a);a.assign(m)}),this.iteration.assign(O(this.iteration,1)),this.accBeta1.assign(N(this.accBeta1,this.beta1))}),this.incrementIterations()}dispose(){this.accBeta1.dispose(),this.iteration.dispose(),this.accumulatedFirstMoment!=null&&ut(this.accumulatedFirstMoment.map(t=>t.variable)),this.accumulatedWeightedInfNorm!=null&&ut(this.accumulatedWeightedInfNorm.map(t=>t.variable))}async getWeights(){throw new Error("getWeights() is not implemented for Adamax yet.")}async setWeights(t){throw new Error("setWeights() is not implemented for Adamax yet.")}getConfig(){return{learningRate:this.learningRate,beta1:this.beta1,beta2:this.beta2,epsilon:this.epsilon,decay:this.decay}}static fromConfig(t,e){return new t(e.learningRate,e.beta1,e.beta2,e.epsilon,e.decay)}}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2478,7 +2478,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class si extends qe{static get className(){return"SGD"}constructor(t){super(),this.learningRate=t,this.setLearningRate(t)}applyGradients(t){(Array.isArray(t)?t.map(s=>s.name):Object.keys(t)).forEach((s,r)=>{const o=Array.isArray(t)?t[r].tensor:t[s];if(o==null)return;const i=A.registeredVariables[s];k(()=>{const a=O(N(this.c,o),i);i.assign(a)})}),this.incrementIterations()}setLearningRate(t){this.learningRate=t,this.c!=null&&this.c.dispose(),this.c=Ln(Ht(-t))}dispose(){this.c.dispose()}async getWeights(){return[await this.saveIterations()]}async setWeights(t){if(t=await this.extractIterations(t),t.length!==0)throw new Error("SGD optimizer does not have settable weights.")}getConfig(){return{learningRate:this.learningRate}}static fromConfig(t,e){return new t(e.learningRate)}}/**
+ */class si extends qe{static get className(){return"SGD"}constructor(t){super(),this.learningRate=t,this.setLearningRate(t)}applyGradients(t){(Array.isArray(t)?t.map(s=>s.name):Object.keys(t)).forEach((s,r)=>{const o=Array.isArray(t)?t[r].tensor:t[s];if(o==null)return;const i=A.registeredVariables[s];k(()=>{const a=O(N(this.c,o),i);i.assign(a)})}),this.incrementIterations()}setLearningRate(t){this.learningRate=t,this.c!=null&&this.c.dispose(),this.c=Ln(Yt(-t))}dispose(){this.c.dispose()}async getWeights(){return[await this.saveIterations()]}async setWeights(t){if(t=await this.extractIterations(t),t.length!==0)throw new Error("SGD optimizer does not have settable weights.")}getConfig(){return{learningRate:this.learningRate}}static fromConfig(t,e){return new t(e.learningRate)}}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2493,7 +2493,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class Xl extends si{static get className(){return"Momentum"}constructor(t,e,s=!1){super(t),this.learningRate=t,this.momentum=e,this.useNesterov=s,this.accumulations=[],this.m=Ht(this.momentum)}applyGradients(t){(Array.isArray(t)?t.map(s=>s.name):Object.keys(t)).forEach((s,r)=>{const o=A.registeredVariables[s];this.accumulations[r]==null&&(this.accumulations[r]={originalName:`${s}/momentum`,variable:k(()=>$e(o).variable(!1))});const i=this.accumulations[r].variable,a=Array.isArray(t)?t[r].tensor:t[s];a!=null&&k(()=>{let l;const u=O(N(this.m,i),a);this.useNesterov?l=O(N(this.c,O(a,N(u,this.m))),o):l=O(N(this.c,u),o),i.assign(u),o.assign(l)})}),this.incrementIterations()}dispose(){this.m.dispose(),this.accumulations!=null&&ut(this.accumulations.map(t=>t.variable))}setMomentum(t){this.momentum=t}async getWeights(){return[await this.saveIterations()].concat(this.accumulations.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(t){t=await this.extractIterations(t);const e=!1;this.accumulations=t.map(s=>({originalName:s.name,variable:s.tensor.variable(e)}))}getConfig(){return{learningRate:this.learningRate,momentum:this.momentum,useNesterov:this.useNesterov}}static fromConfig(t,e){return new t(e.learningRate,e.momentum,e.useNesterov)}}/**
+ */class Xl extends si{static get className(){return"Momentum"}constructor(t,e,s=!1){super(t),this.learningRate=t,this.momentum=e,this.useNesterov=s,this.accumulations=[],this.m=Yt(this.momentum)}applyGradients(t){(Array.isArray(t)?t.map(s=>s.name):Object.keys(t)).forEach((s,r)=>{const o=A.registeredVariables[s];this.accumulations[r]==null&&(this.accumulations[r]={originalName:`${s}/momentum`,variable:k(()=>$e(o).variable(!1))});const i=this.accumulations[r].variable,a=Array.isArray(t)?t[r].tensor:t[s];a!=null&&k(()=>{let l;const u=O(N(this.m,i),a);this.useNesterov?l=O(N(this.c,O(a,N(u,this.m))),o):l=O(N(this.c,u),o),i.assign(u),o.assign(l)})}),this.incrementIterations()}dispose(){this.m.dispose(),this.accumulations!=null&&ut(this.accumulations.map(t=>t.variable))}setMomentum(t){this.momentum=t}async getWeights(){return[await this.saveIterations()].concat(this.accumulations.map(t=>({name:t.originalName,tensor:t.variable})))}async setWeights(t){t=await this.extractIterations(t);const e=!1;this.accumulations=t.map(s=>({originalName:s.name,variable:s.tensor.variable(e)}))}getConfig(){return{learningRate:this.learningRate,momentum:this.momentum,useNesterov:this.useNesterov}}static fromConfig(t,e){return new t(e.learningRate,e.momentum,e.useNesterov)}}/**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2598,7 +2598,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function uy(n,t){const e=n[0].length;n.forEach((r,o)=>{w(r.length===e,()=>`Error in concat${e}D: rank of tensors[${o}] must be the same as the rank of the rest (${e})`)}),w(t>=0&&t<e,()=>`Error in concat${e}D: axis must be between 0 and ${e-1}.`);const s=n[0];n.forEach((r,o)=>{for(let i=0;i<e;i++)w(i===t||r[i]===s[i],()=>`Error in concat${e}D: Shape of tensors[${o}] (${r}) does not match the shape of the rest (${s}) along the non-concatenated axis ${o}.`)})}function ms(n,t){const e=n[0].slice();for(let s=1;s<n.length;s++)e[t]+=n[s][t];return e}/**
+ */function uy(n,t){const e=n[0].length;n.forEach((r,o)=>{w(r.length===e,()=>`Error in concat${e}D: rank of tensors[${o}] must be the same as the rank of the rest (${e})`)}),w(t>=0&&t<e,()=>`Error in concat${e}D: axis must be between 0 and ${e-1}.`);const s=n[0];n.forEach((r,o)=>{for(let i=0;i<e;i++)w(i===t||r[i]===s[i],()=>`Error in concat${e}D: Shape of tensors[${o}] (${r}) does not match the shape of the rest (${s}) along the non-concatenated axis ${o}.`)})}function gs(n,t){const e=n[0].slice();for(let s=1;s<n.length;s++)e[t]+=n[s][t];return e}/**
  * @license
  * Copyright 2022 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2720,7 +2720,7 @@
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function tu(n){try{return n.map(t=>Ks(t))}catch(t){throw new Error(`Failed to decode encoded string bytes into utf-8, error: ${t}`)}}function Ry(n){return n.map(t=>nn(t))}/**
+ */function tu(n){try{return n.map(t=>Ys(t))}catch(t){throw new Error(`Failed to decode encoded string bytes into utf-8, error: ${t}`)}}function Ry(n){return n.map(t=>sn(t))}/**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -2759,12 +2759,12 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function pr(n,t){if(Array.isArray(n)){let e=[];for(let s=0;s<t;s++)e=e.concat(n);return e}else{const e=new Array(t);return e.fill(n),e}}function Ae(n,t){if(!n)throw new ri(t)}function eu(n,t){let e=0;for(const s of n)s===t&&e++;return e}function Ut(n){return n.length===1?n[0]:n}function nt(n){return Array.isArray(n)?n:[n]}function Le(n){const e=n.replace(/(.)([A-Z][a-z0-9]+)/g,"$1_$2").replace(/([a-z])([A-Z])/g,"$1_$2").toLowerCase();return e[0]!=="_"?e:"private"+e}function pn(n){return n.length<=1||n.indexOf("_")===-1?n:n.replace(/[_]+(\w|$)/g,(t,e)=>e.toUpperCase())}let re={};function oi(n){if(n==null)return null;const t={};return t.className=n.getClassName(),t.config=n.getConfig(),t}function ii(n){if(!(n==null||typeof n!="object"))if(Array.isArray(n))n.forEach(t=>ii(t));else{const t=Object.keys(n);for(const e of t){const s=n[e];s!=null&&typeof s=="object"&&(!Array.isArray(s)&&s.type==="ndarray"&&typeof s.value=="number"?n[e]=s.value:ii(s))}}}function gs(n,t={},e={},s="object",r=!1){if(typeof n=="string"){const o=n;let i;if(o in e)i=e[o];else if(o in re)i=re[o];else if(i=t[o],i==null)throw new I(`Unknown ${s}: ${n}. This may be due to one of the following reasons:
+ */function mr(n,t){if(Array.isArray(n)){let e=[];for(let s=0;s<t;s++)e=e.concat(n);return e}else{const e=new Array(t);return e.fill(n),e}}function Ae(n,t){if(!n)throw new ri(t)}function eu(n,t){let e=0;for(const s of n)s===t&&e++;return e}function Ut(n){return n.length===1?n[0]:n}function nt(n){return Array.isArray(n)?n:[n]}function Le(n){const e=n.replace(/(.)([A-Z][a-z0-9]+)/g,"$1_$2").replace(/([a-z])([A-Z])/g,"$1_$2").toLowerCase();return e[0]!=="_"?e:"private"+e}function mn(n){return n.length<=1||n.indexOf("_")===-1?n:n.replace(/[_]+(\w|$)/g,(t,e)=>e.toUpperCase())}let re={};function oi(n){if(n==null)return null;const t={};return t.className=n.getClassName(),t.config=n.getConfig(),t}function ii(n){if(!(n==null||typeof n!="object"))if(Array.isArray(n))n.forEach(t=>ii(t));else{const t=Object.keys(n);for(const e of t){const s=n[e];s!=null&&typeof s=="object"&&(!Array.isArray(s)&&s.type==="ndarray"&&typeof s.value=="number"?n[e]=s.value:ii(s))}}}function bs(n,t={},e={},s="object",r=!1){if(typeof n=="string"){const o=n;let i;if(o in e)i=e[o];else if(o in re)i=re[o];else if(i=t[o],i==null)throw new I(`Unknown ${s}: ${n}. This may be due to one of the following reasons:
 1. The ${s} is defined in Python, in which case it needs to be ported to TensorFlow.js or your JavaScript code.
 2. The custom ${s} is defined in JavaScript, but is not registered properly with tf.serialization.registerClass().`);return i}else{const o=n;if(o.className==null||o.config==null)throw new I(`${s}: Improper config format: ${JSON.stringify(o)}.
 'className' and 'config' must set.`);const i=o.className;let a,l;if(i in e?[a,l]=e[i]:i in re?[a,l]=re.className:i in t&&([a,l]=t[i]),a==null)throw new I(`Unknown ${s}: ${i}. This may be due to one of the following reasons:
 1. The ${s} is defined in Python, in which case it needs to be ported to TensorFlow.js or your JavaScript code.
-2. The custom ${s} is defined in JavaScript, but is not registered properly with tf.serialization.registerClass().`);if(l!=null){const u={};for(const d of Object.keys(re))u[d]=re[d];for(const d of Object.keys(e))u[d]=e[d];const c=o.config;c.customObjects=u;const h=Object.assign({},re);for(const d of Object.keys(e))re[d]=e[d];ii(o.config);const f=l(a,o.config,e,r);return re=Object.assign({},h),f}else{const u=Object.assign({},re);for(const h of Object.keys(e))re[h]=e[h];const c=new a(o.config);return re=Object.assign({},u),c}}}function By(n,t){return n<t?-1:n>t?1:0}function mr(n,t){return-1*By(n,t)}function mn(n){if(n==null)return n;const t=[];for(const e of n)t.indexOf(e)===-1&&t.push(e);return t}function Fy(n){if(n==null)throw new I(`Invalid value in obj: ${JSON.stringify(n)}`);for(const t in n)if(n.hasOwnProperty(t))return!1;return!0}function qn(n,t,e){if(e!=null&&n.indexOf(e)<0)throw new I(`${e} is not a valid ${t}.  Valid values are ${n} or null/undefined.`)}function ai(n,t,e=0,s=1/0){return Ae(e>=0),Ae(s>=e),Array.isArray(n)&&n.length>=e&&n.length<=s&&n.every(r=>typeof r===t)}function Me(n,t){Array.isArray(n)?(w(n.length>0,()=>`${t} is unexpectedly an empty array.`),n.forEach((e,s)=>Me(e,`element ${s+1} of ${t}`))):w(Number.isInteger(n)&&n>0,()=>`Expected ${t} to be a positive integer, but got ${nu(n)}.`)}function nu(n){return n===null?"null":Array.isArray(n)?"["+n.map(t=>nu(t)).join(",")+"]":typeof n=="string"?`"${n}"`:`${n}`}function zy(n,t,e){let s=e!=null?e():Nn(),r;return(...i)=>{const a=e!=null?e():Nn();return a-s<t||(s=a,r=n(...i)),r}}function Uy(n){return n==="relu"?"relu":n==="linear"?"linear":n==="elu"?"elu":null}/**
+2. The custom ${s} is defined in JavaScript, but is not registered properly with tf.serialization.registerClass().`);if(l!=null){const u={};for(const d of Object.keys(re))u[d]=re[d];for(const d of Object.keys(e))u[d]=e[d];const c=o.config;c.customObjects=u;const h=Object.assign({},re);for(const d of Object.keys(e))re[d]=e[d];ii(o.config);const f=l(a,o.config,e,r);return re=Object.assign({},h),f}else{const u=Object.assign({},re);for(const h of Object.keys(e))re[h]=e[h];const c=new a(o.config);return re=Object.assign({},u),c}}}function By(n,t){return n<t?-1:n>t?1:0}function gr(n,t){return-1*By(n,t)}function gn(n){if(n==null)return n;const t=[];for(const e of n)t.indexOf(e)===-1&&t.push(e);return t}function Fy(n){if(n==null)throw new I(`Invalid value in obj: ${JSON.stringify(n)}`);for(const t in n)if(n.hasOwnProperty(t))return!1;return!0}function qn(n,t,e){if(e!=null&&n.indexOf(e)<0)throw new I(`${e} is not a valid ${t}.  Valid values are ${n} or null/undefined.`)}function ai(n,t,e=0,s=1/0){return Ae(e>=0),Ae(s>=e),Array.isArray(n)&&n.length>=e&&n.length<=s&&n.every(r=>typeof r===t)}function Me(n,t){Array.isArray(n)?(w(n.length>0,()=>`${t} is unexpectedly an empty array.`),n.forEach((e,s)=>Me(e,`element ${s+1} of ${t}`))):w(Number.isInteger(n)&&n>0,()=>`Expected ${t} to be a positive integer, but got ${nu(n)}.`)}function nu(n){return n===null?"null":Array.isArray(n)?"["+n.map(t=>nu(t)).join(",")+"]":typeof n=="string"?`"${n}"`:`${n}`}function zy(n,t,e){let s=e!=null?e():Nn(),r;return(...i)=>{const a=e!=null?e():Nn();return a-s<t||(s=a,r=n(...i)),r}}function Uy(n){return n==="relu"?"relu":n==="linear"?"linear":n==="elu"?"elu":null}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2772,7 +2772,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */const jn=new Map;function mt(n){qn(Py,"DataFormat",n)}function Wy(n){qn(Ly,"InterpolationFormat",n)}function oe(n){qn(My,"PaddingMode",n)}function su(n){qn(Oy,"PoolMode",n)}const bs=[],ru="/";function gr(n,t){bs.push(n);try{const e=t();return bs.pop(),e}catch(e){throw bs.pop(),e}}function Gy(){return bs.length===0?"":bs.join(ru)+ru}function ou(n){if(!au(n))throw new Error("Not a valid tensor name: '"+n+"'");return Gy()+n}function iu(n){if(!au(n))throw new Error("Not a valid tensor name: '"+n+"'");jn.has(n)||jn.set(n,0);const t=jn.get(n);if(jn.set(n,jn.get(n)+1),t>0){const e=`${n}_${t}`;return jn.set(e,1),e}else return n}const Vy=new RegExp(/^[A-Za-z0-9][-A-Za-z0-9\._\/]*$/);function au(n){return!!n.match(Vy)}/**
+ */const jn=new Map;function mt(n){qn(Py,"DataFormat",n)}function Wy(n){qn(Ly,"InterpolationFormat",n)}function oe(n){qn(My,"PaddingMode",n)}function su(n){qn(Oy,"PoolMode",n)}const ys=[],ru="/";function br(n,t){ys.push(n);try{const e=t();return ys.pop(),e}catch(e){throw ys.pop(),e}}function Gy(){return ys.length===0?"":ys.join(ru)+ru}function ou(n){if(!au(n))throw new Error("Not a valid tensor name: '"+n+"'");return Gy()+n}function iu(n){if(!au(n))throw new Error("Not a valid tensor name: '"+n+"'");jn.has(n)||jn.set(n,0);const t=jn.get(n);if(jn.set(n,jn.get(n)+1),t>0){const e=`${n}_${t}`;return jn.set(e,1),e}else return n}const Vy=new RegExp(/^[A-Za-z0-9][-A-Za-z0-9\._\/]*$/);function au(n){return!!n.match(Vy)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2780,7 +2780,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function qy(n){return n===parseInt(n.toString(),10)}function ys(n,t,e){t==null&&(t=0),e==null&&(e=n.length);let s=1;for(let r=t;r<e;++r)s*=n[r];return s}function lu(n){if(n.length===0)return Number.NaN;let t=Number.NEGATIVE_INFINITY;for(let e=0;e<n.length;e++){const s=n[e];s>t&&(t=s)}return t}function br(n,t){if(t<n)throw new I(`end (${t}) < begin (${n}) is forbidden.`);const e=[];for(let s=n;s<t;++s)e.push(s);return e}/**
+ */function qy(n){return n===parseInt(n.toString(),10)}function ws(n,t,e){t==null&&(t=0),e==null&&(e=n.length);let s=1;for(let r=t;r<e;++r)s*=n[r];return s}function lu(n){if(n.length===0)return Number.NaN;let t=Number.NEGATIVE_INFINITY;for(let e=0;e<n.length;e++){const s=n[e];s>t&&(t=s)}return t}function yr(n,t){if(t<n)throw new I(`end (${t}) < begin (${n}) is forbidden.`);const e=[];for(let s=n;s<t;++s)e.push(s);return e}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2796,7 +2796,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function uu(n,t){return ot(n,t)}function ui(n,t=-1){const e=n.shape.slice();return t<0&&(t=e.length+t+1),e.splice(t,0,1),L(n,e)}function jy(n){const t=[ys(n.shape)];return L(n,t)}function gn(n,t,e){return k(()=>{switch(n.rank){case 1:return Qo(n,t,e);case 2:return Fl(n,[t,0],[e,n.shape[1]]);case 3:return ti(n,[t,0,0],[e,n.shape[1],n.shape[2]]);case 4:return cs(n,[t,0,0,0],[e,n.shape[1],n.shape[2],n.shape[3]]);case 5:return Et(n,[t,0,0,0,0],[e,n.shape[1],n.shape[2],n.shape[3],n.shape[4]]);case 6:return Et(n,[t,0,0,0,0,0],[e,n.shape[1],n.shape[2],n.shape[3],n.shape[4],n.shape[5]]);default:throw new I(`sliceAlongFirstAxis() received an unsupported tensor rank: ${n.rank}`)}})}function ci(n,t,e){return k(()=>{switch(n.rank){case 1:return Qo(n,t,e);case 2:return Fl(n,[0,t],[n.shape[0],e]);case 3:return ti(n,[0,0,t],[n.shape[0],n.shape[1],e]);case 4:return cs(n,[0,0,0,t],[n.shape[0],n.shape[1],n.shape[2],e]);default:throw new I(`sliceAlongLastAxis() received an unsupported tensor rank: ${n.rank}`)}})}function yr(n,t,e,s){return k(()=>{switch(n.rank){case 1:return Qo(n,t,e);case 2:switch(s){case 1:return gn(n,t,e);case 2:return ci(n,t,e);default:throw new I(`The axis is not within the rank of the tensor ${s}`)}case 3:switch(s){case 1:return gn(n,t,e);case 2:return ti(n,[0,t,0],[n.shape[0],e,n.shape[2]]);case 3:return ci(n,t,e);default:throw new I(`The axis is not within the rank of the tensor ${s}`)}case 4:switch(s){case 1:return gn(n,t,e);case 2:return cs(n,[0,t,0,0],[n.shape[0],e,n.shape[2],n.shape[3]]);case 3:return cs(n,[0,0,t,0],[n.shape[0],n.shape[1],e,n.shape[3]]);case 4:return ci(n,t,e);default:throw new I(`The axis is not within the rank of the tensor ${s}`)}default:throw new I(`sliceAlongLastAxis() received an unsupported tensor rank: ${n.rank}`)}})}function Hy(n,t=-1){let e;return t<0&&(e=n[0].rank,e!==0?t=e:t=0),t===n[0].rank&&(t=-1),on(n,t)}function cu(n,t=0,e=1,s,r){return A0(n,t,e,s,r)}function Ky(n,t,e){return k(()=>(Array.isArray(t)?t=Rt(t,"int32"):t=ot(t,"int32"),Ng(n,t,e)))}function ws(n){return N(n,n)}function Yy(n,t,e){const s=t.shape;if(t.rank!==1&&t.rank!==n)throw new I(`Unexpected bias dimensions: ${t.rank}; expected it to be 1 or ${n}`);if(n===5){if(e==="channelsFirst")return s.length===1?L(t,[1,s[0],1,1,1]):L(t,[1,s[3],s[0],s[1],s[2]]);if(e==="channelsLast")return s.length===1?L(t,[1,1,1,1,s[0]]):L(t,[1].concat(s))}else if(n===4){if(e==="channelsFirst")return s.length===1?L(t,[1,s[0],1,1]):L(t,[1,s[2],s[0],s[1]]);if(e==="channelsLast")return s.length===1?L(t,[1,1,1,s[0]]):L(t,[1].concat(s))}else if(n===3){if(e==="channelsFirst")return s.length===1?L(t,[1,s[0],1]):L(t,[1,s[1],s[0]]);if(e==="channelsLast")return s.length===1?L(t,[1,1,s[0]]):L(t,[1].concat(s))}else if(n<3)return t;throw new I(`Unsupported input rank by biasAdd: ${t.rank}`)}function xs(n,t,e){return k(()=>(e==null&&(e=Hn()),mt(e),O(n,Yy(n.rank,t,e))))}function Xy(n,t=1){if(t!==1)throw new Z(`Support for alpha values other than 1 (${t}) is not implemented yet.`);return Al(n)}function Jy(n){return k(()=>X(n,O(Lt(n),1)))}function Zy(n){return k(()=>{const t=O(.5,N(.2,n));return he(t,0,1)})}/**
+ */function uu(n,t){return ot(n,t)}function ui(n,t=-1){const e=n.shape.slice();return t<0&&(t=e.length+t+1),e.splice(t,0,1),L(n,e)}function jy(n){const t=[ws(n.shape)];return L(n,t)}function bn(n,t,e){return k(()=>{switch(n.rank){case 1:return Qo(n,t,e);case 2:return Fl(n,[t,0],[e,n.shape[1]]);case 3:return ti(n,[t,0,0],[e,n.shape[1],n.shape[2]]);case 4:return hs(n,[t,0,0,0],[e,n.shape[1],n.shape[2],n.shape[3]]);case 5:return Et(n,[t,0,0,0,0],[e,n.shape[1],n.shape[2],n.shape[3],n.shape[4]]);case 6:return Et(n,[t,0,0,0,0,0],[e,n.shape[1],n.shape[2],n.shape[3],n.shape[4],n.shape[5]]);default:throw new I(`sliceAlongFirstAxis() received an unsupported tensor rank: ${n.rank}`)}})}function ci(n,t,e){return k(()=>{switch(n.rank){case 1:return Qo(n,t,e);case 2:return Fl(n,[0,t],[n.shape[0],e]);case 3:return ti(n,[0,0,t],[n.shape[0],n.shape[1],e]);case 4:return hs(n,[0,0,0,t],[n.shape[0],n.shape[1],n.shape[2],e]);default:throw new I(`sliceAlongLastAxis() received an unsupported tensor rank: ${n.rank}`)}})}function wr(n,t,e,s){return k(()=>{switch(n.rank){case 1:return Qo(n,t,e);case 2:switch(s){case 1:return bn(n,t,e);case 2:return ci(n,t,e);default:throw new I(`The axis is not within the rank of the tensor ${s}`)}case 3:switch(s){case 1:return bn(n,t,e);case 2:return ti(n,[0,t,0],[n.shape[0],e,n.shape[2]]);case 3:return ci(n,t,e);default:throw new I(`The axis is not within the rank of the tensor ${s}`)}case 4:switch(s){case 1:return bn(n,t,e);case 2:return hs(n,[0,t,0,0],[n.shape[0],e,n.shape[2],n.shape[3]]);case 3:return hs(n,[0,0,t,0],[n.shape[0],n.shape[1],e,n.shape[3]]);case 4:return ci(n,t,e);default:throw new I(`The axis is not within the rank of the tensor ${s}`)}default:throw new I(`sliceAlongLastAxis() received an unsupported tensor rank: ${n.rank}`)}})}function Hy(n,t=-1){let e;return t<0&&(e=n[0].rank,e!==0?t=e:t=0),t===n[0].rank&&(t=-1),an(n,t)}function cu(n,t=0,e=1,s,r){return A0(n,t,e,s,r)}function Ky(n,t,e){return k(()=>(Array.isArray(t)?t=Rt(t,"int32"):t=ot(t,"int32"),Ng(n,t,e)))}function xs(n){return N(n,n)}function Yy(n,t,e){const s=t.shape;if(t.rank!==1&&t.rank!==n)throw new I(`Unexpected bias dimensions: ${t.rank}; expected it to be 1 or ${n}`);if(n===5){if(e==="channelsFirst")return s.length===1?L(t,[1,s[0],1,1,1]):L(t,[1,s[3],s[0],s[1],s[2]]);if(e==="channelsLast")return s.length===1?L(t,[1,1,1,1,s[0]]):L(t,[1].concat(s))}else if(n===4){if(e==="channelsFirst")return s.length===1?L(t,[1,s[0],1,1]):L(t,[1,s[2],s[0],s[1]]);if(e==="channelsLast")return s.length===1?L(t,[1,1,1,s[0]]):L(t,[1].concat(s))}else if(n===3){if(e==="channelsFirst")return s.length===1?L(t,[1,s[0],1]):L(t,[1,s[1],s[0]]);if(e==="channelsLast")return s.length===1?L(t,[1,1,s[0]]):L(t,[1].concat(s))}else if(n<3)return t;throw new I(`Unsupported input rank by biasAdd: ${t.rank}`)}function Ss(n,t,e){return k(()=>(e==null&&(e=Hn()),mt(e),O(n,Yy(n.rank,t,e))))}function Xy(n,t=1){if(t!==1)throw new Z(`Support for alpha values other than 1 (${t}) is not implemented yet.`);return Al(n)}function Jy(n){return k(()=>X(n,O(Lt(n),1)))}function Zy(n){return k(()=>{const t=O(.5,N(.2,n));return he(t,0,1)})}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2804,7 +2804,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */class Ct extends Gn{getConfig(){return{}}}class hu extends Ct{apply(t,e=1){return Xy(t,e)}}hu.className="elu",M(hu);class fu extends Ct{apply(t){return L0(t)}}fu.className="selu",M(fu);class du extends Ct{apply(t){return ps(t)}}du.className="relu",M(du);class pu extends Ct{apply(t){return k(()=>ur(6,ps(t)))}}pu.className="relu6",M(pu);class mu extends Ct{apply(t){return t}}mu.className="linear",M(mu);class gu extends Ct{apply(t){return Fo(t)}}gu.className="sigmoid",M(gu);class bu extends Ct{apply(t){return Zy(t)}}bu.className="hardSigmoid",M(bu);class yu extends Ct{apply(t){return qo(t)}}yu.className="softplus",M(yu);class wu extends Ct{apply(t){return Jy(t)}}wu.className="softsign",M(wu);class xu extends Ct{apply(t){return zo(t)}}xu.className="tanh",M(xu);class Su extends Ct{apply(t,e=-1){return zl(t,e)}}Su.className="softmax",M(Su);class $u extends Ct{apply(t,e=-1){return Yg(t,e)}}$u.className="logSoftmax",M($u);class vu extends Ct{apply(t){return k(()=>k(()=>{const e=Math.sqrt(2),s=N(.5,O(1,fg(X(t,e))));return N(t,s)}))}}vu.className="gelu",M(vu);class Iu extends Ct{apply(t){return k(()=>N(.5,N(t,O(1,zo(N(fe(X(2,Math.PI)),O(t,N(.044715,ir(t,3)))))))))}}Iu.className="gelu_new",M(Iu);class Au extends Ct{apply(t){return k(()=>N(t,zo(qo(t))))}}Au.className="mish",M(Au);class Eu extends Ct{apply(t,e=1){return k(()=>N(Fo(N(t,e)),t))}}Eu.className="swish",M(Eu);function Qy(n){return n.getClassName()}function hi(n,t={}){return gs(n,se.getMap().classNameMap,t,"activation")}function tw(n){if(n==null){const t={};return t.className="linear",t.config={},hi(t)}if(typeof n=="string"){const t={};return t.className=n,t.config={},hi(t)}else return n instanceof Ct?n:hi(n)}/**
+ */class Ct extends Gn{getConfig(){return{}}}class hu extends Ct{apply(t,e=1){return Xy(t,e)}}hu.className="elu",M(hu);class fu extends Ct{apply(t){return L0(t)}}fu.className="selu",M(fu);class du extends Ct{apply(t){return ms(t)}}du.className="relu",M(du);class pu extends Ct{apply(t){return k(()=>cr(6,ms(t)))}}pu.className="relu6",M(pu);class mu extends Ct{apply(t){return t}}mu.className="linear",M(mu);class gu extends Ct{apply(t){return Fo(t)}}gu.className="sigmoid",M(gu);class bu extends Ct{apply(t){return Zy(t)}}bu.className="hardSigmoid",M(bu);class yu extends Ct{apply(t){return qo(t)}}yu.className="softplus",M(yu);class wu extends Ct{apply(t){return Jy(t)}}wu.className="softsign",M(wu);class xu extends Ct{apply(t){return zo(t)}}xu.className="tanh",M(xu);class Su extends Ct{apply(t,e=-1){return zl(t,e)}}Su.className="softmax",M(Su);class $u extends Ct{apply(t,e=-1){return Yg(t,e)}}$u.className="logSoftmax",M($u);class vu extends Ct{apply(t){return k(()=>k(()=>{const e=Math.sqrt(2),s=N(.5,O(1,fg(X(t,e))));return N(t,s)}))}}vu.className="gelu",M(vu);class Iu extends Ct{apply(t){return k(()=>N(.5,N(t,O(1,zo(N(fe(X(2,Math.PI)),O(t,N(.044715,ar(t,3)))))))))}}Iu.className="gelu_new",M(Iu);class Au extends Ct{apply(t){return k(()=>N(t,zo(qo(t))))}}Au.className="mish",M(Au);class Eu extends Ct{apply(t,e=1){return k(()=>N(Fo(N(t,e)),t))}}Eu.className="swish",M(Eu);function Qy(n){return n.getClassName()}function hi(n,t={}){return bs(n,se.getMap().classNameMap,t,"activation")}function tw(n){if(n==null){const t={};return t.className="linear",t.config={},hi(t)}if(typeof n=="string"){const t={};return t.className=n,t.config={},hi(t)}else return n instanceof Ct?n:hi(n)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2812,7 +2812,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function fi(n,t){return k(()=>fe(et(N(n,n),t,!0)))}class Ss extends Gn{getConfig(){return{}}}class Cu extends Ss{constructor(t){super(),this.defaultMaxValue=2,this.defaultAxis=0,this.maxValue=t.maxValue!=null?t.maxValue:this.defaultMaxValue,this.axis=t.axis!=null?t.axis:this.defaultAxis}apply(t){return k(()=>{const e=fi(t,this.axis),s=he(e,0,this.maxValue);return N(t,X(s,O(gt(),e)))})}getConfig(){return{maxValue:this.maxValue,axis:this.axis}}}Cu.className="MaxNorm",M(Cu);class ku extends Ss{constructor(t){super(),this.defaultAxis=0,this.axis=t.axis!=null?t.axis:this.defaultAxis}apply(t){return k(()=>X(t,O(gt(),fi(t,this.axis))))}getConfig(){return{axis:this.axis}}}ku.className="UnitNorm",M(ku);class _u extends Ss{apply(t){return ps(t)}}_u.className="NonNeg",M(_u);class Tu extends Ss{constructor(t){super(),this.defaultMinValue=0,this.defaultMaxValue=1,this.defaultRate=1,this.defaultAxis=0,this.minValue=t.minValue!=null?t.minValue:this.defaultMinValue,this.maxValue=t.maxValue!=null?t.maxValue:this.defaultMaxValue,this.rate=t.rate!=null?t.rate:this.defaultRate,this.axis=t.axis!=null?t.axis:this.defaultAxis}apply(t){return k(()=>{const e=fi(t,this.axis),s=O(N(this.rate,he(e,this.minValue,this.maxValue)),N(1-this.rate,e));return N(t,X(s,O(gt(),e)))})}getConfig(){return{minValue:this.minValue,maxValue:this.maxValue,rate:this.rate,axis:this.axis}}}Tu.className="MinMaxNorm",M(Tu);const Nu={maxNorm:"MaxNorm",minMaxNorm:"MinMaxNorm",nonNeg:"NonNeg",unitNorm:"UnitNorm"};function wr(n){return oi(n)}function Du(n,t={}){return gs(n,se.getMap().classNameMap,t,"constraint")}function xr(n){if(n==null)return null;if(typeof n=="string"){const e={className:n in Nu?Nu[n]:n,config:{}};return Du(e)}else return n instanceof Ss?n:Du(n)}/**
+ */function fi(n,t){return k(()=>fe(et(N(n,n),t,!0)))}class $s extends Gn{getConfig(){return{}}}class Cu extends $s{constructor(t){super(),this.defaultMaxValue=2,this.defaultAxis=0,this.maxValue=t.maxValue!=null?t.maxValue:this.defaultMaxValue,this.axis=t.axis!=null?t.axis:this.defaultAxis}apply(t){return k(()=>{const e=fi(t,this.axis),s=he(e,0,this.maxValue);return N(t,X(s,O(gt(),e)))})}getConfig(){return{maxValue:this.maxValue,axis:this.axis}}}Cu.className="MaxNorm",M(Cu);class ku extends $s{constructor(t){super(),this.defaultAxis=0,this.axis=t.axis!=null?t.axis:this.defaultAxis}apply(t){return k(()=>X(t,O(gt(),fi(t,this.axis))))}getConfig(){return{axis:this.axis}}}ku.className="UnitNorm",M(ku);class _u extends $s{apply(t){return ms(t)}}_u.className="NonNeg",M(_u);class Tu extends $s{constructor(t){super(),this.defaultMinValue=0,this.defaultMaxValue=1,this.defaultRate=1,this.defaultAxis=0,this.minValue=t.minValue!=null?t.minValue:this.defaultMinValue,this.maxValue=t.maxValue!=null?t.maxValue:this.defaultMaxValue,this.rate=t.rate!=null?t.rate:this.defaultRate,this.axis=t.axis!=null?t.axis:this.defaultAxis}apply(t){return k(()=>{const e=fi(t,this.axis),s=O(N(this.rate,he(e,this.minValue,this.maxValue)),N(1-this.rate,e));return N(t,X(s,O(gt(),e)))})}getConfig(){return{minValue:this.minValue,maxValue:this.maxValue,rate:this.rate,axis:this.axis}}}Tu.className="MinMaxNorm",M(Tu);const Nu={maxNorm:"MaxNorm",minMaxNorm:"MinMaxNorm",nonNeg:"NonNeg",unitNorm:"UnitNorm"};function xr(n){return oi(n)}function Du(n,t={}){return bs(n,se.getMap().classNameMap,t,"constraint")}function Sr(n){if(n==null)return null;if(typeof n=="string"){const e={className:n in Nu?Nu[n]:n,config:{}};return Du(e)}else return n instanceof $s?n:Du(n)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2820,7 +2820,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */let ew=0;function Ru(){return ew++}const Sr={};function di(n=""){return n in Sr||(Sr[n]=0),Sr[n]+=1,n+Sr[n].toString()}/**
+ */let ew=0;function Ru(){return ew++}const $r={};function di(n=""){return n in $r||($r[n]=0),$r[n]+=1,n+$r[n].toString()}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2836,7 +2836,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function rw(n){qn(nw,"FanMode",n)}function ow(n){qn(sw,"Distribution",n)}class Ee extends Gn{fromConfigUsesCustomObjects(){return!1}getConfig(){return{}}}class Pu extends Ee{apply(t,e){return Un(t,e)}}Pu.className="Zeros",M(Pu);class Lu extends Ee{apply(t,e){return jo(t,e)}}Lu.className="Ones",M(Lu);class Mu extends Ee{constructor(t){if(super(),typeof t!="object")throw new I(`Expected argument of type ConstantConfig but got ${t}`);if(t.value===void 0)throw new I(`config must have value set but got ${t}`);this.value=t.value}apply(t,e){return k(()=>N(Ht(this.value),jo(t,e)))}getConfig(){return{value:this.value}}}Mu.className="Constant",M(Mu);class Ou extends Ee{constructor(t){super(),this.DEFAULT_MINVAL=-.05,this.DEFAULT_MAXVAL=.05,this.minval=t.minval||this.DEFAULT_MINVAL,this.maxval=t.maxval||this.DEFAULT_MAXVAL,this.seed=t.seed}apply(t,e){return Bl(t,this.minval,this.maxval,e,this.seed)}getConfig(){return{minval:this.minval,maxval:this.maxval,seed:this.seed}}}Ou.className="RandomUniform",M(Ou);class Bu extends Ee{constructor(t){super(),this.DEFAULT_MEAN=0,this.DEFAULT_STDDEV=.05,this.mean=t.mean||this.DEFAULT_MEAN,this.stddev=t.stddev||this.DEFAULT_STDDEV,this.seed=t.seed}apply(t,e){if(e=e||"float32",e!=="float32"&&e!=="int32")throw new Z(`randomNormal does not support dType ${e}.`);return cu(t,this.mean,this.stddev,e,this.seed)}getConfig(){return{mean:this.mean,stddev:this.stddev,seed:this.seed}}}Bu.className="RandomNormal",M(Bu);class Fu extends Ee{constructor(t){super(),this.DEFAULT_MEAN=0,this.DEFAULT_STDDEV=.05,this.mean=t.mean||this.DEFAULT_MEAN,this.stddev=t.stddev||this.DEFAULT_STDDEV,this.seed=t.seed}apply(t,e){if(e=e||"float32",e!=="float32"&&e!=="int32")throw new Z(`truncatedNormal does not support dType ${e}.`);return Wl(t,this.mean,this.stddev,e,this.seed)}getConfig(){return{mean:this.mean,stddev:this.stddev,seed:this.seed}}}Fu.className="TruncatedNormal",M(Fu);class zu extends Ee{constructor(t){super(),this.gain=t.gain!=null?t.gain:1}apply(t,e){return k(()=>{if(t.length!==2||t[0]!==t[1])throw new I("Identity matrix initializer can only be used for 2D square matrices.");return N(this.gain,Nl(t[0]))})}getConfig(){return{gain:this.gain}}}zu.className="Identity",M(zu);function iw(n,t="channelsLast"){let e,s;if(mt(t),n.length===2)e=n[0],s=n[1];else if([3,4,5].indexOf(n.length)!==-1){if(t==="channelsFirst"){const r=ys(n,2);e=n[1]*r,s=n[0]*r}else if(t==="channelsLast"){const r=ys(n,0,n.length-2);e=n[n.length-2]*r,s=n[n.length-1]*r}}else{const r=ys(n);e=Math.sqrt(r),s=Math.sqrt(r)}return[e,s]}class Wt extends Ee{constructor(t){if(super(),t.scale<0)throw new I(`scale must be a positive float. Got: ${t.scale}`);this.scale=t.scale==null?1:t.scale,this.mode=t.mode==null?"fanIn":t.mode,rw(this.mode),this.distribution=t.distribution==null?"normal":t.distribution,ow(this.distribution),this.seed=t.seed}apply(t,e){const s=iw(t),r=s[0],o=s[1];let i=this.scale;if(this.mode==="fanIn"?i/=Math.max(1,r):this.mode==="fanOut"?i/=Math.max(1,o):i/=Math.max(1,(r+o)/2),this.distribution==="normal"){const a=Math.sqrt(i);if(e=e||"float32",e!=="float32"&&e!=="int32")throw new Z(`${this.getClassName()} does not support dType ${e}.`);return Wl(t,0,a,e,this.seed)}else{const a=Math.sqrt(3*i);return Bl(t,-a,a,e,this.seed)}}getConfig(){return{scale:this.scale,mode:this.mode,distribution:this.distribution,seed:this.seed}}}Wt.className="VarianceScaling",M(Wt);class pi extends Wt{constructor(t){super({scale:1,mode:"fanAvg",distribution:"uniform",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}pi.className="GlorotUniform",M(pi);class mi extends Wt{constructor(t){super({scale:1,mode:"fanAvg",distribution:"normal",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}mi.className="GlorotNormal",M(mi);class gi extends Wt{constructor(t){super({scale:2,mode:"fanIn",distribution:"normal",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}gi.className="HeNormal",M(gi);class bi extends Wt{constructor(t){super({scale:2,mode:"fanIn",distribution:"uniform",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}bi.className="HeUniform",M(bi);class yi extends Wt{constructor(t){super({scale:1,mode:"fanIn",distribution:"normal",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}yi.className="LeCunNormal",M(yi);class wi extends Wt{constructor(t){super({scale:1,mode:"fanIn",distribution:"uniform",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}wi.className="LeCunUniform",M(wi);class Uu extends Ee{constructor(t){super(),this.DEFAULT_GAIN=1,this.ELEMENTS_WARN_SLOW=2e3,this.gain=t.gain==null?this.DEFAULT_GAIN:t.gain,this.seed=t.seed}apply(t,e){return k(()=>{if(t.length<2)throw new Z("Shape must be at least 2D.");if(e!=="int32"&&e!=="float32"&&e!==void 0)throw new TypeError(`Unsupported data type ${e}.`);e=e;const s=z(t.slice(0,-1)),r=t[t.length-1],o=s*r;o>this.ELEMENTS_WARN_SLOW&&console.warn(`Orthogonal initializer is being called on a matrix with more than ${this.ELEMENTS_WARN_SLOW} (${o}) elements: Slowness may result.`);const i=[Math.max(r,s),Math.min(r,s)],a=cu(i,0,1,e,this.seed),l=Jb.qr(a,!1);let u=l[0];const h=l[1].flatten().stridedSlice([0],[Math.min(r,s)*Math.min(r,s)],[Math.min(r,s)+1]);return u=N(u,h.sign()),s<r&&(u=u.transpose()),N(Ht(this.gain),u.reshape(t))})}getConfig(){return{gain:this.gain,seed:this.seed}}}Uu.className="Orthogonal",M(Uu);const Wu={constant:"Constant",glorotNormal:"GlorotNormal",glorotUniform:"GlorotUniform",heNormal:"HeNormal",heUniform:"HeUniform",identity:"Identity",leCunNormal:"LeCunNormal",leCunUniform:"LeCunUniform",ones:"Ones",orthogonal:"Orthogonal",randomNormal:"RandomNormal",randomUniform:"RandomUniform",truncatedNormal:"TruncatedNormal",varianceScaling:"VarianceScaling",zeros:"Zeros"};function Gu(n,t={}){return gs(n,se.getMap().classNameMap,t,"initializer")}function $r(n){return oi(n)}function $s(n){if(typeof n=="string"){const t=n in Wu?Wu[n]:n;if(t==="GlorotNormal")return new mi;if(t==="GlorotUniform")return new pi;if(t==="HeNormal")return new gi;if(t==="HeUniform")return new bi;if(t==="LeCunNormal")return new yi;if(t==="LeCunUniform")return new wi;{const e={};return e.className=t,e.config={},Gu(e)}}else return n instanceof Ee?n:Gu(n)}/**
+ */function rw(n){qn(nw,"FanMode",n)}function ow(n){qn(sw,"Distribution",n)}class Ee extends Gn{fromConfigUsesCustomObjects(){return!1}getConfig(){return{}}}class Pu extends Ee{apply(t,e){return Un(t,e)}}Pu.className="Zeros",M(Pu);class Lu extends Ee{apply(t,e){return jo(t,e)}}Lu.className="Ones",M(Lu);class Mu extends Ee{constructor(t){if(super(),typeof t!="object")throw new I(`Expected argument of type ConstantConfig but got ${t}`);if(t.value===void 0)throw new I(`config must have value set but got ${t}`);this.value=t.value}apply(t,e){return k(()=>N(Yt(this.value),jo(t,e)))}getConfig(){return{value:this.value}}}Mu.className="Constant",M(Mu);class Ou extends Ee{constructor(t){super(),this.DEFAULT_MINVAL=-.05,this.DEFAULT_MAXVAL=.05,this.minval=t.minval||this.DEFAULT_MINVAL,this.maxval=t.maxval||this.DEFAULT_MAXVAL,this.seed=t.seed}apply(t,e){return Bl(t,this.minval,this.maxval,e,this.seed)}getConfig(){return{minval:this.minval,maxval:this.maxval,seed:this.seed}}}Ou.className="RandomUniform",M(Ou);class Bu extends Ee{constructor(t){super(),this.DEFAULT_MEAN=0,this.DEFAULT_STDDEV=.05,this.mean=t.mean||this.DEFAULT_MEAN,this.stddev=t.stddev||this.DEFAULT_STDDEV,this.seed=t.seed}apply(t,e){if(e=e||"float32",e!=="float32"&&e!=="int32")throw new Z(`randomNormal does not support dType ${e}.`);return cu(t,this.mean,this.stddev,e,this.seed)}getConfig(){return{mean:this.mean,stddev:this.stddev,seed:this.seed}}}Bu.className="RandomNormal",M(Bu);class Fu extends Ee{constructor(t){super(),this.DEFAULT_MEAN=0,this.DEFAULT_STDDEV=.05,this.mean=t.mean||this.DEFAULT_MEAN,this.stddev=t.stddev||this.DEFAULT_STDDEV,this.seed=t.seed}apply(t,e){if(e=e||"float32",e!=="float32"&&e!=="int32")throw new Z(`truncatedNormal does not support dType ${e}.`);return Wl(t,this.mean,this.stddev,e,this.seed)}getConfig(){return{mean:this.mean,stddev:this.stddev,seed:this.seed}}}Fu.className="TruncatedNormal",M(Fu);class zu extends Ee{constructor(t){super(),this.gain=t.gain!=null?t.gain:1}apply(t,e){return k(()=>{if(t.length!==2||t[0]!==t[1])throw new I("Identity matrix initializer can only be used for 2D square matrices.");return N(this.gain,Nl(t[0]))})}getConfig(){return{gain:this.gain}}}zu.className="Identity",M(zu);function iw(n,t="channelsLast"){let e,s;if(mt(t),n.length===2)e=n[0],s=n[1];else if([3,4,5].indexOf(n.length)!==-1){if(t==="channelsFirst"){const r=ws(n,2);e=n[1]*r,s=n[0]*r}else if(t==="channelsLast"){const r=ws(n,0,n.length-2);e=n[n.length-2]*r,s=n[n.length-1]*r}}else{const r=ws(n);e=Math.sqrt(r),s=Math.sqrt(r)}return[e,s]}class Wt extends Ee{constructor(t){if(super(),t.scale<0)throw new I(`scale must be a positive float. Got: ${t.scale}`);this.scale=t.scale==null?1:t.scale,this.mode=t.mode==null?"fanIn":t.mode,rw(this.mode),this.distribution=t.distribution==null?"normal":t.distribution,ow(this.distribution),this.seed=t.seed}apply(t,e){const s=iw(t),r=s[0],o=s[1];let i=this.scale;if(this.mode==="fanIn"?i/=Math.max(1,r):this.mode==="fanOut"?i/=Math.max(1,o):i/=Math.max(1,(r+o)/2),this.distribution==="normal"){const a=Math.sqrt(i);if(e=e||"float32",e!=="float32"&&e!=="int32")throw new Z(`${this.getClassName()} does not support dType ${e}.`);return Wl(t,0,a,e,this.seed)}else{const a=Math.sqrt(3*i);return Bl(t,-a,a,e,this.seed)}}getConfig(){return{scale:this.scale,mode:this.mode,distribution:this.distribution,seed:this.seed}}}Wt.className="VarianceScaling",M(Wt);class pi extends Wt{constructor(t){super({scale:1,mode:"fanAvg",distribution:"uniform",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}pi.className="GlorotUniform",M(pi);class mi extends Wt{constructor(t){super({scale:1,mode:"fanAvg",distribution:"normal",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}mi.className="GlorotNormal",M(mi);class gi extends Wt{constructor(t){super({scale:2,mode:"fanIn",distribution:"normal",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}gi.className="HeNormal",M(gi);class bi extends Wt{constructor(t){super({scale:2,mode:"fanIn",distribution:"uniform",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}bi.className="HeUniform",M(bi);class yi extends Wt{constructor(t){super({scale:1,mode:"fanIn",distribution:"normal",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}yi.className="LeCunNormal",M(yi);class wi extends Wt{constructor(t){super({scale:1,mode:"fanIn",distribution:"uniform",seed:t==null?null:t.seed})}getClassName(){return Wt.className}}wi.className="LeCunUniform",M(wi);class Uu extends Ee{constructor(t){super(),this.DEFAULT_GAIN=1,this.ELEMENTS_WARN_SLOW=2e3,this.gain=t.gain==null?this.DEFAULT_GAIN:t.gain,this.seed=t.seed}apply(t,e){return k(()=>{if(t.length<2)throw new Z("Shape must be at least 2D.");if(e!=="int32"&&e!=="float32"&&e!==void 0)throw new TypeError(`Unsupported data type ${e}.`);e=e;const s=z(t.slice(0,-1)),r=t[t.length-1],o=s*r;o>this.ELEMENTS_WARN_SLOW&&console.warn(`Orthogonal initializer is being called on a matrix with more than ${this.ELEMENTS_WARN_SLOW} (${o}) elements: Slowness may result.`);const i=[Math.max(r,s),Math.min(r,s)],a=cu(i,0,1,e,this.seed),l=Jb.qr(a,!1);let u=l[0];const h=l[1].flatten().stridedSlice([0],[Math.min(r,s)*Math.min(r,s)],[Math.min(r,s)+1]);return u=N(u,h.sign()),s<r&&(u=u.transpose()),N(Yt(this.gain),u.reshape(t))})}getConfig(){return{gain:this.gain,seed:this.seed}}}Uu.className="Orthogonal",M(Uu);const Wu={constant:"Constant",glorotNormal:"GlorotNormal",glorotUniform:"GlorotUniform",heNormal:"HeNormal",heUniform:"HeUniform",identity:"Identity",leCunNormal:"LeCunNormal",leCunUniform:"LeCunUniform",ones:"Ones",orthogonal:"Orthogonal",randomNormal:"RandomNormal",randomUniform:"RandomUniform",truncatedNormal:"TruncatedNormal",varianceScaling:"VarianceScaling",zeros:"Zeros"};function Gu(n,t={}){return bs(n,se.getMap().classNameMap,t,"initializer")}function vr(n){return oi(n)}function vs(n){if(typeof n=="string"){const t=n in Wu?Wu[n]:n;if(t==="GlorotNormal")return new mi;if(t==="GlorotUniform")return new pi;if(t==="HeNormal")return new gi;if(t==="HeUniform")return new bi;if(t==="LeCunNormal")return new yi;if(t==="LeCunUniform")return new wi;{const e={};return e.className=t,e.config={},Gu(e)}}else return n instanceof Ee?n:Gu(n)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2844,7 +2844,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function vr(n){return n.length===0?[]:Array.isArray(n[0])?n:[n]}function Gt(n){let t;if(Array.isArray(n)){if(n.length!==1)throw new I(`Expected Tensor length to be 1; got ${n.length}`);t=n[0]}else t=n;return t}function de(n){if(Array.isArray(n)&&Array.isArray(n[0])){if(n.length===1)return n=n,n[0];throw new I(`Expected exactly 1 Shape; got ${n.length}`)}else return n}/**
+ */function Ir(n){return n.length===0?[]:Array.isArray(n[0])?n:[n]}function Gt(n){let t;if(Array.isArray(n)){if(n.length!==1)throw new I(`Expected Tensor length to be 1; got ${n.length}`);t=n[0]}else t=n;return t}function de(n){if(Array.isArray(n)&&Array.isArray(n[0])){if(n.length===1)return n=n,n[0];throw new I(`Expected exactly 1 Shape; got ${n.length}`)}else return n}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2852,7 +2852,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Ir(n){let t=0;for(const e of n)e.shape.length===0?t+=1:t+=e.shape.reduce((s,r)=>s*r);return t}/**
+ */function Ar(n){let t=0;for(const e of n)e.shape.length===0?t+=1:t+=e.shape.reduce((s,r)=>s*r);return t}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2868,7 +2868,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */class Ce{constructor(t){this.dtype=t.dtype,this.shape=t.shape,t.shape!=null?this.ndim=t.shape.length:this.ndim=t.ndim,this.maxNDim=t.maxNDim,this.minNDim=t.minNDim,this.axes=t.axes||{}}}class bn{constructor(t,e,s,r,o,i,a){this.dtype=t,this.shape=e,this.sourceLayer=s,this.inputs=r,this.callArgs=o,this.outputTensorIndex=a,this.id=Ru(),i!=null&&(this.originalName=ou(i),this.name=iu(this.originalName)),this.rank=e.length}}let uw=0;class xi{constructor(t,e){this.callArgs=e,this.id=uw++,this.outboundLayer=t.outboundLayer,this.inboundLayers=t.inboundLayers,this.nodeIndices=t.nodeIndices,this.tensorIndices=t.tensorIndices,this.inputTensors=t.inputTensors,this.outputTensors=t.outputTensors,this.inputMasks=t.inputMasks,this.outputMasks=t.outputMasks,this.inputShapes=t.inputShapes,this.outputShapes=t.outputShapes;for(const s of t.inboundLayers)s!=null&&s.outboundNodes.push(this);t.outboundLayer.inboundNodes.push(this)}getConfig(){const t=[];for(const e of this.inboundLayers)e!=null?t.push(e.name):t.push(null);return{outboundLayer:this.outboundLayer?this.outboundLayer.name:null,inboundLayers:t,nodeIndices:this.nodeIndices,tensorIndices:this.tensorIndices}}}let cw=0;class pe extends Gn{constructor(t={}){super(),this._callHook=null,this._addedWeightNames=[],this._stateful=!1,this.id=cw++,this.activityRegularizer=null,this.inputSpec=null,this.supportsMasking=!1,this._trainableWeights=[],this._nonTrainableWeights=[],this._losses=[],this._updates=[],this._built=!1,this.inboundNodes=[],this.outboundNodes=[];let e=t.name;if(!e){const s=this.getClassName();e=Le(s)+"_"+di(s)}if(this.name=e,this.trainable_=t.trainable==null?!0:t.trainable,t.inputShape!=null||t.batchInputShape!=null){let s;if(t.batchInputShape!=null)s=t.batchInputShape;else if(t.inputShape!=null){let o=null;t.batchSize!=null&&(o=t.batchSize),s=[o].concat(t.inputShape)}this.batchInputShape=s;let r=t.dtype;r==null&&(r=t.inputDType),r==null&&(r="float32"),this.dtype=r}t.weights!=null?this.initialWeights=t.weights:this.initialWeights=null,this._refCount=null,this.fastWeightInitDuringBuild=!1}static nodeKey(t,e){return t.name+"_ib-"+e.toString()}getNodeAtIndex(t,e){if(this.inboundNodes.length===0)throw new He(`The layer has never been called and thus has no defined ${e}.`);if(this.inboundNodes.length<=t)throw new I(`Asked to get ${e} at node ${t}, but the layer has only ${this.inboundNodes.length} inbound nodes.`);return this.inboundNodes[t]}getInputAt(t){return Ut(this.getNodeAtIndex(t,"input").inputTensors)}getOutputAt(t){return Ut(this.getNodeAtIndex(t,"output").outputTensors)}get input(){if(this.inboundNodes.length>1)throw new je(`Layer ${this.name} has multiple inbound nodes, hence the notion of "layer input" is ill-defined. Use \`getInputAt(nodeIndex)\` instead.`);if(this.inboundNodes.length===0)throw new je(`Layer ${this.name} is not connected, no input to return.`);return Ut(this.getNodeAtIndex(0,"input").inputTensors)}get output(){if(this.inboundNodes.length===0)throw new je(`Layer ${this.name} has no inbound nodes.`);if(this.inboundNodes.length>1)throw new je(`Layer ${this.name} has multiple inbound nodes, hence the notion of "layer output" is ill-defined. Use \`getOutputAt(nodeIndex)\` instead.`);return Ut(this.getNodeAtIndex(0,"output").outputTensors)}get losses(){return this._losses}calculateLosses(){return this.losses.map(t=>t())}get updates(){return this._updates}get built(){return this._built}set built(t){this._built=t}get trainable(){return this.trainable_}set trainable(t){this._trainableWeights.forEach(e=>e.trainable=t),this.trainable_=t}get trainableWeights(){return this.trainable_?this._trainableWeights.filter(t=>t.trainable):[]}set trainableWeights(t){this._trainableWeights=t}get nonTrainableWeights(){return this.trainable?this._trainableWeights.filter(t=>!t.trainable).concat(this._nonTrainableWeights):this._trainableWeights.concat(this._nonTrainableWeights)}set nonTrainableWeights(t){this._nonTrainableWeights=t}get weights(){return this.trainableWeights.concat(this.nonTrainableWeights)}get stateful(){return this._stateful}resetStates(){if(!this.stateful)throw new Error("Cannot call the resetStates() method of a non-stateful Layer object.")}assertInputCompatibility(t){const e=nt(t);if(this.inputSpec==null||this.inputSpec.length===0)return;const s=nt(this.inputSpec);if(e.length!==s.length)throw new I(`Layer ${this.name} expects ${s.length} inputs, but it received ${e.length} input tensors. Input received: ${t}`);for(let r=0;r<e.length;r++){const o=e[r],i=s[r];if(i==null)continue;const a=o.rank;if(i.ndim!=null&&a!==i.ndim)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected ndim=${i.ndim}, found ndim=${a}`);if(i.maxNDim!=null&&a>i.maxNDim)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected max_ndim=${i.maxNDim}, found ndim=${a}`);if(i.minNDim!=null&&a<i.minNDim)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected min_ndim=${i.minNDim}, found ndim=${a}.`);if(i.dtype!=null&&o.dtype!==i.dtype)throw new I(`Input ${r} is incompatible with layer ${this.name} : expected dtype=${i.dtype}, found dtype=${o.dtype}.`);if(i.axes){const l=o.shape;for(const u in i.axes){const c=Number(u),h=i.axes[u],f=c>=0?l[c]:l[l.length+c];if(h!=null&&[h,null].indexOf(f)===-1)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected axis ${c} of input shape to have value ${h} but got shape ${l}.`)}}if(i.shape!=null)for(let l=0;l<i.shape.length;++l){const u=i.shape[l],c=o.shape[l];if(u!=null&&c!=null&&u!==c)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected shape=${i.shape}, found shape=${o.shape}.`)}}}call(t,e){return t}invokeCallHook(t,e){this._callHook!=null&&this._callHook(t,e)}setCallHook(t){this._callHook=t}clearCallHook(){this._callHook=null}apply(t,e){e=e||{},this.assertNotDisposed();const s=nt(t),r=dw(t),o=pw(t);if(r===o)throw new I("Arguments to apply() must be all SymbolicTensors or all Tensors");return gr(this.name,()=>{if(!this.built){this.assertInputCompatibility(t);const i=[];for(const a of nt(t))i.push(a.shape);this.build(Ut(i)),this.built=!0,this.initialWeights&&this.setWeights(this.initialWeights),this._refCount===null&&o&&(this._refCount=1)}if(this.assertInputCompatibility(t),o){let i=this.call(t,e);this.supportsMasking&&this.setMaskMetadata(t,i);const a=nt(i),l=[];for(let u of a)s.indexOf(u)!==-1&&(u=u.clone()),l.push(u);if(i=Ut(l),this.activityRegularizer!=null)throw new Z("Layer invocation in the presence of activity regularizer(s) is not supported yet.");return i}else{const i=hw(t),a=this.computeOutputShape(i);let l;const u=fw(t);if(this.warnOnIncompatibleInputShape(Array.isArray(t)?i[0]:i),a!=null&&a.length>0&&Array.isArray(a[0])?l=a.map((c,h)=>new bn(u,c,this,nt(t),e,this.name,h)):l=new bn(u,a,this,nt(t),e,this.name),this.addInboundNode(t,l,null,null,i,a,e),this._refCount++,this.activityRegularizer!=null)throw new Z("Layer invocation in the presence of activity regularizer(s) is not supported yet.");return l}})}warnOnIncompatibleInputShape(t){if(this.batchInputShape!=null)if(t.length!==this.batchInputShape.length)console.warn(`The rank of the input tensor provided (shape: ${JSON.stringify(t)}) does not match that of the batchInputShape (${JSON.stringify(this.batchInputShape)}) of the layer ${this.name}`);else{let e=!1;this.batchInputShape.forEach((s,r)=>{s!=null&&t[r]!=null&&t[r]!==s&&(e=!0)}),e&&console.warn(`The shape of the input tensor (${JSON.stringify(t)}) does not match the expectation of layer ${this.name}: ${JSON.stringify(this.batchInputShape)}`)}}get outputShape(){if(this.inboundNodes==null||this.inboundNodes.length===0)throw new je(`The layer ${this.name} has never been called and thus has no defined output shape.`);const t=[];for(const e of this.inboundNodes){const s=JSON.stringify(e.outputShapes);t.indexOf(s)===-1&&t.push(s)}if(t.length===1){const e=this.inboundNodes[0].outputShapes;return Array.isArray(e)&&Array.isArray(e[0])&&e.length===1?e[0]:e}else throw new je(`The layer ${this.name} has multiple inbound nodes with different output shapes. Hence the notion of "output shape" is ill-defined for the layer.`)}countParams(){if(!this.built)throw new He(`You tried to call countParams() on ${this.name}, but the layer is not built yet. Build it first by calling build(batchInputShape).`);return Ir(this.weights)}build(t){this.built=!0}getWeights(t=!1){return qu(t?this.trainableWeights:this.weights)}setWeights(t){k(()=>{const e=this.weights;if(e.length!==t.length)throw new I(`You called setWeights(weights) on layer "${this.name}" with a weight list of length ${t.length}, but the layer was expecting ${e.length} weights. Provided weights: ${t}...`);if(e.length===0)return;const s=[],r=qu(e);for(let o=0;o<r.length;++o){const i=r[o],a=e[o],l=t[o];if(!Zt(i.shape,l.shape))throw new I(`Layer weight shape ${i.shape} not compatible with provided weight shape ${l.shape}`);s.push([a,l])}ju(s)})}addWeight(t,e,s,r,o,i,a,l){if(this._addedWeightNames.indexOf(t)!==-1)throw new I(`Duplicate weight name ${t} for layer ${this.name}`);this._addedWeightNames.push(t),s==null&&(s="float32"),this.fastWeightInitDuringBuild&&(r=l!=null?l():$s("zeros"));const u=r.apply(e,s),c=new aw(u,s,t,i,a);return u.dispose(),o!=null&&this.addLoss(()=>o.apply(c.read())),i==null&&(i=!0),i?this._trainableWeights.push(c):this._nonTrainableWeights.push(c),c}setFastWeightInitDuringBuild(t){this.fastWeightInitDuringBuild=t}addLoss(t){t==null||Array.isArray(t)&&t.length===0||(t=nt(t),this._losses!==void 0&&this._losses!==null&&this.losses.push(...t))}computeOutputShape(t){return t}computeMask(t,e){if(!this.supportsMasking){if(e!=null)if(Array.isArray(e))e.forEach(s=>{if(s!=null)throw new TypeError(`Layer ${this.name} does not support masking, but was passed an inputMask.`)});else throw new TypeError(`Layer ${this.name} does not support masking, but was passed an inputMask.`);return null}return e}setMaskMetadata(t,e,s){if(!this.supportsMasking)return;const r=this.computeMask(t,s),o=nt(e),i=nt(r);if(o.length!==i.length)throw new Error(`${this.name} outputs ${o.length} tensors but ${o.length} masks for those tensors`);for(let a=0;a<o.length;a++)o[a].kerasMask=i[a]}addInboundNode(t,e,s,r,o,i,a=null){const l=nt(t);e=nt(e),s=nt(s),r=nt(r),o=vr(o),i=vr(i);const u=[],c=[],h=[];for(const f of l)u.push(f.sourceLayer),c.push(f.nodeIndex),h.push(f.tensorIndex);new xi({outboundLayer:this,inboundLayers:u,nodeIndices:c,tensorIndices:h,inputTensors:l,outputTensors:e,inputMasks:s,outputMasks:r,inputShapes:o,outputShapes:i},a);for(let f=0;f<e.length;f++)e[f].sourceLayer=this,e[f].nodeIndex=this.inboundNodes.length-1,e[f].tensorIndex=f}getConfig(){const t={name:this.name,trainable:this.trainable};return this.batchInputShape!=null&&(t.batchInputShape=this.batchInputShape),this.dtype!=null&&(t.dtype=this.dtype),t}disposeWeights(){return this.weights.forEach(t=>t.dispose()),this.weights.length}assertNotDisposed(){if(this._refCount===0)throw new Error(`Layer '${this.name}' is already disposed.`)}dispose(){if(!this.built)throw new Error(`Cannot dispose Layer ${this.name} because it has not been built yet.`);if(this._refCount===null)throw new Error(`Cannot dispose Layer ${this.name} because it has not been used yet.`);this.assertNotDisposed();let t=0;return--this._refCount===0&&(t=this.disposeWeights()),{refCountAfterDispose:this._refCount,numDisposedVariables:t}}}function hw(n){n=nt(n);const t=[];for(const e of n)t.push(e.shape);return Ut(t)}function fw(n){return"float32"}function dw(n){let t=!0;for(const e of nt(n))if(!(e instanceof bn)){t=!1;break}return t}function pw(n){let t=!0;for(const e of nt(n))if(e instanceof bn){t=!1;break}return t}/**
+ */class Ce{constructor(t){this.dtype=t.dtype,this.shape=t.shape,t.shape!=null?this.ndim=t.shape.length:this.ndim=t.ndim,this.maxNDim=t.maxNDim,this.minNDim=t.minNDim,this.axes=t.axes||{}}}class yn{constructor(t,e,s,r,o,i,a){this.dtype=t,this.shape=e,this.sourceLayer=s,this.inputs=r,this.callArgs=o,this.outputTensorIndex=a,this.id=Ru(),i!=null&&(this.originalName=ou(i),this.name=iu(this.originalName)),this.rank=e.length}}let uw=0;class xi{constructor(t,e){this.callArgs=e,this.id=uw++,this.outboundLayer=t.outboundLayer,this.inboundLayers=t.inboundLayers,this.nodeIndices=t.nodeIndices,this.tensorIndices=t.tensorIndices,this.inputTensors=t.inputTensors,this.outputTensors=t.outputTensors,this.inputMasks=t.inputMasks,this.outputMasks=t.outputMasks,this.inputShapes=t.inputShapes,this.outputShapes=t.outputShapes;for(const s of t.inboundLayers)s!=null&&s.outboundNodes.push(this);t.outboundLayer.inboundNodes.push(this)}getConfig(){const t=[];for(const e of this.inboundLayers)e!=null?t.push(e.name):t.push(null);return{outboundLayer:this.outboundLayer?this.outboundLayer.name:null,inboundLayers:t,nodeIndices:this.nodeIndices,tensorIndices:this.tensorIndices}}}let cw=0;class pe extends Gn{constructor(t={}){super(),this._callHook=null,this._addedWeightNames=[],this._stateful=!1,this.id=cw++,this.activityRegularizer=null,this.inputSpec=null,this.supportsMasking=!1,this._trainableWeights=[],this._nonTrainableWeights=[],this._losses=[],this._updates=[],this._built=!1,this.inboundNodes=[],this.outboundNodes=[];let e=t.name;if(!e){const s=this.getClassName();e=Le(s)+"_"+di(s)}if(this.name=e,this.trainable_=t.trainable==null?!0:t.trainable,t.inputShape!=null||t.batchInputShape!=null){let s;if(t.batchInputShape!=null)s=t.batchInputShape;else if(t.inputShape!=null){let o=null;t.batchSize!=null&&(o=t.batchSize),s=[o].concat(t.inputShape)}this.batchInputShape=s;let r=t.dtype;r==null&&(r=t.inputDType),r==null&&(r="float32"),this.dtype=r}t.weights!=null?this.initialWeights=t.weights:this.initialWeights=null,this._refCount=null,this.fastWeightInitDuringBuild=!1}static nodeKey(t,e){return t.name+"_ib-"+e.toString()}getNodeAtIndex(t,e){if(this.inboundNodes.length===0)throw new He(`The layer has never been called and thus has no defined ${e}.`);if(this.inboundNodes.length<=t)throw new I(`Asked to get ${e} at node ${t}, but the layer has only ${this.inboundNodes.length} inbound nodes.`);return this.inboundNodes[t]}getInputAt(t){return Ut(this.getNodeAtIndex(t,"input").inputTensors)}getOutputAt(t){return Ut(this.getNodeAtIndex(t,"output").outputTensors)}get input(){if(this.inboundNodes.length>1)throw new je(`Layer ${this.name} has multiple inbound nodes, hence the notion of "layer input" is ill-defined. Use \`getInputAt(nodeIndex)\` instead.`);if(this.inboundNodes.length===0)throw new je(`Layer ${this.name} is not connected, no input to return.`);return Ut(this.getNodeAtIndex(0,"input").inputTensors)}get output(){if(this.inboundNodes.length===0)throw new je(`Layer ${this.name} has no inbound nodes.`);if(this.inboundNodes.length>1)throw new je(`Layer ${this.name} has multiple inbound nodes, hence the notion of "layer output" is ill-defined. Use \`getOutputAt(nodeIndex)\` instead.`);return Ut(this.getNodeAtIndex(0,"output").outputTensors)}get losses(){return this._losses}calculateLosses(){return this.losses.map(t=>t())}get updates(){return this._updates}get built(){return this._built}set built(t){this._built=t}get trainable(){return this.trainable_}set trainable(t){this._trainableWeights.forEach(e=>e.trainable=t),this.trainable_=t}get trainableWeights(){return this.trainable_?this._trainableWeights.filter(t=>t.trainable):[]}set trainableWeights(t){this._trainableWeights=t}get nonTrainableWeights(){return this.trainable?this._trainableWeights.filter(t=>!t.trainable).concat(this._nonTrainableWeights):this._trainableWeights.concat(this._nonTrainableWeights)}set nonTrainableWeights(t){this._nonTrainableWeights=t}get weights(){return this.trainableWeights.concat(this.nonTrainableWeights)}get stateful(){return this._stateful}resetStates(){if(!this.stateful)throw new Error("Cannot call the resetStates() method of a non-stateful Layer object.")}assertInputCompatibility(t){const e=nt(t);if(this.inputSpec==null||this.inputSpec.length===0)return;const s=nt(this.inputSpec);if(e.length!==s.length)throw new I(`Layer ${this.name} expects ${s.length} inputs, but it received ${e.length} input tensors. Input received: ${t}`);for(let r=0;r<e.length;r++){const o=e[r],i=s[r];if(i==null)continue;const a=o.rank;if(i.ndim!=null&&a!==i.ndim)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected ndim=${i.ndim}, found ndim=${a}`);if(i.maxNDim!=null&&a>i.maxNDim)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected max_ndim=${i.maxNDim}, found ndim=${a}`);if(i.minNDim!=null&&a<i.minNDim)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected min_ndim=${i.minNDim}, found ndim=${a}.`);if(i.dtype!=null&&o.dtype!==i.dtype)throw new I(`Input ${r} is incompatible with layer ${this.name} : expected dtype=${i.dtype}, found dtype=${o.dtype}.`);if(i.axes){const l=o.shape;for(const u in i.axes){const c=Number(u),h=i.axes[u],f=c>=0?l[c]:l[l.length+c];if(h!=null&&[h,null].indexOf(f)===-1)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected axis ${c} of input shape to have value ${h} but got shape ${l}.`)}}if(i.shape!=null)for(let l=0;l<i.shape.length;++l){const u=i.shape[l],c=o.shape[l];if(u!=null&&c!=null&&u!==c)throw new I(`Input ${r} is incompatible with layer ${this.name}: expected shape=${i.shape}, found shape=${o.shape}.`)}}}call(t,e){return t}invokeCallHook(t,e){this._callHook!=null&&this._callHook(t,e)}setCallHook(t){this._callHook=t}clearCallHook(){this._callHook=null}apply(t,e){e=e||{},this.assertNotDisposed();const s=nt(t),r=dw(t),o=pw(t);if(r===o)throw new I("Arguments to apply() must be all SymbolicTensors or all Tensors");return br(this.name,()=>{if(!this.built){this.assertInputCompatibility(t);const i=[];for(const a of nt(t))i.push(a.shape);this.build(Ut(i)),this.built=!0,this.initialWeights&&this.setWeights(this.initialWeights),this._refCount===null&&o&&(this._refCount=1)}if(this.assertInputCompatibility(t),o){let i=this.call(t,e);this.supportsMasking&&this.setMaskMetadata(t,i);const a=nt(i),l=[];for(let u of a)s.indexOf(u)!==-1&&(u=u.clone()),l.push(u);if(i=Ut(l),this.activityRegularizer!=null)throw new Z("Layer invocation in the presence of activity regularizer(s) is not supported yet.");return i}else{const i=hw(t),a=this.computeOutputShape(i);let l;const u=fw(t);if(this.warnOnIncompatibleInputShape(Array.isArray(t)?i[0]:i),a!=null&&a.length>0&&Array.isArray(a[0])?l=a.map((c,h)=>new yn(u,c,this,nt(t),e,this.name,h)):l=new yn(u,a,this,nt(t),e,this.name),this.addInboundNode(t,l,null,null,i,a,e),this._refCount++,this.activityRegularizer!=null)throw new Z("Layer invocation in the presence of activity regularizer(s) is not supported yet.");return l}})}warnOnIncompatibleInputShape(t){if(this.batchInputShape!=null)if(t.length!==this.batchInputShape.length)console.warn(`The rank of the input tensor provided (shape: ${JSON.stringify(t)}) does not match that of the batchInputShape (${JSON.stringify(this.batchInputShape)}) of the layer ${this.name}`);else{let e=!1;this.batchInputShape.forEach((s,r)=>{s!=null&&t[r]!=null&&t[r]!==s&&(e=!0)}),e&&console.warn(`The shape of the input tensor (${JSON.stringify(t)}) does not match the expectation of layer ${this.name}: ${JSON.stringify(this.batchInputShape)}`)}}get outputShape(){if(this.inboundNodes==null||this.inboundNodes.length===0)throw new je(`The layer ${this.name} has never been called and thus has no defined output shape.`);const t=[];for(const e of this.inboundNodes){const s=JSON.stringify(e.outputShapes);t.indexOf(s)===-1&&t.push(s)}if(t.length===1){const e=this.inboundNodes[0].outputShapes;return Array.isArray(e)&&Array.isArray(e[0])&&e.length===1?e[0]:e}else throw new je(`The layer ${this.name} has multiple inbound nodes with different output shapes. Hence the notion of "output shape" is ill-defined for the layer.`)}countParams(){if(!this.built)throw new He(`You tried to call countParams() on ${this.name}, but the layer is not built yet. Build it first by calling build(batchInputShape).`);return Ar(this.weights)}build(t){this.built=!0}getWeights(t=!1){return qu(t?this.trainableWeights:this.weights)}setWeights(t){k(()=>{const e=this.weights;if(e.length!==t.length)throw new I(`You called setWeights(weights) on layer "${this.name}" with a weight list of length ${t.length}, but the layer was expecting ${e.length} weights. Provided weights: ${t}...`);if(e.length===0)return;const s=[],r=qu(e);for(let o=0;o<r.length;++o){const i=r[o],a=e[o],l=t[o];if(!Zt(i.shape,l.shape))throw new I(`Layer weight shape ${i.shape} not compatible with provided weight shape ${l.shape}`);s.push([a,l])}ju(s)})}addWeight(t,e,s,r,o,i,a,l){if(this._addedWeightNames.indexOf(t)!==-1)throw new I(`Duplicate weight name ${t} for layer ${this.name}`);this._addedWeightNames.push(t),s==null&&(s="float32"),this.fastWeightInitDuringBuild&&(r=l!=null?l():vs("zeros"));const u=r.apply(e,s),c=new aw(u,s,t,i,a);return u.dispose(),o!=null&&this.addLoss(()=>o.apply(c.read())),i==null&&(i=!0),i?this._trainableWeights.push(c):this._nonTrainableWeights.push(c),c}setFastWeightInitDuringBuild(t){this.fastWeightInitDuringBuild=t}addLoss(t){t==null||Array.isArray(t)&&t.length===0||(t=nt(t),this._losses!==void 0&&this._losses!==null&&this.losses.push(...t))}computeOutputShape(t){return t}computeMask(t,e){if(!this.supportsMasking){if(e!=null)if(Array.isArray(e))e.forEach(s=>{if(s!=null)throw new TypeError(`Layer ${this.name} does not support masking, but was passed an inputMask.`)});else throw new TypeError(`Layer ${this.name} does not support masking, but was passed an inputMask.`);return null}return e}setMaskMetadata(t,e,s){if(!this.supportsMasking)return;const r=this.computeMask(t,s),o=nt(e),i=nt(r);if(o.length!==i.length)throw new Error(`${this.name} outputs ${o.length} tensors but ${o.length} masks for those tensors`);for(let a=0;a<o.length;a++)o[a].kerasMask=i[a]}addInboundNode(t,e,s,r,o,i,a=null){const l=nt(t);e=nt(e),s=nt(s),r=nt(r),o=Ir(o),i=Ir(i);const u=[],c=[],h=[];for(const f of l)u.push(f.sourceLayer),c.push(f.nodeIndex),h.push(f.tensorIndex);new xi({outboundLayer:this,inboundLayers:u,nodeIndices:c,tensorIndices:h,inputTensors:l,outputTensors:e,inputMasks:s,outputMasks:r,inputShapes:o,outputShapes:i},a);for(let f=0;f<e.length;f++)e[f].sourceLayer=this,e[f].nodeIndex=this.inboundNodes.length-1,e[f].tensorIndex=f}getConfig(){const t={name:this.name,trainable:this.trainable};return this.batchInputShape!=null&&(t.batchInputShape=this.batchInputShape),this.dtype!=null&&(t.dtype=this.dtype),t}disposeWeights(){return this.weights.forEach(t=>t.dispose()),this.weights.length}assertNotDisposed(){if(this._refCount===0)throw new Error(`Layer '${this.name}' is already disposed.`)}dispose(){if(!this.built)throw new Error(`Cannot dispose Layer ${this.name} because it has not been built yet.`);if(this._refCount===null)throw new Error(`Cannot dispose Layer ${this.name} because it has not been used yet.`);this.assertNotDisposed();let t=0;return--this._refCount===0&&(t=this.disposeWeights()),{refCountAfterDispose:this._refCount,numDisposedVariables:t}}}function hw(n){n=nt(n);const t=[];for(const e of n)t.push(e.shape);return Ut(t)}function fw(n){return"float32"}function dw(n){let t=!0;for(const e of nt(n))if(!(e instanceof yn)){t=!1;break}return t}function pw(n){let t=!0;for(const e of nt(n))if(e instanceof yn){t=!1;break}return t}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2876,7 +2876,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function mw(n){if(n!=null&&typeof n!="object")throw new Error(`Argument to L1L2 regularizer's constructor is expected to be an object, but received: ${n}`)}class Hu extends Gn{}class Ku extends Hu{constructor(t){super(),mw(t),this.l1=t==null||t.l1==null?.01:t.l1,this.l2=t==null||t.l2==null?.01:t.l2,this.hasL1=this.l1!==0,this.hasL2=this.l2!==0}apply(t){return k(()=>{let e=Un([1]);return this.hasL1&&(e=O(e,et(N(this.l1,Lt(t))))),this.hasL2&&(e=O(e,et(N(this.l2,ws(t))))),L(e,[])})}getConfig(){return{l1:this.l1,l2:this.l2}}static fromConfig(t,e){return new t({l1:e.l1,l2:e.l2})}}Ku.className="L1L2",M(Ku);const Yu={l1l2:"L1L2"};function vs(n){return oi(n)}function Xu(n,t={}){return gs(n,se.getMap().classNameMap,t,"regularizer")}function Is(n){if(n==null)return null;if(typeof n=="string"){const e={className:n in Yu?Yu[n]:n,config:{}};return Xu(e)}else return n instanceof Hu?n:Xu(n)}/**
+ */function mw(n){if(n!=null&&typeof n!="object")throw new Error(`Argument to L1L2 regularizer's constructor is expected to be an object, but received: ${n}`)}class Hu extends Gn{}class Ku extends Hu{constructor(t){super(),mw(t),this.l1=t==null||t.l1==null?.01:t.l1,this.l2=t==null||t.l2==null?.01:t.l2,this.hasL1=this.l1!==0,this.hasL2=this.l2!==0}apply(t){return k(()=>{let e=Un([1]);return this.hasL1&&(e=O(e,et(N(this.l1,Lt(t))))),this.hasL2&&(e=O(e,et(N(this.l2,xs(t))))),L(e,[])})}getConfig(){return{l1:this.l1,l2:this.l2}}static fromConfig(t,e){return new t({l1:e.l1,l2:e.l2})}}Ku.className="L1L2",M(Ku);const Yu={l1l2:"L1L2"};function Is(n){return oi(n)}function Xu(n,t={}){return bs(n,se.getMap().classNameMap,t,"regularizer")}function As(n){if(n==null)return null;if(typeof n=="string"){const e={className:n in Yu?Yu[n]:n,config:{}};return Xu(e)}else return n instanceof Hu?n:Xu(n)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2884,7 +2884,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Si(n,t,e){if(typeof n=="number")return pr(n,t);if(n.length!==t)throw new I(`The ${e} argument must be an integer or tuple of ${t} integers. Received: ${n.length} elements.`);for(let s=0;s<t;++s){const r=n[s];if(!qy(r))throw new I(`The ${e} argument must be an integer or tuple of ${t} integers. Received: ${JSON.stringify(n)} including a non-integer number ${r}`)}return n}function yn(n,t,e,s,r=1){if(n==null)return n;const o=t+(t-1)*(r-1);let i;return e==="same"?i=n:i=n-o+1,Math.floor((i+s-1)/s)}function ke(n,t,e,s){if(n==null)return null;if(s==="valid")n=n*t+lu([e-t,0]);else if(s==="same")n=n*t;else throw new I(`Unsupport padding mode: ${s}.`);return n}/**
+ */function Si(n,t,e){if(typeof n=="number")return mr(n,t);if(n.length!==t)throw new I(`The ${e} argument must be an integer or tuple of ${t} integers. Received: ${n.length} elements.`);for(let s=0;s<t;++s){const r=n[s];if(!qy(r))throw new I(`The ${e} argument must be an integer or tuple of ${t} integers. Received: ${JSON.stringify(n)} including a non-integer number ${r}`)}return n}function wn(n,t,e,s,r=1){if(n==null)return n;const o=t+(t-1)*(r-1);let i;return e==="same"?i=n:i=n-o+1,Math.floor((i+s-1)/s)}function ke(n,t,e,s){if(n==null)return null;if(s==="valid")n=n*t+lu([e-t,0]);else if(s==="same")n=n*t;else throw new I(`Unsupport padding mode: ${s}.`);return n}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2892,7 +2892,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Ju(n,t){return k(()=>(mt(t),t==="channelsFirst"?pt(n,[0,2,3,1]):n))}function Zu(n,t){return k(()=>(mt(t),t==="channelsFirst"?pt(n,[0,2,3,4,1]):n))}function gw(n,t,e,s=1,r="valid",o,i=1){return k(()=>{if(o==null&&(o=Hn()),mt(o),n.shape.length!==3)throw new I(`The input of a conv1dWithBias operation should be 3, but is ${n.shape.length} instead.`);if(t.shape.length!==3)throw new I(`The kernel for a conv1dWithBias operation should be 3, but is ${t.shape.length} instead`);if(e!=null&&e.shape.length!==1)throw new I(`The bias for a conv1dWithBias operation should be 1, but is ${e.shape.length} instead`);if(o==="channelsFirst"&&(n=pt(n,[0,2,1])),r==="causal")throw new Z("The support for CAUSAL padding mode in conv1dWithBias is not implemented yet.");let a=Hm(n,t,s,r==="same"?"same":"valid","NWC",i);return e!=null&&(a=xs(a,e)),a})}function Qu(n,t,e,s=[1,1],r="valid",o,i,a=null){return k(()=>{if(o==null&&(o=Hn()),mt(o),n.rank!==3&&n.rank!==4)throw new I(`conv2dWithBiasActivation expects input to be of rank 3 or 4, but received ${n.rank}.`);if(t.rank!==3&&t.rank!==4)throw new I(`conv2dWithBiasActivation expects kernel to be of rank 3 or 4, but received ${n.rank}.`);let l=Ju(n,o);if(r==="causal")throw new Z("The support for CAUSAL padding mode in conv1dWithBias is not implemented yet.");return l=rb({x:l,filter:t,strides:s,pad:r==="same"?"same":"valid",dilations:i,dataFormat:"NHWC",bias:e,activation:a}),o==="channelsFirst"&&(l=pt(l,[0,3,1,2])),l})}function bw(n,t,e,s=[1,1,1],r="valid",o,i){return k(()=>{if(o==null&&(o=Hn()),mt(o),n.rank!==4&&n.rank!==5)throw new I(`conv3dWithBias expects input to be of rank 4 or 5, but received ${n.rank}.`);if(t.rank!==4&&t.rank!==5)throw new I(`conv3dWithBias expects kernel to be of rank 4 or 5, but received ${n.rank}.`);let a=Zu(n,o);if(r==="causal")throw new Z("The support for CAUSAL padding mode in conv3dWithBias is not implemented yet.");return a=Zm(a,t,s,r==="same"?"same":"valid","NDHWC",i),e!=null&&(a=xs(a,e)),o==="channelsFirst"&&(a=pt(a,[0,4,1,2,3])),a})}class $i extends pe{constructor(t,e){if(super(e),this.bias=null,this.DEFAULT_KERNEL_INITIALIZER="glorotNormal",this.DEFAULT_BIAS_INITIALIZER="zeros",$i.verifyArgs(e),this.rank=t,Me(this.rank,"rank"),this.rank!==1&&this.rank!==2&&this.rank!==3)throw new Z(`Convolution layer for rank other than 1, 2, or 3 (${this.rank}) is not implemented yet.`);if(this.kernelSize=Si(e.kernelSize,t,"kernelSize"),this.strides=Si(e.strides==null?1:e.strides,t,"strides"),this.padding=e.padding==null?"valid":e.padding,oe(this.padding),this.dataFormat=e.dataFormat==null?"channelsLast":e.dataFormat,mt(this.dataFormat),this.activation=tw(e.activation),this.useBias=e.useBias==null?!0:e.useBias,this.biasInitializer=$s(e.biasInitializer||this.DEFAULT_BIAS_INITIALIZER),this.biasConstraint=xr(e.biasConstraint),this.biasRegularizer=Is(e.biasRegularizer),this.activityRegularizer=Is(e.activityRegularizer),this.dilationRate=Si(e.dilationRate==null?1:e.dilationRate,t,"dilationRate"),this.rank===1&&Array.isArray(this.dilationRate)&&this.dilationRate.length!==1)throw new I(`dilationRate must be a number or an array of a single number for 1D convolution, but received ${JSON.stringify(this.dilationRate)}`);if(this.rank===2){if(typeof this.dilationRate=="number")this.dilationRate=[this.dilationRate,this.dilationRate];else if(this.dilationRate.length!==2)throw new I(`dilationRate must be a number or array of two numbers for 2D convolution, but received ${JSON.stringify(this.dilationRate)}`)}else if(this.rank===3){if(typeof this.dilationRate=="number")this.dilationRate=[this.dilationRate,this.dilationRate,this.dilationRate];else if(this.dilationRate.length!==3)throw new I(`dilationRate must be a number or array of three numbers for 3D convolution, but received ${JSON.stringify(this.dilationRate)}`)}}static verifyArgs(t){if(Ae("kernelSize"in t,"required key 'kernelSize' not in config"),typeof t.kernelSize!="number"&&!ai(t.kernelSize,"number",1,3))throw new I(`BaseConv expects config.kernelSize to be number or number[] with length 1, 2, or 3, but received ${JSON.stringify(t.kernelSize)}.`)}getConfig(){const t={kernelSize:this.kernelSize,strides:this.strides,padding:this.padding,dataFormat:this.dataFormat,dilationRate:this.dilationRate,activation:Qy(this.activation),useBias:this.useBias,biasInitializer:$r(this.biasInitializer),biasRegularizer:vs(this.biasRegularizer),activityRegularizer:vs(this.activityRegularizer),biasConstraint:wr(this.biasConstraint)},e=super.getConfig();return Object.assign(t,e),t}}class Kn extends $i{constructor(t,e){super(t,e),this.kernel=null,Kn.verifyArgs(e),this.filters=e.filters,Me(this.filters,"filters"),this.kernelInitializer=$s(e.kernelInitializer||this.DEFAULT_KERNEL_INITIALIZER),this.kernelConstraint=xr(e.kernelConstraint),this.kernelRegularizer=Is(e.kernelRegularizer)}build(t){t=de(t);const e=this.dataFormat==="channelsFirst"?1:t.length-1;if(t[e]==null)throw new I(`The channel dimension of the input should be defined. Found ${t[e]}`);const s=t[e],r=this.kernelSize.concat([s,this.filters]);this.kernel=this.addWeight("kernel",r,null,this.kernelInitializer,this.kernelRegularizer,!0,this.kernelConstraint),this.useBias&&(this.bias=this.addWeight("bias",[this.filters],null,this.biasInitializer,this.biasRegularizer,!0,this.biasConstraint)),this.inputSpec=[{ndim:this.rank+2,axes:{[e]:s}}],this.built=!0}call(t,e){return k(()=>{t=Gt(t);let s;const r=this.bias==null?null:this.bias.read(),o=Uy(this.activation.getClassName());if(o!=null&&this.rank===2)s=Qu(t,this.kernel.read(),r,this.strides,this.padding,this.dataFormat,this.dilationRate,o);else{if(this.rank===1)s=gw(t,this.kernel.read(),r,this.strides[0],this.padding,this.dataFormat,this.dilationRate[0]);else if(this.rank===2)s=Qu(t,this.kernel.read(),r,this.strides,this.padding,this.dataFormat,this.dilationRate);else if(this.rank===3)s=bw(t,this.kernel.read(),r,this.strides,this.padding,this.dataFormat,this.dilationRate);else throw new Z("convolutions greater than 3D are not implemented yet.");this.activation!=null&&(s=this.activation.apply(s))}return s})}computeOutputShape(t){t=de(t);const e=[],s=this.dataFormat==="channelsLast"?t.slice(1,t.length-1):t.slice(2);for(let o=0;o<s.length;++o){const i=yn(s[o],this.kernelSize[o],this.padding,this.strides[o],typeof this.dilationRate=="number"?this.dilationRate:this.dilationRate[o]);e.push(i)}let r=[t[0]];return this.dataFormat==="channelsLast"?(r=r.concat(e),r.push(this.filters)):(r.push(this.filters),r=r.concat(e)),r}getConfig(){const t={filters:this.filters,kernelInitializer:$r(this.kernelInitializer),kernelRegularizer:vs(this.kernelRegularizer),kernelConstraint:wr(this.kernelConstraint)},e=super.getConfig();return Object.assign(t,e),t}static verifyArgs(t){if(!("filters"in t)||typeof t.filters!="number"||t.filters<1)throw new I(`Convolution layer expected config.filters to be a 'number' > 0 but got ${JSON.stringify(t.filters)}`)}}class Yn extends Kn{constructor(t){super(2,t),Yn.verifyArgs(t)}getConfig(){const t=super.getConfig();return delete t.rank,t}static verifyArgs(t){if(typeof t.kernelSize!="number"&&!ai(t.kernelSize,"number",1,2))throw new I(`Conv2D expects config.kernelSize to be number or number[] with length 1 or 2, but received ${JSON.stringify(t.kernelSize)}.`)}}Yn.className="Conv2D",M(Yn);class As extends Kn{constructor(t){super(3,t),As.verifyArgs(t)}getConfig(){const t=super.getConfig();return delete t.rank,t}static verifyArgs(t){if(typeof t.kernelSize!="number"&&!(Array.isArray(t.kernelSize)&&(t.kernelSize.length===1||t.kernelSize.length===3)))throw new I(`Conv3D expects config.kernelSize to be number or [number, number, number], but received ${JSON.stringify(t.kernelSize)}.`)}}As.className="Conv3D",M(As);class tc extends Yn{constructor(t){if(super(t),this.inputSpec=[new Ce({ndim:4})],this.padding!=="same"&&this.padding!=="valid")throw new I(`Conv2DTranspose currently supports only padding modes 'same' and 'valid', but received padding mode ${this.padding}`)}build(t){if(t=de(t),t.length!==4)throw new I("Input should have rank 4; Received input shape: "+JSON.stringify(t));const e=this.dataFormat==="channelsFirst"?1:t.length-1;if(t[e]==null)throw new I("The channel dimension of the inputs should be defined. Found `None`.");const s=t[e],r=this.kernelSize.concat([this.filters,s]);this.kernel=this.addWeight("kernel",r,"float32",this.kernelInitializer,this.kernelRegularizer,!0,this.kernelConstraint),this.useBias&&(this.bias=this.addWeight("bias",[this.filters],"float32",this.biasInitializer,this.biasRegularizer,!0,this.biasConstraint)),this.inputSpec=[new Ce({ndim:4,axes:{[e]:s}})],this.built=!0}call(t,e){return k(()=>{let s=Gt(t);if(s.shape.length!==4)throw new I(`Conv2DTranspose.call() expects input tensor to be rank-4, but received a tensor of rank-${s.shape.length}`);const r=s.shape,o=r[0];let i,a;this.dataFormat==="channelsFirst"?(i=2,a=3):(i=1,a=2);const l=r[i],u=r[a],c=this.kernelSize[0],h=this.kernelSize[1],f=this.strides[0],d=this.strides[1],p=ke(l,f,c,this.padding),g=ke(u,d,h,this.padding),m=[o,p,g,this.filters];this.dataFormat!=="channelsLast"&&(s=pt(s,[0,2,3,1]));let b=Xm(s,this.kernel.read(),m,this.strides,this.padding);return this.dataFormat!=="channelsLast"&&(b=pt(b,[0,3,1,2])),this.bias!=null&&(b=xs(b,this.bias.read(),this.dataFormat)),this.activation!=null&&(b=this.activation.apply(b)),b})}computeOutputShape(t){t=de(t);const e=t.slice();let s,r,o;this.dataFormat==="channelsFirst"?(s=1,r=2,o=3):(s=3,r=1,o=2);const i=this.kernelSize[0],a=this.kernelSize[1],l=this.strides[0],u=this.strides[1];return e[s]=this.filters,e[r]=ke(e[r],l,i,this.padding),e[o]=ke(e[o],u,a,this.padding),e}getConfig(){const t=super.getConfig();return delete t.dilationRate,t}}tc.className="Conv2DTranspose",M(tc);class ec extends As{constructor(t){if(super(t),this.inputSpec=[new Ce({ndim:5})],this.padding!=="same"&&this.padding!=="valid")throw new I(`Conv3DTranspose currently supports only padding modes 'same' and 'valid', but received padding mode ${this.padding}`)}build(t){if(t=de(t),t.length!==5)throw new I("Input should have rank 5; Received input shape: "+JSON.stringify(t));const e=this.dataFormat==="channelsFirst"?1:t.length-1;if(t[e]==null)throw new I("The channel dimension of the inputs should be defined. Found `None`.");const s=t[e],r=this.kernelSize.concat([this.filters,s]);this.kernel=this.addWeight("kernel",r,"float32",this.kernelInitializer,this.kernelRegularizer,!0,this.kernelConstraint),this.useBias&&(this.bias=this.addWeight("bias",[this.filters],"float32",this.biasInitializer,this.biasRegularizer,!0,this.biasConstraint)),this.inputSpec=[new Ce({ndim:5,axes:{[e]:s}})],this.built=!0}call(t,e){return k(()=>{let s=Gt(t);if(s.shape.length!==5)throw new I(`Conv3DTranspose.call() expects input tensor to be rank-4, but received a tensor of rank-${s.shape.length}`);const r=s.shape,o=r[0];let i,a,l;this.dataFormat==="channelsFirst"?(l=2,i=3,a=4):(l=1,i=2,a=3);const u=r[l],c=r[i],h=r[a],f=this.kernelSize[0],d=this.kernelSize[1],p=this.kernelSize[2],g=this.strides[0],m=this.strides[1],b=this.strides[2],y=ke(u,g,f,this.padding),S=ke(c,m,d,this.padding),x=ke(h,b,p,this.padding),$=[o,y,S,x,this.filters];this.dataFormat!=="channelsLast"&&(s=pt(s,[0,2,3,4,1]));let E=ng(s,this.kernel.read(),$,this.strides,this.padding);return this.dataFormat!=="channelsLast"&&(E=pt(E,[0,4,1,2,3])),this.bias!==null&&(E=xs(E,this.bias.read(),this.dataFormat)),this.activation!==null&&(E=this.activation.apply(E)),E})}computeOutputShape(t){t=de(t);const e=t.slice();let s,r,o,i;this.dataFormat==="channelsFirst"?(s=1,r=2,o=3,i=4):(s=4,r=1,o=2,i=3);const a=this.kernelSize[0],l=this.kernelSize[1],u=this.kernelSize[2],c=this.strides[0],h=this.strides[1],f=this.strides[2];return e[s]=this.filters,e[r]=ke(e[r],c,a,this.padding),e[o]=ke(e[o],h,l,this.padding),e[i]=ke(e[i],f,u,this.padding),e}getConfig(){const t=super.getConfig();return delete t.dilationRate,t}}ec.className="Conv3DTranspose",M(ec);class nc extends Kn{constructor(t,e){if(super(t,e),this.DEFAULT_DEPTHWISE_INITIALIZER="glorotUniform",this.DEFAULT_POINTWISE_INITIALIZER="glorotUniform",this.depthwiseKernel=null,this.pointwiseKernel=null,e.filters==null)throw new I("The `filters` configuration field is required by SeparableConv, but is unspecified.");if(e.kernelInitializer!=null||e.kernelRegularizer!=null||e.kernelConstraint!=null)throw new I("Fields kernelInitializer, kernelRegularizer and kernelConstraint are invalid for SeparableConv2D. Use depthwiseInitializer, depthwiseRegularizer, depthwiseConstraint, pointwiseInitializer, pointwiseRegularizer and pointwiseConstraint instead.");if(e.padding!=null&&e.padding!=="same"&&e.padding!=="valid")throw new I(`SeparableConv${this.rank}D supports only padding modes: 'same' and 'valid', but received ${JSON.stringify(e.padding)}`);this.depthMultiplier=e.depthMultiplier==null?1:e.depthMultiplier,this.depthwiseInitializer=$s(e.depthwiseInitializer||this.DEFAULT_DEPTHWISE_INITIALIZER),this.depthwiseRegularizer=Is(e.depthwiseRegularizer),this.depthwiseConstraint=xr(e.depthwiseConstraint),this.pointwiseInitializer=$s(e.depthwiseInitializer||this.DEFAULT_POINTWISE_INITIALIZER),this.pointwiseRegularizer=Is(e.pointwiseRegularizer),this.pointwiseConstraint=xr(e.pointwiseConstraint)}build(t){if(t=de(t),t.length<this.rank+2)throw new I(`Inputs to SeparableConv${this.rank}D should have rank ${this.rank+2}, but received input shape: ${JSON.stringify(t)}`);const e=this.dataFormat==="channelsFirst"?1:t.length-1;if(t[e]==null||t[e]<0)throw new I(`The channel dimension of the inputs should be defined, but found ${JSON.stringify(t[e])}`);const s=t[e],r=this.kernelSize.concat([s,this.depthMultiplier]),o=[];for(let a=0;a<this.rank;++a)o.push(1);o.push(s*this.depthMultiplier,this.filters);const i=!0;this.depthwiseKernel=this.addWeight("depthwise_kernel",r,"float32",this.depthwiseInitializer,this.depthwiseRegularizer,i,this.depthwiseConstraint),this.pointwiseKernel=this.addWeight("pointwise_kernel",o,"float32",this.pointwiseInitializer,this.pointwiseRegularizer,i,this.pointwiseConstraint),this.useBias?this.bias=this.addWeight("bias",[this.filters],"float32",this.biasInitializer,this.biasRegularizer,i,this.biasConstraint):this.bias=null,this.inputSpec=[new Ce({ndim:this.rank+2,axes:{[e]:s}})],this.built=!0}call(t,e){return k(()=>{t=Gt(t);let s;if(this.rank===1)throw new Z("1D separable convolution is not implemented yet.");return this.rank===2&&(this.dataFormat==="channelsFirst"&&(t=pt(t,[0,2,3,1])),s=O0(t,this.depthwiseKernel.read(),this.pointwiseKernel.read(),this.strides,this.padding,this.dilationRate,"NHWC")),this.useBias&&(s=xs(s,this.bias.read(),this.dataFormat)),this.activation!=null&&(s=this.activation.apply(s)),this.dataFormat==="channelsFirst"&&(s=pt(s,[0,3,1,2])),s})}getConfig(){const t=super.getConfig();return delete t.rank,delete t.kernelInitializer,delete t.kernelRegularizer,delete t.kernelConstraint,t.depthwiseInitializer=$r(this.depthwiseInitializer),t.pointwiseInitializer=$r(this.pointwiseInitializer),t.depthwiseRegularizer=vs(this.depthwiseRegularizer),t.pointwiseRegularizer=vs(this.pointwiseRegularizer),t.depthwiseConstraint=wr(this.depthwiseConstraint),t.pointwiseConstraint=wr(this.pointwiseConstraint),t}}nc.className="SeparableConv";class sc extends nc{constructor(t){super(2,t)}}sc.className="SeparableConv2D",M(sc);class Ar extends Kn{constructor(t){super(1,t),Ar.verifyArgs(t),this.inputSpec=[{ndim:3}]}getConfig(){const t=super.getConfig();return delete t.rank,delete t.dataFormat,t}static verifyArgs(t){if(typeof t.kernelSize!="number"&&!ai(t.kernelSize,"number",1,1))throw new I(`Conv1D expects config.kernelSize to be number or number[] with length 1, but received ${JSON.stringify(t.kernelSize)}.`)}}Ar.className="Conv1D",M(Ar);class rc extends pe{constructor(t){super(t),typeof t.cropping=="number"?this.cropping=[[t.cropping,t.cropping],[t.cropping,t.cropping]]:typeof t.cropping[0]=="number"?this.cropping=[[t.cropping[0],t.cropping[0]],[t.cropping[1],t.cropping[1]]]:this.cropping=t.cropping,this.dataFormat=t.dataFormat===void 0?"channelsLast":t.dataFormat,this.inputSpec=[{ndim:4}]}computeOutputShape(t){return this.dataFormat==="channelsFirst"?[t[0],t[1],t[2]-this.cropping[0][0]-this.cropping[0][1],t[3]-this.cropping[1][0]-this.cropping[1][1]]:[t[0],t[1]-this.cropping[0][0]-this.cropping[0][1],t[2]-this.cropping[1][0]-this.cropping[1][1],t[3]]}call(t,e){return k(()=>{if(t=Gt(t),this.dataFormat==="channelsLast"){const s=yr(t,this.cropping[0][0],t.shape[1]-this.cropping[0][0]-this.cropping[0][1],2);return yr(s,this.cropping[1][0],t.shape[2]-this.cropping[1][1]-this.cropping[1][0],3)}else{const s=yr(t,this.cropping[0][0],t.shape[2]-this.cropping[0][0]-this.cropping[0][1],3);return yr(s,this.cropping[1][0],t.shape[3]-this.cropping[1][1]-this.cropping[1][0],4)}})}getConfig(){const t={cropping:this.cropping,dataFormat:this.dataFormat},e=super.getConfig();return Object.assign(t,e),t}}rc.className="Cropping2D",M(rc);class vi extends pe{constructor(t){super(t),this.DEFAULT_SIZE=[2,2],this.inputSpec=[{ndim:4}],this.size=t.size==null?this.DEFAULT_SIZE:t.size,this.dataFormat=t.dataFormat==null?"channelsLast":t.dataFormat,mt(this.dataFormat),this.interpolation=t.interpolation==null?"nearest":t.interpolation,Wy(this.interpolation)}computeOutputShape(t){if(this.dataFormat==="channelsFirst"){const e=t[2]==null?null:this.size[0]*t[2],s=t[3]==null?null:this.size[1]*t[3];return[t[0],t[1],e,s]}else{const e=t[1]==null?null:this.size[0]*t[1],s=t[2]==null?null:this.size[1]*t[2];return[t[0],e,s,t[3]]}}call(t,e){return k(()=>{let s=Gt(t);const r=s.shape;if(this.dataFormat==="channelsFirst"){s=pt(s,[0,2,3,1]);const o=this.size[0]*r[2],i=this.size[1]*r[3],a=this.interpolation==="nearest"?dr.resizeNearestNeighbor(s,[o,i]):dr.resizeBilinear(s,[o,i]);return pt(a,[0,3,1,2])}else{const o=this.size[0]*r[1],i=this.size[1]*r[2];return this.interpolation==="nearest"?dr.resizeNearestNeighbor(s,[o,i]):dr.resizeBilinear(s,[o,i])}})}getConfig(){const t={size:this.size,dataFormat:this.dataFormat,interpolation:this.interpolation},e=super.getConfig();return Object.assign(t,e),t}}vi.className="UpSampling2D",M(vi);/**
+ */function Ju(n,t){return k(()=>(mt(t),t==="channelsFirst"?pt(n,[0,2,3,1]):n))}function Zu(n,t){return k(()=>(mt(t),t==="channelsFirst"?pt(n,[0,2,3,4,1]):n))}function gw(n,t,e,s=1,r="valid",o,i=1){return k(()=>{if(o==null&&(o=Hn()),mt(o),n.shape.length!==3)throw new I(`The input of a conv1dWithBias operation should be 3, but is ${n.shape.length} instead.`);if(t.shape.length!==3)throw new I(`The kernel for a conv1dWithBias operation should be 3, but is ${t.shape.length} instead`);if(e!=null&&e.shape.length!==1)throw new I(`The bias for a conv1dWithBias operation should be 1, but is ${e.shape.length} instead`);if(o==="channelsFirst"&&(n=pt(n,[0,2,1])),r==="causal")throw new Z("The support for CAUSAL padding mode in conv1dWithBias is not implemented yet.");let a=Hm(n,t,s,r==="same"?"same":"valid","NWC",i);return e!=null&&(a=Ss(a,e)),a})}function Qu(n,t,e,s=[1,1],r="valid",o,i,a=null){return k(()=>{if(o==null&&(o=Hn()),mt(o),n.rank!==3&&n.rank!==4)throw new I(`conv2dWithBiasActivation expects input to be of rank 3 or 4, but received ${n.rank}.`);if(t.rank!==3&&t.rank!==4)throw new I(`conv2dWithBiasActivation expects kernel to be of rank 3 or 4, but received ${n.rank}.`);let l=Ju(n,o);if(r==="causal")throw new Z("The support for CAUSAL padding mode in conv1dWithBias is not implemented yet.");return l=rb({x:l,filter:t,strides:s,pad:r==="same"?"same":"valid",dilations:i,dataFormat:"NHWC",bias:e,activation:a}),o==="channelsFirst"&&(l=pt(l,[0,3,1,2])),l})}function bw(n,t,e,s=[1,1,1],r="valid",o,i){return k(()=>{if(o==null&&(o=Hn()),mt(o),n.rank!==4&&n.rank!==5)throw new I(`conv3dWithBias expects input to be of rank 4 or 5, but received ${n.rank}.`);if(t.rank!==4&&t.rank!==5)throw new I(`conv3dWithBias expects kernel to be of rank 4 or 5, but received ${n.rank}.`);let a=Zu(n,o);if(r==="causal")throw new Z("The support for CAUSAL padding mode in conv3dWithBias is not implemented yet.");return a=Zm(a,t,s,r==="same"?"same":"valid","NDHWC",i),e!=null&&(a=Ss(a,e)),o==="channelsFirst"&&(a=pt(a,[0,4,1,2,3])),a})}class $i extends pe{constructor(t,e){if(super(e),this.bias=null,this.DEFAULT_KERNEL_INITIALIZER="glorotNormal",this.DEFAULT_BIAS_INITIALIZER="zeros",$i.verifyArgs(e),this.rank=t,Me(this.rank,"rank"),this.rank!==1&&this.rank!==2&&this.rank!==3)throw new Z(`Convolution layer for rank other than 1, 2, or 3 (${this.rank}) is not implemented yet.`);if(this.kernelSize=Si(e.kernelSize,t,"kernelSize"),this.strides=Si(e.strides==null?1:e.strides,t,"strides"),this.padding=e.padding==null?"valid":e.padding,oe(this.padding),this.dataFormat=e.dataFormat==null?"channelsLast":e.dataFormat,mt(this.dataFormat),this.activation=tw(e.activation),this.useBias=e.useBias==null?!0:e.useBias,this.biasInitializer=vs(e.biasInitializer||this.DEFAULT_BIAS_INITIALIZER),this.biasConstraint=Sr(e.biasConstraint),this.biasRegularizer=As(e.biasRegularizer),this.activityRegularizer=As(e.activityRegularizer),this.dilationRate=Si(e.dilationRate==null?1:e.dilationRate,t,"dilationRate"),this.rank===1&&Array.isArray(this.dilationRate)&&this.dilationRate.length!==1)throw new I(`dilationRate must be a number or an array of a single number for 1D convolution, but received ${JSON.stringify(this.dilationRate)}`);if(this.rank===2){if(typeof this.dilationRate=="number")this.dilationRate=[this.dilationRate,this.dilationRate];else if(this.dilationRate.length!==2)throw new I(`dilationRate must be a number or array of two numbers for 2D convolution, but received ${JSON.stringify(this.dilationRate)}`)}else if(this.rank===3){if(typeof this.dilationRate=="number")this.dilationRate=[this.dilationRate,this.dilationRate,this.dilationRate];else if(this.dilationRate.length!==3)throw new I(`dilationRate must be a number or array of three numbers for 3D convolution, but received ${JSON.stringify(this.dilationRate)}`)}}static verifyArgs(t){if(Ae("kernelSize"in t,"required key 'kernelSize' not in config"),typeof t.kernelSize!="number"&&!ai(t.kernelSize,"number",1,3))throw new I(`BaseConv expects config.kernelSize to be number or number[] with length 1, 2, or 3, but received ${JSON.stringify(t.kernelSize)}.`)}getConfig(){const t={kernelSize:this.kernelSize,strides:this.strides,padding:this.padding,dataFormat:this.dataFormat,dilationRate:this.dilationRate,activation:Qy(this.activation),useBias:this.useBias,biasInitializer:vr(this.biasInitializer),biasRegularizer:Is(this.biasRegularizer),activityRegularizer:Is(this.activityRegularizer),biasConstraint:xr(this.biasConstraint)},e=super.getConfig();return Object.assign(t,e),t}}class Kn extends $i{constructor(t,e){super(t,e),this.kernel=null,Kn.verifyArgs(e),this.filters=e.filters,Me(this.filters,"filters"),this.kernelInitializer=vs(e.kernelInitializer||this.DEFAULT_KERNEL_INITIALIZER),this.kernelConstraint=Sr(e.kernelConstraint),this.kernelRegularizer=As(e.kernelRegularizer)}build(t){t=de(t);const e=this.dataFormat==="channelsFirst"?1:t.length-1;if(t[e]==null)throw new I(`The channel dimension of the input should be defined. Found ${t[e]}`);const s=t[e],r=this.kernelSize.concat([s,this.filters]);this.kernel=this.addWeight("kernel",r,null,this.kernelInitializer,this.kernelRegularizer,!0,this.kernelConstraint),this.useBias&&(this.bias=this.addWeight("bias",[this.filters],null,this.biasInitializer,this.biasRegularizer,!0,this.biasConstraint)),this.inputSpec=[{ndim:this.rank+2,axes:{[e]:s}}],this.built=!0}call(t,e){return k(()=>{t=Gt(t);let s;const r=this.bias==null?null:this.bias.read(),o=Uy(this.activation.getClassName());if(o!=null&&this.rank===2)s=Qu(t,this.kernel.read(),r,this.strides,this.padding,this.dataFormat,this.dilationRate,o);else{if(this.rank===1)s=gw(t,this.kernel.read(),r,this.strides[0],this.padding,this.dataFormat,this.dilationRate[0]);else if(this.rank===2)s=Qu(t,this.kernel.read(),r,this.strides,this.padding,this.dataFormat,this.dilationRate);else if(this.rank===3)s=bw(t,this.kernel.read(),r,this.strides,this.padding,this.dataFormat,this.dilationRate);else throw new Z("convolutions greater than 3D are not implemented yet.");this.activation!=null&&(s=this.activation.apply(s))}return s})}computeOutputShape(t){t=de(t);const e=[],s=this.dataFormat==="channelsLast"?t.slice(1,t.length-1):t.slice(2);for(let o=0;o<s.length;++o){const i=wn(s[o],this.kernelSize[o],this.padding,this.strides[o],typeof this.dilationRate=="number"?this.dilationRate:this.dilationRate[o]);e.push(i)}let r=[t[0]];return this.dataFormat==="channelsLast"?(r=r.concat(e),r.push(this.filters)):(r.push(this.filters),r=r.concat(e)),r}getConfig(){const t={filters:this.filters,kernelInitializer:vr(this.kernelInitializer),kernelRegularizer:Is(this.kernelRegularizer),kernelConstraint:xr(this.kernelConstraint)},e=super.getConfig();return Object.assign(t,e),t}static verifyArgs(t){if(!("filters"in t)||typeof t.filters!="number"||t.filters<1)throw new I(`Convolution layer expected config.filters to be a 'number' > 0 but got ${JSON.stringify(t.filters)}`)}}class Yn extends Kn{constructor(t){super(2,t),Yn.verifyArgs(t)}getConfig(){const t=super.getConfig();return delete t.rank,t}static verifyArgs(t){if(typeof t.kernelSize!="number"&&!ai(t.kernelSize,"number",1,2))throw new I(`Conv2D expects config.kernelSize to be number or number[] with length 1 or 2, but received ${JSON.stringify(t.kernelSize)}.`)}}Yn.className="Conv2D",M(Yn);class Es extends Kn{constructor(t){super(3,t),Es.verifyArgs(t)}getConfig(){const t=super.getConfig();return delete t.rank,t}static verifyArgs(t){if(typeof t.kernelSize!="number"&&!(Array.isArray(t.kernelSize)&&(t.kernelSize.length===1||t.kernelSize.length===3)))throw new I(`Conv3D expects config.kernelSize to be number or [number, number, number], but received ${JSON.stringify(t.kernelSize)}.`)}}Es.className="Conv3D",M(Es);class tc extends Yn{constructor(t){if(super(t),this.inputSpec=[new Ce({ndim:4})],this.padding!=="same"&&this.padding!=="valid")throw new I(`Conv2DTranspose currently supports only padding modes 'same' and 'valid', but received padding mode ${this.padding}`)}build(t){if(t=de(t),t.length!==4)throw new I("Input should have rank 4; Received input shape: "+JSON.stringify(t));const e=this.dataFormat==="channelsFirst"?1:t.length-1;if(t[e]==null)throw new I("The channel dimension of the inputs should be defined. Found `None`.");const s=t[e],r=this.kernelSize.concat([this.filters,s]);this.kernel=this.addWeight("kernel",r,"float32",this.kernelInitializer,this.kernelRegularizer,!0,this.kernelConstraint),this.useBias&&(this.bias=this.addWeight("bias",[this.filters],"float32",this.biasInitializer,this.biasRegularizer,!0,this.biasConstraint)),this.inputSpec=[new Ce({ndim:4,axes:{[e]:s}})],this.built=!0}call(t,e){return k(()=>{let s=Gt(t);if(s.shape.length!==4)throw new I(`Conv2DTranspose.call() expects input tensor to be rank-4, but received a tensor of rank-${s.shape.length}`);const r=s.shape,o=r[0];let i,a;this.dataFormat==="channelsFirst"?(i=2,a=3):(i=1,a=2);const l=r[i],u=r[a],c=this.kernelSize[0],h=this.kernelSize[1],f=this.strides[0],d=this.strides[1],p=ke(l,f,c,this.padding),g=ke(u,d,h,this.padding),m=[o,p,g,this.filters];this.dataFormat!=="channelsLast"&&(s=pt(s,[0,2,3,1]));let b=Xm(s,this.kernel.read(),m,this.strides,this.padding);return this.dataFormat!=="channelsLast"&&(b=pt(b,[0,3,1,2])),this.bias!=null&&(b=Ss(b,this.bias.read(),this.dataFormat)),this.activation!=null&&(b=this.activation.apply(b)),b})}computeOutputShape(t){t=de(t);const e=t.slice();let s,r,o;this.dataFormat==="channelsFirst"?(s=1,r=2,o=3):(s=3,r=1,o=2);const i=this.kernelSize[0],a=this.kernelSize[1],l=this.strides[0],u=this.strides[1];return e[s]=this.filters,e[r]=ke(e[r],l,i,this.padding),e[o]=ke(e[o],u,a,this.padding),e}getConfig(){const t=super.getConfig();return delete t.dilationRate,t}}tc.className="Conv2DTranspose",M(tc);class ec extends Es{constructor(t){if(super(t),this.inputSpec=[new Ce({ndim:5})],this.padding!=="same"&&this.padding!=="valid")throw new I(`Conv3DTranspose currently supports only padding modes 'same' and 'valid', but received padding mode ${this.padding}`)}build(t){if(t=de(t),t.length!==5)throw new I("Input should have rank 5; Received input shape: "+JSON.stringify(t));const e=this.dataFormat==="channelsFirst"?1:t.length-1;if(t[e]==null)throw new I("The channel dimension of the inputs should be defined. Found `None`.");const s=t[e],r=this.kernelSize.concat([this.filters,s]);this.kernel=this.addWeight("kernel",r,"float32",this.kernelInitializer,this.kernelRegularizer,!0,this.kernelConstraint),this.useBias&&(this.bias=this.addWeight("bias",[this.filters],"float32",this.biasInitializer,this.biasRegularizer,!0,this.biasConstraint)),this.inputSpec=[new Ce({ndim:5,axes:{[e]:s}})],this.built=!0}call(t,e){return k(()=>{let s=Gt(t);if(s.shape.length!==5)throw new I(`Conv3DTranspose.call() expects input tensor to be rank-4, but received a tensor of rank-${s.shape.length}`);const r=s.shape,o=r[0];let i,a,l;this.dataFormat==="channelsFirst"?(l=2,i=3,a=4):(l=1,i=2,a=3);const u=r[l],c=r[i],h=r[a],f=this.kernelSize[0],d=this.kernelSize[1],p=this.kernelSize[2],g=this.strides[0],m=this.strides[1],b=this.strides[2],y=ke(u,g,f,this.padding),S=ke(c,m,d,this.padding),x=ke(h,b,p,this.padding),$=[o,y,S,x,this.filters];this.dataFormat!=="channelsLast"&&(s=pt(s,[0,2,3,4,1]));let E=ng(s,this.kernel.read(),$,this.strides,this.padding);return this.dataFormat!=="channelsLast"&&(E=pt(E,[0,4,1,2,3])),this.bias!==null&&(E=Ss(E,this.bias.read(),this.dataFormat)),this.activation!==null&&(E=this.activation.apply(E)),E})}computeOutputShape(t){t=de(t);const e=t.slice();let s,r,o,i;this.dataFormat==="channelsFirst"?(s=1,r=2,o=3,i=4):(s=4,r=1,o=2,i=3);const a=this.kernelSize[0],l=this.kernelSize[1],u=this.kernelSize[2],c=this.strides[0],h=this.strides[1],f=this.strides[2];return e[s]=this.filters,e[r]=ke(e[r],c,a,this.padding),e[o]=ke(e[o],h,l,this.padding),e[i]=ke(e[i],f,u,this.padding),e}getConfig(){const t=super.getConfig();return delete t.dilationRate,t}}ec.className="Conv3DTranspose",M(ec);class nc extends Kn{constructor(t,e){if(super(t,e),this.DEFAULT_DEPTHWISE_INITIALIZER="glorotUniform",this.DEFAULT_POINTWISE_INITIALIZER="glorotUniform",this.depthwiseKernel=null,this.pointwiseKernel=null,e.filters==null)throw new I("The `filters` configuration field is required by SeparableConv, but is unspecified.");if(e.kernelInitializer!=null||e.kernelRegularizer!=null||e.kernelConstraint!=null)throw new I("Fields kernelInitializer, kernelRegularizer and kernelConstraint are invalid for SeparableConv2D. Use depthwiseInitializer, depthwiseRegularizer, depthwiseConstraint, pointwiseInitializer, pointwiseRegularizer and pointwiseConstraint instead.");if(e.padding!=null&&e.padding!=="same"&&e.padding!=="valid")throw new I(`SeparableConv${this.rank}D supports only padding modes: 'same' and 'valid', but received ${JSON.stringify(e.padding)}`);this.depthMultiplier=e.depthMultiplier==null?1:e.depthMultiplier,this.depthwiseInitializer=vs(e.depthwiseInitializer||this.DEFAULT_DEPTHWISE_INITIALIZER),this.depthwiseRegularizer=As(e.depthwiseRegularizer),this.depthwiseConstraint=Sr(e.depthwiseConstraint),this.pointwiseInitializer=vs(e.depthwiseInitializer||this.DEFAULT_POINTWISE_INITIALIZER),this.pointwiseRegularizer=As(e.pointwiseRegularizer),this.pointwiseConstraint=Sr(e.pointwiseConstraint)}build(t){if(t=de(t),t.length<this.rank+2)throw new I(`Inputs to SeparableConv${this.rank}D should have rank ${this.rank+2}, but received input shape: ${JSON.stringify(t)}`);const e=this.dataFormat==="channelsFirst"?1:t.length-1;if(t[e]==null||t[e]<0)throw new I(`The channel dimension of the inputs should be defined, but found ${JSON.stringify(t[e])}`);const s=t[e],r=this.kernelSize.concat([s,this.depthMultiplier]),o=[];for(let a=0;a<this.rank;++a)o.push(1);o.push(s*this.depthMultiplier,this.filters);const i=!0;this.depthwiseKernel=this.addWeight("depthwise_kernel",r,"float32",this.depthwiseInitializer,this.depthwiseRegularizer,i,this.depthwiseConstraint),this.pointwiseKernel=this.addWeight("pointwise_kernel",o,"float32",this.pointwiseInitializer,this.pointwiseRegularizer,i,this.pointwiseConstraint),this.useBias?this.bias=this.addWeight("bias",[this.filters],"float32",this.biasInitializer,this.biasRegularizer,i,this.biasConstraint):this.bias=null,this.inputSpec=[new Ce({ndim:this.rank+2,axes:{[e]:s}})],this.built=!0}call(t,e){return k(()=>{t=Gt(t);let s;if(this.rank===1)throw new Z("1D separable convolution is not implemented yet.");return this.rank===2&&(this.dataFormat==="channelsFirst"&&(t=pt(t,[0,2,3,1])),s=O0(t,this.depthwiseKernel.read(),this.pointwiseKernel.read(),this.strides,this.padding,this.dilationRate,"NHWC")),this.useBias&&(s=Ss(s,this.bias.read(),this.dataFormat)),this.activation!=null&&(s=this.activation.apply(s)),this.dataFormat==="channelsFirst"&&(s=pt(s,[0,3,1,2])),s})}getConfig(){const t=super.getConfig();return delete t.rank,delete t.kernelInitializer,delete t.kernelRegularizer,delete t.kernelConstraint,t.depthwiseInitializer=vr(this.depthwiseInitializer),t.pointwiseInitializer=vr(this.pointwiseInitializer),t.depthwiseRegularizer=Is(this.depthwiseRegularizer),t.pointwiseRegularizer=Is(this.pointwiseRegularizer),t.depthwiseConstraint=xr(this.depthwiseConstraint),t.pointwiseConstraint=xr(this.pointwiseConstraint),t}}nc.className="SeparableConv";class sc extends nc{constructor(t){super(2,t)}}sc.className="SeparableConv2D",M(sc);class Er extends Kn{constructor(t){super(1,t),Er.verifyArgs(t),this.inputSpec=[{ndim:3}]}getConfig(){const t=super.getConfig();return delete t.rank,delete t.dataFormat,t}static verifyArgs(t){if(typeof t.kernelSize!="number"&&!ai(t.kernelSize,"number",1,1))throw new I(`Conv1D expects config.kernelSize to be number or number[] with length 1, but received ${JSON.stringify(t.kernelSize)}.`)}}Er.className="Conv1D",M(Er);class rc extends pe{constructor(t){super(t),typeof t.cropping=="number"?this.cropping=[[t.cropping,t.cropping],[t.cropping,t.cropping]]:typeof t.cropping[0]=="number"?this.cropping=[[t.cropping[0],t.cropping[0]],[t.cropping[1],t.cropping[1]]]:this.cropping=t.cropping,this.dataFormat=t.dataFormat===void 0?"channelsLast":t.dataFormat,this.inputSpec=[{ndim:4}]}computeOutputShape(t){return this.dataFormat==="channelsFirst"?[t[0],t[1],t[2]-this.cropping[0][0]-this.cropping[0][1],t[3]-this.cropping[1][0]-this.cropping[1][1]]:[t[0],t[1]-this.cropping[0][0]-this.cropping[0][1],t[2]-this.cropping[1][0]-this.cropping[1][1],t[3]]}call(t,e){return k(()=>{if(t=Gt(t),this.dataFormat==="channelsLast"){const s=wr(t,this.cropping[0][0],t.shape[1]-this.cropping[0][0]-this.cropping[0][1],2);return wr(s,this.cropping[1][0],t.shape[2]-this.cropping[1][1]-this.cropping[1][0],3)}else{const s=wr(t,this.cropping[0][0],t.shape[2]-this.cropping[0][0]-this.cropping[0][1],3);return wr(s,this.cropping[1][0],t.shape[3]-this.cropping[1][1]-this.cropping[1][0],4)}})}getConfig(){const t={cropping:this.cropping,dataFormat:this.dataFormat},e=super.getConfig();return Object.assign(t,e),t}}rc.className="Cropping2D",M(rc);class vi extends pe{constructor(t){super(t),this.DEFAULT_SIZE=[2,2],this.inputSpec=[{ndim:4}],this.size=t.size==null?this.DEFAULT_SIZE:t.size,this.dataFormat=t.dataFormat==null?"channelsLast":t.dataFormat,mt(this.dataFormat),this.interpolation=t.interpolation==null?"nearest":t.interpolation,Wy(this.interpolation)}computeOutputShape(t){if(this.dataFormat==="channelsFirst"){const e=t[2]==null?null:this.size[0]*t[2],s=t[3]==null?null:this.size[1]*t[3];return[t[0],t[1],e,s]}else{const e=t[1]==null?null:this.size[0]*t[1],s=t[2]==null?null:this.size[1]*t[2];return[t[0],e,s,t[3]]}}call(t,e){return k(()=>{let s=Gt(t);const r=s.shape;if(this.dataFormat==="channelsFirst"){s=pt(s,[0,2,3,1]);const o=this.size[0]*r[2],i=this.size[1]*r[3],a=this.interpolation==="nearest"?pr.resizeNearestNeighbor(s,[o,i]):pr.resizeBilinear(s,[o,i]);return pt(a,[0,3,1,2])}else{const o=this.size[0]*r[1],i=this.size[1]*r[2];return this.interpolation==="nearest"?pr.resizeNearestNeighbor(s,[o,i]):pr.resizeBilinear(s,[o,i])}})}getConfig(){const t={size:this.size,dataFormat:this.dataFormat,interpolation:this.interpolation},e=super.getConfig();return Object.assign(t,e),t}}vi.className="UpSampling2D",M(vi);/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2900,7 +2900,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Er(n,t,e,s,r,o){return k(()=>{mt(r),su(o),oe(s),e==null&&(e=[1,1]),s==null&&(s="valid"),r==null&&(r=Hn()),o==null&&(o="max"),n=Ju(n,r);let i;const a=s==="same"?"same":"valid";return o==="max"?i=Zg(n,t,e,a):i=Lm(n,t,e,a),r==="channelsFirst"&&(i=pt(i,[0,3,1,2])),i})}function oc(n,t,e,s,r,o){return k(()=>{mt(r),su(o),oe(s),e==null&&(e=[1,1,1]),s==null&&(s="valid"),r==null&&(r=Hn()),o==null&&(o="max"),n=Zu(n,r);let i;const a=s==="same"?"same":"valid";return o==="max"?i=t0(n,t,e,a):i=Om(n,t,e,a),r==="channelsFirst"&&(i=pt(i,[0,4,1,2,3])),i})}class ic extends pe{constructor(t){if(t.poolSize==null&&(t.poolSize=2),super(t),typeof t.poolSize=="number")this.poolSize=[t.poolSize];else if(Array.isArray(t.poolSize)&&t.poolSize.length===1&&typeof t.poolSize[0]=="number")this.poolSize=t.poolSize;else throw new I(`poolSize for 1D convolutional layer must be a number or an Array of a single number, but received ${JSON.stringify(t.poolSize)}`);if(Me(this.poolSize,"poolSize"),t.strides==null)this.strides=this.poolSize;else if(typeof t.strides=="number")this.strides=[t.strides];else if(Array.isArray(t.strides)&&t.strides.length===1&&typeof t.strides[0]=="number")this.strides=t.strides;else throw new I(`strides for 1D convolutional layer must be a number or an Array of a single number, but received ${JSON.stringify(t.strides)}`);Me(this.strides,"strides"),this.padding=t.padding==null?"valid":t.padding,oe(this.padding),this.inputSpec=[new Ce({ndim:3})]}computeOutputShape(t){t=de(t);const e=yn(t[1],this.poolSize[0],this.padding,this.strides[0]);return[t[0],e,t[2]]}call(t,e){return k(()=>{this.invokeCallHook(t,e),t=ui(Gt(t),2);const s=this.poolingFunction(Gt(t),[this.poolSize[0],1],[this.strides[0],1],this.padding,"channelsLast");return hr(s,[2])})}getConfig(){const t={poolSize:this.poolSize,padding:this.padding,strides:this.strides},e=super.getConfig();return Object.assign(t,e),t}}class ac extends ic{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),Er(t,e,s,r,o,"max")}}ac.className="MaxPooling1D",M(ac);class lc extends ic{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),Er(t,e,s,r,o,"avg")}}lc.className="AveragePooling1D",M(lc);class uc extends pe{constructor(t){if(t.poolSize==null&&(t.poolSize=[2,2]),super(t),this.poolSize=Array.isArray(t.poolSize)?t.poolSize:[t.poolSize,t.poolSize],t.strides==null)this.strides=this.poolSize;else if(Array.isArray(t.strides)){if(t.strides.length!==2)throw new I(`If the strides property of a 2D pooling layer is an Array, it is expected to have a length of 2, but received length ${t.strides.length}.`);this.strides=t.strides}else this.strides=[t.strides,t.strides];Me(this.poolSize,"poolSize"),Me(this.strides,"strides"),this.padding=t.padding==null?"valid":t.padding,this.dataFormat=t.dataFormat==null?"channelsLast":t.dataFormat,mt(this.dataFormat),oe(this.padding),this.inputSpec=[new Ce({ndim:4})]}computeOutputShape(t){t=de(t);let e=this.dataFormat==="channelsFirst"?t[2]:t[1],s=this.dataFormat==="channelsFirst"?t[3]:t[2];return e=yn(e,this.poolSize[0],this.padding,this.strides[0]),s=yn(s,this.poolSize[1],this.padding,this.strides[1]),this.dataFormat==="channelsFirst"?[t[0],t[1],e,s]:[t[0],e,s,t[3]]}call(t,e){return k(()=>(this.invokeCallHook(t,e),this.poolingFunction(Gt(t),this.poolSize,this.strides,this.padding,this.dataFormat)))}getConfig(){const t={poolSize:this.poolSize,padding:this.padding,strides:this.strides,dataFormat:this.dataFormat},e=super.getConfig();return Object.assign(t,e),t}}class Ii extends uc{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),Er(t,e,s,r,o,"max")}}Ii.className="MaxPooling2D",M(Ii);class cc extends uc{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),Er(t,e,s,r,o,"avg")}}cc.className="AveragePooling2D",M(cc);class hc extends pe{constructor(t){if(t.poolSize==null&&(t.poolSize=[2,2,2]),super(t),this.poolSize=Array.isArray(t.poolSize)?t.poolSize:[t.poolSize,t.poolSize,t.poolSize],t.strides==null)this.strides=this.poolSize;else if(Array.isArray(t.strides)){if(t.strides.length!==3)throw new I(`If the strides property of a 3D pooling layer is an Array, it is expected to have a length of 3, but received length ${t.strides.length}.`);this.strides=t.strides}else this.strides=[t.strides,t.strides,t.strides];Me(this.poolSize,"poolSize"),Me(this.strides,"strides"),this.padding=t.padding==null?"valid":t.padding,this.dataFormat=t.dataFormat==null?"channelsLast":t.dataFormat,mt(this.dataFormat),oe(this.padding),this.inputSpec=[new Ce({ndim:5})]}computeOutputShape(t){t=de(t);let e=this.dataFormat==="channelsFirst"?t[2]:t[1],s=this.dataFormat==="channelsFirst"?t[3]:t[2],r=this.dataFormat==="channelsFirst"?t[4]:t[3];return e=yn(e,this.poolSize[0],this.padding,this.strides[0]),s=yn(s,this.poolSize[1],this.padding,this.strides[1]),r=yn(r,this.poolSize[2],this.padding,this.strides[2]),this.dataFormat==="channelsFirst"?[t[0],t[1],e,s,r]:[t[0],e,s,r,t[4]]}call(t,e){return k(()=>(this.invokeCallHook(t,e),this.poolingFunction(Gt(t),this.poolSize,this.strides,this.padding,this.dataFormat)))}getConfig(){const t={poolSize:this.poolSize,padding:this.padding,strides:this.strides,dataFormat:this.dataFormat},e=super.getConfig();return Object.assign(t,e),t}}class fc extends hc{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),oc(t,e,s,r,o,"max")}}fc.className="MaxPooling3D",M(fc);class dc extends hc{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),oc(t,e,s,r,o,"avg")}}dc.className="AveragePooling3D",M(dc);class pc extends pe{constructor(t){super(t),this.inputSpec=[new Ce({ndim:3})]}computeOutputShape(t){return[t[0],t[2]]}call(t,e){throw new Z}}class mc extends pc{constructor(t){super(t||{})}call(t,e){return k(()=>{const s=Gt(t);return St(s,1)})}}mc.className="GlobalAveragePooling1D",M(mc);class gc extends pc{constructor(t){super(t||{})}call(t,e){return k(()=>{const s=Gt(t);return Ge(s,1)})}}gc.className="GlobalMaxPooling1D",M(gc);class bc extends pe{constructor(t){super(t),this.dataFormat=t.dataFormat==null?"channelsLast":t.dataFormat,mt(this.dataFormat),this.inputSpec=[new Ce({ndim:4})]}computeOutputShape(t){return t=t,this.dataFormat==="channelsLast"?[t[0],t[3]]:[t[0],t[1]]}call(t,e){throw new Z}getConfig(){const t={dataFormat:this.dataFormat},e=super.getConfig();return Object.assign(t,e),t}}class yc extends bc{call(t,e){return k(()=>{const s=Gt(t);return this.dataFormat==="channelsLast"?St(s,[1,2]):St(s,[2,3])})}}yc.className="GlobalAveragePooling2D",M(yc);class wc extends bc{call(t,e){return k(()=>{const s=Gt(t);return this.dataFormat==="channelsLast"?Ge(s,[1,2]):Ge(s,[2,3])})}}wc.className="GlobalMaxPooling2D",M(wc);/**
+ */function Cr(n,t,e,s,r,o){return k(()=>{mt(r),su(o),oe(s),e==null&&(e=[1,1]),s==null&&(s="valid"),r==null&&(r=Hn()),o==null&&(o="max"),n=Ju(n,r);let i;const a=s==="same"?"same":"valid";return o==="max"?i=Zg(n,t,e,a):i=Lm(n,t,e,a),r==="channelsFirst"&&(i=pt(i,[0,3,1,2])),i})}function oc(n,t,e,s,r,o){return k(()=>{mt(r),su(o),oe(s),e==null&&(e=[1,1,1]),s==null&&(s="valid"),r==null&&(r=Hn()),o==null&&(o="max"),n=Zu(n,r);let i;const a=s==="same"?"same":"valid";return o==="max"?i=t0(n,t,e,a):i=Om(n,t,e,a),r==="channelsFirst"&&(i=pt(i,[0,4,1,2,3])),i})}class ic extends pe{constructor(t){if(t.poolSize==null&&(t.poolSize=2),super(t),typeof t.poolSize=="number")this.poolSize=[t.poolSize];else if(Array.isArray(t.poolSize)&&t.poolSize.length===1&&typeof t.poolSize[0]=="number")this.poolSize=t.poolSize;else throw new I(`poolSize for 1D convolutional layer must be a number or an Array of a single number, but received ${JSON.stringify(t.poolSize)}`);if(Me(this.poolSize,"poolSize"),t.strides==null)this.strides=this.poolSize;else if(typeof t.strides=="number")this.strides=[t.strides];else if(Array.isArray(t.strides)&&t.strides.length===1&&typeof t.strides[0]=="number")this.strides=t.strides;else throw new I(`strides for 1D convolutional layer must be a number or an Array of a single number, but received ${JSON.stringify(t.strides)}`);Me(this.strides,"strides"),this.padding=t.padding==null?"valid":t.padding,oe(this.padding),this.inputSpec=[new Ce({ndim:3})]}computeOutputShape(t){t=de(t);const e=wn(t[1],this.poolSize[0],this.padding,this.strides[0]);return[t[0],e,t[2]]}call(t,e){return k(()=>{this.invokeCallHook(t,e),t=ui(Gt(t),2);const s=this.poolingFunction(Gt(t),[this.poolSize[0],1],[this.strides[0],1],this.padding,"channelsLast");return fr(s,[2])})}getConfig(){const t={poolSize:this.poolSize,padding:this.padding,strides:this.strides},e=super.getConfig();return Object.assign(t,e),t}}class ac extends ic{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),Cr(t,e,s,r,o,"max")}}ac.className="MaxPooling1D",M(ac);class lc extends ic{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),Cr(t,e,s,r,o,"avg")}}lc.className="AveragePooling1D",M(lc);class uc extends pe{constructor(t){if(t.poolSize==null&&(t.poolSize=[2,2]),super(t),this.poolSize=Array.isArray(t.poolSize)?t.poolSize:[t.poolSize,t.poolSize],t.strides==null)this.strides=this.poolSize;else if(Array.isArray(t.strides)){if(t.strides.length!==2)throw new I(`If the strides property of a 2D pooling layer is an Array, it is expected to have a length of 2, but received length ${t.strides.length}.`);this.strides=t.strides}else this.strides=[t.strides,t.strides];Me(this.poolSize,"poolSize"),Me(this.strides,"strides"),this.padding=t.padding==null?"valid":t.padding,this.dataFormat=t.dataFormat==null?"channelsLast":t.dataFormat,mt(this.dataFormat),oe(this.padding),this.inputSpec=[new Ce({ndim:4})]}computeOutputShape(t){t=de(t);let e=this.dataFormat==="channelsFirst"?t[2]:t[1],s=this.dataFormat==="channelsFirst"?t[3]:t[2];return e=wn(e,this.poolSize[0],this.padding,this.strides[0]),s=wn(s,this.poolSize[1],this.padding,this.strides[1]),this.dataFormat==="channelsFirst"?[t[0],t[1],e,s]:[t[0],e,s,t[3]]}call(t,e){return k(()=>(this.invokeCallHook(t,e),this.poolingFunction(Gt(t),this.poolSize,this.strides,this.padding,this.dataFormat)))}getConfig(){const t={poolSize:this.poolSize,padding:this.padding,strides:this.strides,dataFormat:this.dataFormat},e=super.getConfig();return Object.assign(t,e),t}}class Ii extends uc{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),Cr(t,e,s,r,o,"max")}}Ii.className="MaxPooling2D",M(Ii);class cc extends uc{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),Cr(t,e,s,r,o,"avg")}}cc.className="AveragePooling2D",M(cc);class hc extends pe{constructor(t){if(t.poolSize==null&&(t.poolSize=[2,2,2]),super(t),this.poolSize=Array.isArray(t.poolSize)?t.poolSize:[t.poolSize,t.poolSize,t.poolSize],t.strides==null)this.strides=this.poolSize;else if(Array.isArray(t.strides)){if(t.strides.length!==3)throw new I(`If the strides property of a 3D pooling layer is an Array, it is expected to have a length of 3, but received length ${t.strides.length}.`);this.strides=t.strides}else this.strides=[t.strides,t.strides,t.strides];Me(this.poolSize,"poolSize"),Me(this.strides,"strides"),this.padding=t.padding==null?"valid":t.padding,this.dataFormat=t.dataFormat==null?"channelsLast":t.dataFormat,mt(this.dataFormat),oe(this.padding),this.inputSpec=[new Ce({ndim:5})]}computeOutputShape(t){t=de(t);let e=this.dataFormat==="channelsFirst"?t[2]:t[1],s=this.dataFormat==="channelsFirst"?t[3]:t[2],r=this.dataFormat==="channelsFirst"?t[4]:t[3];return e=wn(e,this.poolSize[0],this.padding,this.strides[0]),s=wn(s,this.poolSize[1],this.padding,this.strides[1]),r=wn(r,this.poolSize[2],this.padding,this.strides[2]),this.dataFormat==="channelsFirst"?[t[0],t[1],e,s,r]:[t[0],e,s,r,t[4]]}call(t,e){return k(()=>(this.invokeCallHook(t,e),this.poolingFunction(Gt(t),this.poolSize,this.strides,this.padding,this.dataFormat)))}getConfig(){const t={poolSize:this.poolSize,padding:this.padding,strides:this.strides,dataFormat:this.dataFormat},e=super.getConfig();return Object.assign(t,e),t}}class fc extends hc{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),oc(t,e,s,r,o,"max")}}fc.className="MaxPooling3D",M(fc);class dc extends hc{constructor(t){super(t)}poolingFunction(t,e,s,r,o){return mt(o),oe(r),oc(t,e,s,r,o,"avg")}}dc.className="AveragePooling3D",M(dc);class pc extends pe{constructor(t){super(t),this.inputSpec=[new Ce({ndim:3})]}computeOutputShape(t){return[t[0],t[2]]}call(t,e){throw new Z}}class mc extends pc{constructor(t){super(t||{})}call(t,e){return k(()=>{const s=Gt(t);return St(s,1)})}}mc.className="GlobalAveragePooling1D",M(mc);class gc extends pc{constructor(t){super(t||{})}call(t,e){return k(()=>{const s=Gt(t);return Ge(s,1)})}}gc.className="GlobalMaxPooling1D",M(gc);class bc extends pe{constructor(t){super(t),this.dataFormat=t.dataFormat==null?"channelsLast":t.dataFormat,mt(this.dataFormat),this.inputSpec=[new Ce({ndim:4})]}computeOutputShape(t){return t=t,this.dataFormat==="channelsLast"?[t[0],t[3]]:[t[0],t[1]]}call(t,e){throw new Z}getConfig(){const t={dataFormat:this.dataFormat},e=super.getConfig();return Object.assign(t,e),t}}class yc extends bc{call(t,e){return k(()=>{const s=Gt(t);return this.dataFormat==="channelsLast"?St(s,[1,2]):St(s,[2,3])})}}yc.className="GlobalAveragePooling2D",M(yc);class wc extends bc{call(t,e){return k(()=>{const s=Gt(t);return this.dataFormat==="channelsLast"?Ge(s,[1,2]):Ge(s,[2,3])})}}wc.className="GlobalMaxPooling2D",M(wc);/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2908,7 +2908,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Cr(n,t){return k(()=>{n.dtype!=="float32"&&(n=ot(n,"float32"));const e=et(ws(n),t,!0),s=rr(e.shape,gt()),r=fe(zn(e,s));return X(n,r)})}function kr(n,t){return k(()=>St(ws(J(t,n)),-1))}function Ai(n,t){return k(()=>St(Lt(J(t,n)),-1))}function Ei(n,t){return k(()=>{const e=J(n,t),s=he(Lt(n),gt(),Number.MAX_VALUE),r=Lt(X(e,s));return N(100,St(r,-1))})}function yw(n,t){return k(()=>{const e=he(t,gt(),Number.MAX_VALUE),s=fn(O(1,e)),r=he(n,gt(),Number.MAX_VALUE),o=fn(O(1,r));return St(ws(J(s,o)),-1)})}function ww(n,t){return k(()=>{const e=zn(0,J(1,N(n,t)));return St(ws(e),-1)})}function xw(n,t){return k(()=>{const e=zn(0,J(1,N(n,t)));return St(e,-1)})}function Sw(n,t){return k(()=>{const e=et(N(n,t),-1),s=Ge(N(J(1,n),t),-1);return zn(0,O(1,J(s,e)))})}function $w(n,t){return k(()=>{const e=Math.log(2),s=J(t,n),r=J(O(s,qo(N(-2,s))),e);return St(r,-1)})}function Es(n,t,e=!1){return k(()=>{if(e)t=zl(t);else{const s=et(t,t.shape.length-1,!0);t=X(t,s)}return t=he(t,gt(),1-gt()),Fn(et(N(ot(n,"float32"),fn(t)),t.shape.length-1))})}function _r(n,t,e=!1){return k(()=>{const s=ot(_g(jy(n)),"int32");t=he(t,gt(),1-gt());const r=t.shape,o=L(o0(s,r[r.length-1]),r);return Es(o,t,e)})}function vw(n,t){if(!Zt(n.shape,t.shape))throw new I(`logits and labels must have the same shape, but got shapes ${JSON.stringify(n.shape)} and ${JSON.stringify(t.shape)}`);return k(()=>{const e=ps(t),s=Fn(Lt(t));return O(J(e,N(t,n)),Gg(Go(s)))})}function Tr(n,t){return k(()=>{let e;return e=he(t,gt(),1-gt()),e=fn(X(e,J(1,e))),St(vw(n,e),-1)})}function Iw(n,t){return k(()=>{const e=he(n,gt(),1),s=he(t,gt(),1);return et(N(n,fn(X(e,s))),-1)})}function Aw(n,t){return k(()=>{const e=fn(O(gt(),t));return St(J(t,N(n,e)),-1)})}function xc(n,t){return k(()=>{const e=Cr(n,-1),s=Cr(t,-1),r=N(e,s);return Fn(et(r,-1))})}const Nr={meanSquaredError:kr,meanAbsoluteError:Ai,meanAbsolutePercentageError:Ei,meanSquaredLogarithmicError:yw,squaredHinge:ww,hinge:xw,categoricalHinge:Sw,logcosh:$w,categoricalCrossentropy:Es,sparseCategoricalCrossentropy:_r,binaryCrossentropy:Tr,kullbackLeiblerDivergence:Iw,poisson:Aw,cosineProximity:xc};function Ci(n){if(typeof n=="string"){if(n in Nr)return Nr[n];let t=`Unknown loss ${n}`;throw n.toLowerCase().includes("softmaxcrossentropy")&&(t=`Unknown loss ${n}. Use "categoricalCrossentropy" as the string name for tf.losses.softmaxCrossEntropy`),new I(t)}else return n}/**
+ */function kr(n,t){return k(()=>{n.dtype!=="float32"&&(n=ot(n,"float32"));const e=et(xs(n),t,!0),s=or(e.shape,gt()),r=fe(zn(e,s));return X(n,r)})}function _r(n,t){return k(()=>St(xs(J(t,n)),-1))}function Ai(n,t){return k(()=>St(Lt(J(t,n)),-1))}function Ei(n,t){return k(()=>{const e=J(n,t),s=he(Lt(n),gt(),Number.MAX_VALUE),r=Lt(X(e,s));return N(100,St(r,-1))})}function yw(n,t){return k(()=>{const e=he(t,gt(),Number.MAX_VALUE),s=dn(O(1,e)),r=he(n,gt(),Number.MAX_VALUE),o=dn(O(1,r));return St(xs(J(s,o)),-1)})}function ww(n,t){return k(()=>{const e=zn(0,J(1,N(n,t)));return St(xs(e),-1)})}function xw(n,t){return k(()=>{const e=zn(0,J(1,N(n,t)));return St(e,-1)})}function Sw(n,t){return k(()=>{const e=et(N(n,t),-1),s=Ge(N(J(1,n),t),-1);return zn(0,O(1,J(s,e)))})}function $w(n,t){return k(()=>{const e=Math.log(2),s=J(t,n),r=J(O(s,qo(N(-2,s))),e);return St(r,-1)})}function Cs(n,t,e=!1){return k(()=>{if(e)t=zl(t);else{const s=et(t,t.shape.length-1,!0);t=X(t,s)}return t=he(t,gt(),1-gt()),Fn(et(N(ot(n,"float32"),dn(t)),t.shape.length-1))})}function Tr(n,t,e=!1){return k(()=>{const s=ot(_g(jy(n)),"int32");t=he(t,gt(),1-gt());const r=t.shape,o=L(o0(s,r[r.length-1]),r);return Cs(o,t,e)})}function vw(n,t){if(!Zt(n.shape,t.shape))throw new I(`logits and labels must have the same shape, but got shapes ${JSON.stringify(n.shape)} and ${JSON.stringify(t.shape)}`);return k(()=>{const e=ms(t),s=Fn(Lt(t));return O(J(e,N(t,n)),Gg(Go(s)))})}function Nr(n,t){return k(()=>{let e;return e=he(t,gt(),1-gt()),e=dn(X(e,J(1,e))),St(vw(n,e),-1)})}function Iw(n,t){return k(()=>{const e=he(n,gt(),1),s=he(t,gt(),1);return et(N(n,dn(X(e,s))),-1)})}function Aw(n,t){return k(()=>{const e=dn(O(gt(),t));return St(J(t,N(n,e)),-1)})}function xc(n,t){return k(()=>{const e=kr(n,-1),s=kr(t,-1),r=N(e,s);return Fn(et(r,-1))})}const Dr={meanSquaredError:_r,meanAbsoluteError:Ai,meanAbsolutePercentageError:Ei,meanSquaredLogarithmicError:yw,squaredHinge:ww,hinge:xw,categoricalHinge:Sw,logcosh:$w,categoricalCrossentropy:Cs,sparseCategoricalCrossentropy:Tr,binaryCrossentropy:Nr,kullbackLeiblerDivergence:Iw,poisson:Aw,cosineProximity:xc};function Ci(n){if(typeof n=="string"){if(n in Dr)return Dr[n];let t=`Unknown loss ${n}`;throw n.toLowerCase().includes("softmaxcrossentropy")&&(t=`Unknown loss ${n}. Use "categoricalCrossentropy" as the string name for tf.losses.softmaxCrossEntropy`),new I(t)}else return n}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2916,7 +2916,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */class wn extends pe{constructor(t){super(t||{}),this.supportsMasking=!0}mergeFunction(t){throw new Z}computeElementwiseOpOutputShape(t,e){if(t==null||e==null)return null;if(t.length<e.length)return this.computeElementwiseOpOutputShape(e,t);if(e.length===0)return t;const s=t.slice(0,t.length-e.length);for(let r=0;r<e.length;++r){const o=t[t.length-e.length+r],i=e[r];if(o==null||i==null||o<0||i<0)s.push(null);else if(o===1)s.push(i);else if(i===1)s.push(o);else{if(o!==i)throw new I("Operands could not be broadcast together with shapes "+JSON.stringify(t)+" "+JSON.stringify(e));s.push(o)}}return s}build(t){if(Array.isArray(t)&&!Array.isArray(t[0])&&(t=[de(t)]),t=t,t.length<2)throw new I(`A merge layer should be called on an Array of at least 2 inputs. Got ${t.length} input(s).`);let e=[];for(const o of t)o!=null&&o[0]!==null&&e.push(o[0]);if(e=mn(e),e.length>1)throw new I(`Can not merge tensors with different batch sizes. Got tensors with shapes: ${JSON.stringify(t)}.`);let s=t[0]==null?null:t[0].slice(1);for(let o=1;o<t.length;++o){const i=t[o]==null?null:t[o].slice(1);s=this.computeElementwiseOpOutputShape(s,i)}const r=t.map(o=>o.length);t.indexOf(null)===-1&&mn(r).length===1?this.reshapeRequired=!1:this.reshapeRequired=!0}call(t,e){return k(()=>{if(t=t,this.reshapeRequired){const s=[],r=t.map(o=>o.rank);if(r.indexOf(null)===-1){const o=lu(r);for(let i of t){const a=i.rank;for(let l=0;l<o-a;++l)i=ui(i,1);s.push(i)}return this.mergeFunction(s)}else{let o=!1;for(const l of t){const u=l.rank;if(u==null){const c=l.shape,h=c[0],f=c.slice(1).concat([h]);let d=L(l,[h].concat(ys(c.slice(1))));d=pt(d,[1,0]),d=L(d,f),s.push(d),o=!0}else if(u>1){const c=br(1,u).concat([0]);s.push(pt(l,c)),o=!0}else s.push(l)}let i=this.mergeFunction(s);const a=i.rank;if(o){if(a==null){const l=i.shape,u=l.length,c=l[u-1],h=[c].concat(l.slice(0,l.length-1));i=L(pt(L(i,[-1,c]),[1,0]),h)}else if(a>1){const l=[a-1].concat(br(0,a-1));i=pt(i,l)}}return i}}else return this.mergeFunction(t)})}computeOutputShape(t){t=t;let e;t[0]==null?e=null:e=t[0].slice(1);for(let r=1;r<t.length;++r){const o=t[r]==null?null:t[r].slice(1);e=this.computeElementwiseOpOutputShape(e,o)}let s=[];for(const r of t)r!=null&&r[0]!==null&&s.push(r[0]);return s=mn(s),s.length===1?e=s.concat(e):e=[null].concat(e),e}computeMask(t,e){return k(()=>{if(e==null)return null;if(!Array.isArray(e))throw new I("`mask` should be an Array");if(!Array.isArray(t))throw new I("`inputs` should be an Array");if(e.length!==t.length)throw new I(`The Array 'inputs' and 'mask' are expected to have the same length, but have different lengths (${t.length} vs ${e.length})`);if(e.every(r=>r==null))return null;e=e.map(r=>r==null?r:ve(r,0));let s=e[0];for(let r=1;r<e.length-1;++r)s=lr(s,e[r]);return s})}}class Sc extends wn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0].clone();for(let s=1;s<t.length;++s)e=O(e,t[s]);return e})}}Sc.className="Add",M(Sc);class $c extends wn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0].clone();for(let s=1;s<t.length;++s)e=N(e,t[s]);return e})}}$c.className="Multiply",M($c);class vc extends wn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0].clone();for(let s=1;s<t.length;++s)e=O(e,t[s]);return N(1/t.length,e)})}}vc.className="Average",M(vc);class Ic extends wn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0];for(let s=1;s<t.length;++s)e=zn(e,t[s]);return e})}}Ic.className="Maximum",M(Ic);class Ac extends wn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0];for(let s=1;s<t.length;++s)e=ur(e,t[s]);return e})}}Ac.className="Minimum",M(Ac);class ki extends wn{constructor(t){super(t),this.DEFAULT_AXIS=-1,t==null&&(t={}),this.axis=t.axis==null?this.DEFAULT_AXIS:t.axis,this.supportsMasking=!0,this.reshapeRequired=!1}build(t){if(!(Array.isArray(t)&&Array.isArray(t[0]))||t.length===1)throw new I("A `Concatenate` layer should be called on a list of at least 2 inputs");t=t;let e=!0;for(const r of t)if(r!=null){e=!1;break}if(e)return;const s=[];for(let r=0;r<t.length;++r){const o=t[r].slice();o.splice(this.axis,1);let i=!1;for(const a of s)if(Zt(a,o)){i=!0;break}i||s.push(o)}if(s.length>1)throw new I("A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got input shapes: "+JSON.stringify(t))}mergeFunction(t){return k(()=>Hy(t,this.axis))}computeOutputShape(t){if(!(Array.isArray(t)&&Array.isArray(t[0])))throw new I("A `Concatenate` layer should be called on a list of inputs.");const e=t,s=e[0].slice(),r=this.axis<0?s.length+this.axis:this.axis;for(const o of e.slice(1)){if(s[r]==null||o[r]==null){s[r]=null;break}s[r]+=o[r]}return s}computeMask(t,e){if(e==null)return null;if(!Array.isArray(e))throw new I("`mask` should be an array for Concatenate");if(!Array.isArray(t))throw new I("`inputs` should be an array for Concatenate");if(e.length!==t.length)throw new I(`Mismatch in the length of mask (${e.length}) and the legnth of inputs (${t.length})`);return k(()=>{let s=!0;if(e.forEach(i=>{if(i!=null){s=!1;return}}),s)return null;const r=[];for(let i=0;i<t.length;++i)e[i]==null?r.push(ot(Pl(t[i]),"bool")):e[i].rank<t[i].rank?r.push(ve(e[i],-1)):r.push(e[i]);const o=on(r,this.axis);return Em(o,-1,!1)})}getConfig(){const t={axis:this.axis},e=super.getConfig();return Object.assign(t,e),t}}ki.className="Concatenate",M(ki);function Cs(n,t){for(;n<0;)n+=t;return n}function Ew(n,t,e){if(n.shape.length>3||t.shape.length>3)throw new Z("batchDot is not implemented for tensors of 4D or higher rank yet");if(w(n.shape.length>=2,()=>`batchDot requires the rank of x to be >= 2, but got ${n.shape.length}`),w(n.shape.length>=2,()=>`batchDot requires the rank of y to be >= 2, but got ${t.shape.length}`),typeof e=="number"&&(e=[e,e]),n.dtype==="complex64"||t.dtype==="complex64")throw new Z("batchDot is not implemented for complex64-type Tensors yet.");const s=n.shape.length,r=t.shape.length;e==null&&(e=[s-1,r-2]);const o=e;return k(()=>{let i;if(s>r){i=s-r;const l=[];for(let u=0;u<i;++u)l.push(1);t=L(t,t.shape.concat(l))}else if(r>s){i=r-s;const l=[];for(let u=0;u<i;++u)l.push(1);n=L(n,n.shape.concat(l))}else i=0;let a;if(n.shape.length===2&&t.shape.length===2)o[0]===o[1]?a=et(N(n,t),o[0]):a=et(N(pt(n,[1,0]),t),o[1]);else{const l=o[0]!==n.shape.length-1,u=o[1]===t.shape.length-1;a=Se(n,t,l,u)}if(i>0){let l;s>r?l=s+r-3:l=s-1;const u=[];for(let c=l;c<l+i;++c)u.push(c);a=hr(a,u)}return a.shape.length===1&&(a=ve(a,1)),a})}class Ec extends wn{constructor(t){super(t),this.axes=t.axes,this.normalize=t.normalize==null?!1:t.normalize,this.supportsMasking=!0,this.reshapeRequired=!1}build(t){w(Array.isArray(t)&&t.length===2&&Array.isArray(t[0])&&Array.isArray(t[1]),()=>"A `Dot` layer should be called on a list of exactly 2 inputs.");const e=t[0],s=t[1];if(e.length>3||s.length>3)throw new Z("Dot layer does not support tensors of 4D or higher rank yet.");const r=this.interpretAxes(e,s);if(e[r[0]]!==s[r[1]])throw new I(`Dimension incompatibility: ${e[r[0]]} !== ${s[r[1]]}`)}mergeFunction(t){if(t.length!==2)throw new I(`A \`Dot\` layer must be called on exactly 2 inputs, but received ${t.length} input(s).`);let e=t[0],s=t[1],r;return Array.isArray(this.axes)?r=this.axes.map((o,i)=>Cs(o,t[i].shape.length)):r=[Cs(this.axes,e.shape.length),Cs(this.axes,s.shape.length)],this.normalize&&(e=Cr(e,r[0]),s=Cr(s,r[1])),Ew(e,s,r)}interpretAxes(t,e){let s;return Array.isArray(this.axes)?s=this.axes:s=[Cs(this.axes,t.length),Cs(this.axes,e.length)],s}computeOutputShape(t){w(Array.isArray(t)&&t.length===2&&Array.isArray(t[0])&&Array.isArray(t[1]),()=>"A `Dot` layer should be called on a list of exactly 2 inputs.");const e=t[0].slice(),s=t[1].slice();if(e.length>3||s.length>3)throw new Z("Dot layer does not support tensors of 4D or higher rank yet.");const r=this.interpretAxes(e,s);e.splice(r[0],1),s.splice(r[1],1),s.splice(0,1);const o=e.concat(s);return o.length===1&&o.push(1),o}computeMask(t,e){return null}getConfig(){const t={axes:this.axes,normalize:this.normalize},e=super.getConfig();return Object.assign(t,e),t}}Ec.className="Dot",M(Ec);/**
+ */class xn extends pe{constructor(t){super(t||{}),this.supportsMasking=!0}mergeFunction(t){throw new Z}computeElementwiseOpOutputShape(t,e){if(t==null||e==null)return null;if(t.length<e.length)return this.computeElementwiseOpOutputShape(e,t);if(e.length===0)return t;const s=t.slice(0,t.length-e.length);for(let r=0;r<e.length;++r){const o=t[t.length-e.length+r],i=e[r];if(o==null||i==null||o<0||i<0)s.push(null);else if(o===1)s.push(i);else if(i===1)s.push(o);else{if(o!==i)throw new I("Operands could not be broadcast together with shapes "+JSON.stringify(t)+" "+JSON.stringify(e));s.push(o)}}return s}build(t){if(Array.isArray(t)&&!Array.isArray(t[0])&&(t=[de(t)]),t=t,t.length<2)throw new I(`A merge layer should be called on an Array of at least 2 inputs. Got ${t.length} input(s).`);let e=[];for(const o of t)o!=null&&o[0]!==null&&e.push(o[0]);if(e=gn(e),e.length>1)throw new I(`Can not merge tensors with different batch sizes. Got tensors with shapes: ${JSON.stringify(t)}.`);let s=t[0]==null?null:t[0].slice(1);for(let o=1;o<t.length;++o){const i=t[o]==null?null:t[o].slice(1);s=this.computeElementwiseOpOutputShape(s,i)}const r=t.map(o=>o.length);t.indexOf(null)===-1&&gn(r).length===1?this.reshapeRequired=!1:this.reshapeRequired=!0}call(t,e){return k(()=>{if(t=t,this.reshapeRequired){const s=[],r=t.map(o=>o.rank);if(r.indexOf(null)===-1){const o=lu(r);for(let i of t){const a=i.rank;for(let l=0;l<o-a;++l)i=ui(i,1);s.push(i)}return this.mergeFunction(s)}else{let o=!1;for(const l of t){const u=l.rank;if(u==null){const c=l.shape,h=c[0],f=c.slice(1).concat([h]);let d=L(l,[h].concat(ws(c.slice(1))));d=pt(d,[1,0]),d=L(d,f),s.push(d),o=!0}else if(u>1){const c=yr(1,u).concat([0]);s.push(pt(l,c)),o=!0}else s.push(l)}let i=this.mergeFunction(s);const a=i.rank;if(o){if(a==null){const l=i.shape,u=l.length,c=l[u-1],h=[c].concat(l.slice(0,l.length-1));i=L(pt(L(i,[-1,c]),[1,0]),h)}else if(a>1){const l=[a-1].concat(yr(0,a-1));i=pt(i,l)}}return i}}else return this.mergeFunction(t)})}computeOutputShape(t){t=t;let e;t[0]==null?e=null:e=t[0].slice(1);for(let r=1;r<t.length;++r){const o=t[r]==null?null:t[r].slice(1);e=this.computeElementwiseOpOutputShape(e,o)}let s=[];for(const r of t)r!=null&&r[0]!==null&&s.push(r[0]);return s=gn(s),s.length===1?e=s.concat(e):e=[null].concat(e),e}computeMask(t,e){return k(()=>{if(e==null)return null;if(!Array.isArray(e))throw new I("`mask` should be an Array");if(!Array.isArray(t))throw new I("`inputs` should be an Array");if(e.length!==t.length)throw new I(`The Array 'inputs' and 'mask' are expected to have the same length, but have different lengths (${t.length} vs ${e.length})`);if(e.every(r=>r==null))return null;e=e.map(r=>r==null?r:ve(r,0));let s=e[0];for(let r=1;r<e.length-1;++r)s=ur(s,e[r]);return s})}}class Sc extends xn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0].clone();for(let s=1;s<t.length;++s)e=O(e,t[s]);return e})}}Sc.className="Add",M(Sc);class $c extends xn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0].clone();for(let s=1;s<t.length;++s)e=N(e,t[s]);return e})}}$c.className="Multiply",M($c);class vc extends xn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0].clone();for(let s=1;s<t.length;++s)e=O(e,t[s]);return N(1/t.length,e)})}}vc.className="Average",M(vc);class Ic extends xn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0];for(let s=1;s<t.length;++s)e=zn(e,t[s]);return e})}}Ic.className="Maximum",M(Ic);class Ac extends xn{constructor(t){super(t)}mergeFunction(t){return k(()=>{let e=t[0];for(let s=1;s<t.length;++s)e=cr(e,t[s]);return e})}}Ac.className="Minimum",M(Ac);class ki extends xn{constructor(t){super(t),this.DEFAULT_AXIS=-1,t==null&&(t={}),this.axis=t.axis==null?this.DEFAULT_AXIS:t.axis,this.supportsMasking=!0,this.reshapeRequired=!1}build(t){if(!(Array.isArray(t)&&Array.isArray(t[0]))||t.length===1)throw new I("A `Concatenate` layer should be called on a list of at least 2 inputs");t=t;let e=!0;for(const r of t)if(r!=null){e=!1;break}if(e)return;const s=[];for(let r=0;r<t.length;++r){const o=t[r].slice();o.splice(this.axis,1);let i=!1;for(const a of s)if(Zt(a,o)){i=!0;break}i||s.push(o)}if(s.length>1)throw new I("A `Concatenate` layer requires inputs with matching shapes except for the concat axis. Got input shapes: "+JSON.stringify(t))}mergeFunction(t){return k(()=>Hy(t,this.axis))}computeOutputShape(t){if(!(Array.isArray(t)&&Array.isArray(t[0])))throw new I("A `Concatenate` layer should be called on a list of inputs.");const e=t,s=e[0].slice(),r=this.axis<0?s.length+this.axis:this.axis;for(const o of e.slice(1)){if(s[r]==null||o[r]==null){s[r]=null;break}s[r]+=o[r]}return s}computeMask(t,e){if(e==null)return null;if(!Array.isArray(e))throw new I("`mask` should be an array for Concatenate");if(!Array.isArray(t))throw new I("`inputs` should be an array for Concatenate");if(e.length!==t.length)throw new I(`Mismatch in the length of mask (${e.length}) and the legnth of inputs (${t.length})`);return k(()=>{let s=!0;if(e.forEach(i=>{if(i!=null){s=!1;return}}),s)return null;const r=[];for(let i=0;i<t.length;++i)e[i]==null?r.push(ot(Pl(t[i]),"bool")):e[i].rank<t[i].rank?r.push(ve(e[i],-1)):r.push(e[i]);const o=an(r,this.axis);return Em(o,-1,!1)})}getConfig(){const t={axis:this.axis},e=super.getConfig();return Object.assign(t,e),t}}ki.className="Concatenate",M(ki);function ks(n,t){for(;n<0;)n+=t;return n}function Ew(n,t,e){if(n.shape.length>3||t.shape.length>3)throw new Z("batchDot is not implemented for tensors of 4D or higher rank yet");if(w(n.shape.length>=2,()=>`batchDot requires the rank of x to be >= 2, but got ${n.shape.length}`),w(n.shape.length>=2,()=>`batchDot requires the rank of y to be >= 2, but got ${t.shape.length}`),typeof e=="number"&&(e=[e,e]),n.dtype==="complex64"||t.dtype==="complex64")throw new Z("batchDot is not implemented for complex64-type Tensors yet.");const s=n.shape.length,r=t.shape.length;e==null&&(e=[s-1,r-2]);const o=e;return k(()=>{let i;if(s>r){i=s-r;const l=[];for(let u=0;u<i;++u)l.push(1);t=L(t,t.shape.concat(l))}else if(r>s){i=r-s;const l=[];for(let u=0;u<i;++u)l.push(1);n=L(n,n.shape.concat(l))}else i=0;let a;if(n.shape.length===2&&t.shape.length===2)o[0]===o[1]?a=et(N(n,t),o[0]):a=et(N(pt(n,[1,0]),t),o[1]);else{const l=o[0]!==n.shape.length-1,u=o[1]===t.shape.length-1;a=Se(n,t,l,u)}if(i>0){let l;s>r?l=s+r-3:l=s-1;const u=[];for(let c=l;c<l+i;++c)u.push(c);a=fr(a,u)}return a.shape.length===1&&(a=ve(a,1)),a})}class Ec extends xn{constructor(t){super(t),this.axes=t.axes,this.normalize=t.normalize==null?!1:t.normalize,this.supportsMasking=!0,this.reshapeRequired=!1}build(t){w(Array.isArray(t)&&t.length===2&&Array.isArray(t[0])&&Array.isArray(t[1]),()=>"A `Dot` layer should be called on a list of exactly 2 inputs.");const e=t[0],s=t[1];if(e.length>3||s.length>3)throw new Z("Dot layer does not support tensors of 4D or higher rank yet.");const r=this.interpretAxes(e,s);if(e[r[0]]!==s[r[1]])throw new I(`Dimension incompatibility: ${e[r[0]]} !== ${s[r[1]]}`)}mergeFunction(t){if(t.length!==2)throw new I(`A \`Dot\` layer must be called on exactly 2 inputs, but received ${t.length} input(s).`);let e=t[0],s=t[1],r;return Array.isArray(this.axes)?r=this.axes.map((o,i)=>ks(o,t[i].shape.length)):r=[ks(this.axes,e.shape.length),ks(this.axes,s.shape.length)],this.normalize&&(e=kr(e,r[0]),s=kr(s,r[1])),Ew(e,s,r)}interpretAxes(t,e){let s;return Array.isArray(this.axes)?s=this.axes:s=[ks(this.axes,t.length),ks(this.axes,e.length)],s}computeOutputShape(t){w(Array.isArray(t)&&t.length===2&&Array.isArray(t[0])&&Array.isArray(t[1]),()=>"A `Dot` layer should be called on a list of exactly 2 inputs.");const e=t[0].slice(),s=t[1].slice();if(e.length>3||s.length>3)throw new Z("Dot layer does not support tensors of 4D or higher rank yet.");const r=this.interpretAxes(e,s);e.splice(r[0],1),s.splice(r[1],1),s.splice(0,1);const o=e.concat(s);return o.length===1&&o.push(1),o}computeMask(t,e){return null}getConfig(){const t={axes:this.axes,normalize:this.normalize},e=super.getConfig();return Object.assign(t,e),t}}Ec.className="Dot",M(Ec);/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2924,7 +2924,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */async function xn(n){if(n==null)return;const t=[],e=[],s=[];for(const r in n){const o=n[r];if(typeof o!="number"){const i=o;t.push(i.data()),e.push(r),s.push(i)}}if(t.length>0){const r=await Promise.all(t);for(let o=0;o<r.length;++o)n[e[o]]=r[o][0];ut(s)}}function Cc(n){if(n!=null)for(const t in n){const e=n[t];typeof e!="number"&&e.dispose()}}/**
+ */async function Sn(n){if(n==null)return;const t=[],e=[],s=[];for(const r in n){const o=n[r];if(typeof o!="number"){const i=o;t.push(i.data()),e.push(r),s.push(i)}}if(t.length>0){const r=await Promise.all(t);for(let o=0;o<r.length;++o)n[e[o]]=r[o][0];ut(s)}}function Cc(n){if(n!=null)for(const t in n){const e=n[t];typeof e!="number"&&e.dispose()}}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2932,7 +2932,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */var kc;(function(n){n[n.SILENT=0]="SILENT",n[n.VERBOSE=1]="VERBOSE"})(kc||(kc={}));const Cw=125;class ks{constructor(){this.validationData=null}setParams(t){this.params=t}async onEpochBegin(t,e){}async onEpochEnd(t,e){}async onBatchBegin(t,e){}async onBatchEnd(t,e){}async onTrainBegin(t){}async onTrainEnd(t){}setModel(t){}}class kw{constructor(t,e=10){t==null&&(t=[]),this.callbacks=t,this.queueLength=e}append(t){this.callbacks.push(t)}setParams(t){for(const e of this.callbacks)e.setParams(t)}setModel(t){for(const e of this.callbacks)e.setModel(t)}async onEpochBegin(t,e){e==null&&(e={});for(const s of this.callbacks)await s.onEpochBegin(t,e)}async onEpochEnd(t,e){e==null&&(e={});for(const s of this.callbacks)await s.onEpochEnd(t,e)}async onBatchBegin(t,e){e==null&&(e={});for(const s of this.callbacks)await s.onBatchBegin(t,e)}async onBatchEnd(t,e){e==null&&(e={});for(const s of this.callbacks)await s.onBatchEnd(t,e)}async onTrainBegin(t){t==null&&(t={});for(const e of this.callbacks)await e.onTrainBegin(t)}async onTrainEnd(t){t==null&&(t={});for(const e of this.callbacks)await e.onTrainEnd(t)}}class _w extends ks{constructor(){super()}async onEpochBegin(t){this.seen=0,this.totals={}}async onBatchEnd(t,e){e==null&&(e={});const s=e.size==null?0:e.size;this.seen+=s;for(const r in e){const o=e[r];if(typeof o=="number")this.totals.hasOwnProperty(r)||(this.totals[r]=0),this.totals[r]=this.totals[r]+o*s;else{let i;r in this.totals?i=this.totals[r]:this.totals[r]=0;const a=k(()=>O(this.totals[r],N(o,s)));this.totals[r]=a,i!=null&&i.dispose()}}}async onEpochEnd(t,e){if(e!=null)for(const s of this.params.metrics)this.totals[s]!=null&&(typeof this.totals[s]=="number"?e[s]=this.totals[s]/this.seen:k(()=>{const r=N(X(1,this.seen),this.totals[s]);e[s]=r,this.totals[s].dispose(),Ln(e[s])}))}}class Tw extends ks{async onTrainBegin(t){this.epoch=[],this.history={}}async onEpochEnd(t,e){e==null&&(e={}),this.epoch.push(t);for(const s in e)this.history[s]==null&&(this.history[s]=[]),this.history[s].push(e[s])}async syncData(){const t=[],e=[],s=[];for(const o in this.history){const i=this.history[o];for(let a=0;a<i.length;++a)if(typeof i[a]!="number"){const l=i[a];t.push(l.data()),e.push(o),s.push(a)}}const r=await Promise.all(t);for(let o=0;o<r.length;++o)this.history[e[o]][s[o]].dispose(),this.history[e[o]][s[o]]=r[o][0]}}class Nw extends ks{constructor(t,e){if(super(),this.currentEpoch=0,this.nowFunc=t.nowFunc,this.nextFrameFunc=t.nextFrameFunc||ly,this.yieldEvery=e||"auto",this.yieldEvery==="auto"&&(this.yieldEvery=Cw),this.yieldEvery==="never"&&t.onYield!=null)throw new Error("yieldEvery is `never` but you provided an `onYield` callback. Either change `yieldEvery` or remove the callback");ao(this.yieldEvery)&&(this.maybeWait=zy(this.maybeWait.bind(this),this.yieldEvery,this.nowFunc)),this.trainBegin=t.onTrainBegin,this.trainEnd=t.onTrainEnd,this.epochBegin=t.onEpochBegin,this.epochEnd=t.onEpochEnd,this.batchBegin=t.onBatchBegin,this.batchEnd=t.onBatchEnd,this.yield=t.onYield}async maybeWait(t,e,s){const r=[];this.yield!=null&&(await xn(s),r.push(this.yield(t,e,s))),r.push(this.nextFrameFunc()),await Promise.all(r)}async onEpochBegin(t,e){this.currentEpoch=t,this.epochBegin!=null&&(await xn(e),await this.epochBegin(t,e))}async onEpochEnd(t,e){const s=[];this.epochEnd!=null&&(await xn(e),s.push(this.epochEnd(t,e))),this.yieldEvery==="epoch"&&s.push(this.nextFrameFunc()),await Promise.all(s)}async onBatchBegin(t,e){this.batchBegin!=null&&(await xn(e),await this.batchBegin(t,e))}async onBatchEnd(t,e){const s=[];this.batchEnd!=null&&(await xn(e),s.push(this.batchEnd(t,e))),this.yieldEvery==="batch"?s.push(this.nextFrameFunc()):ao(this.yieldEvery)&&s.push(this.maybeWait(this.currentEpoch,t,e)),await Promise.all(s)}async onTrainBegin(t){this.trainBegin!=null&&(await xn(t),await this.trainBegin(t))}async onTrainEnd(t){this.trainEnd!=null&&(await xn(t),await this.trainEnd(t))}}function _c(n,t){return n==null&&(n={}),n instanceof ks?[n]:Array.isArray(n)&&n[0]instanceof ks?n:nt(n).map(s=>new Nw(s,t))}class ie{constructor(){}static registerCallbackConstructor(t,e){w(t>=0&&Number.isInteger(t),()=>`Verbosity level is expected to be an integer >= 0, but got ${t}`),ie.checkForDuplicate(e),ie.constructors[t]==null&&(ie.constructors[t]=[]),ie.constructors[t].push(e)}static checkForDuplicate(t){for(const e in ie.constructors)ie.constructors[+e].forEach(r=>{if(r===t)throw new I("Duplicate callback constructor.")})}static clear(){ie.constructors={}}static createCallbacks(t){const e=[];for(const s in ie.constructors){const r=+s;t>=r&&e.push(...ie.constructors[r])}return e.map(s=>new s)}}ie.constructors={};function Tc(n,t,e,s,r,o,i,a,l){const u=new Tw,c=[new _w,...ie.createCallbacks(t)];n!=null&&c.push(...n),c.push(u);const h=new kw(c);return h.setParams({epochs:e,initialEpoch:s,samples:r,steps:o,batchSize:i,verbose:t,doValidation:a,metrics:l}),{callbackList:h,history:u}}/**
+ */var kc;(function(n){n[n.SILENT=0]="SILENT",n[n.VERBOSE=1]="VERBOSE"})(kc||(kc={}));const Cw=125;class _s{constructor(){this.validationData=null}setParams(t){this.params=t}async onEpochBegin(t,e){}async onEpochEnd(t,e){}async onBatchBegin(t,e){}async onBatchEnd(t,e){}async onTrainBegin(t){}async onTrainEnd(t){}setModel(t){}}class kw{constructor(t,e=10){t==null&&(t=[]),this.callbacks=t,this.queueLength=e}append(t){this.callbacks.push(t)}setParams(t){for(const e of this.callbacks)e.setParams(t)}setModel(t){for(const e of this.callbacks)e.setModel(t)}async onEpochBegin(t,e){e==null&&(e={});for(const s of this.callbacks)await s.onEpochBegin(t,e)}async onEpochEnd(t,e){e==null&&(e={});for(const s of this.callbacks)await s.onEpochEnd(t,e)}async onBatchBegin(t,e){e==null&&(e={});for(const s of this.callbacks)await s.onBatchBegin(t,e)}async onBatchEnd(t,e){e==null&&(e={});for(const s of this.callbacks)await s.onBatchEnd(t,e)}async onTrainBegin(t){t==null&&(t={});for(const e of this.callbacks)await e.onTrainBegin(t)}async onTrainEnd(t){t==null&&(t={});for(const e of this.callbacks)await e.onTrainEnd(t)}}class _w extends _s{constructor(){super()}async onEpochBegin(t){this.seen=0,this.totals={}}async onBatchEnd(t,e){e==null&&(e={});const s=e.size==null?0:e.size;this.seen+=s;for(const r in e){const o=e[r];if(typeof o=="number")this.totals.hasOwnProperty(r)||(this.totals[r]=0),this.totals[r]=this.totals[r]+o*s;else{let i;r in this.totals?i=this.totals[r]:this.totals[r]=0;const a=k(()=>O(this.totals[r],N(o,s)));this.totals[r]=a,i!=null&&i.dispose()}}}async onEpochEnd(t,e){if(e!=null)for(const s of this.params.metrics)this.totals[s]!=null&&(typeof this.totals[s]=="number"?e[s]=this.totals[s]/this.seen:k(()=>{const r=N(X(1,this.seen),this.totals[s]);e[s]=r,this.totals[s].dispose(),Ln(e[s])}))}}class Tw extends _s{async onTrainBegin(t){this.epoch=[],this.history={}}async onEpochEnd(t,e){e==null&&(e={}),this.epoch.push(t);for(const s in e)this.history[s]==null&&(this.history[s]=[]),this.history[s].push(e[s])}async syncData(){const t=[],e=[],s=[];for(const o in this.history){const i=this.history[o];for(let a=0;a<i.length;++a)if(typeof i[a]!="number"){const l=i[a];t.push(l.data()),e.push(o),s.push(a)}}const r=await Promise.all(t);for(let o=0;o<r.length;++o)this.history[e[o]][s[o]].dispose(),this.history[e[o]][s[o]]=r[o][0]}}class Nw extends _s{constructor(t,e){if(super(),this.currentEpoch=0,this.nowFunc=t.nowFunc,this.nextFrameFunc=t.nextFrameFunc||ly,this.yieldEvery=e||"auto",this.yieldEvery==="auto"&&(this.yieldEvery=Cw),this.yieldEvery==="never"&&t.onYield!=null)throw new Error("yieldEvery is `never` but you provided an `onYield` callback. Either change `yieldEvery` or remove the callback");ao(this.yieldEvery)&&(this.maybeWait=zy(this.maybeWait.bind(this),this.yieldEvery,this.nowFunc)),this.trainBegin=t.onTrainBegin,this.trainEnd=t.onTrainEnd,this.epochBegin=t.onEpochBegin,this.epochEnd=t.onEpochEnd,this.batchBegin=t.onBatchBegin,this.batchEnd=t.onBatchEnd,this.yield=t.onYield}async maybeWait(t,e,s){const r=[];this.yield!=null&&(await Sn(s),r.push(this.yield(t,e,s))),r.push(this.nextFrameFunc()),await Promise.all(r)}async onEpochBegin(t,e){this.currentEpoch=t,this.epochBegin!=null&&(await Sn(e),await this.epochBegin(t,e))}async onEpochEnd(t,e){const s=[];this.epochEnd!=null&&(await Sn(e),s.push(this.epochEnd(t,e))),this.yieldEvery==="epoch"&&s.push(this.nextFrameFunc()),await Promise.all(s)}async onBatchBegin(t,e){this.batchBegin!=null&&(await Sn(e),await this.batchBegin(t,e))}async onBatchEnd(t,e){const s=[];this.batchEnd!=null&&(await Sn(e),s.push(this.batchEnd(t,e))),this.yieldEvery==="batch"?s.push(this.nextFrameFunc()):ao(this.yieldEvery)&&s.push(this.maybeWait(this.currentEpoch,t,e)),await Promise.all(s)}async onTrainBegin(t){this.trainBegin!=null&&(await Sn(t),await this.trainBegin(t))}async onTrainEnd(t){this.trainEnd!=null&&(await Sn(t),await this.trainEnd(t))}}function _c(n,t){return n==null&&(n={}),n instanceof _s?[n]:Array.isArray(n)&&n[0]instanceof _s?n:nt(n).map(s=>new Nw(s,t))}class ie{constructor(){}static registerCallbackConstructor(t,e){w(t>=0&&Number.isInteger(t),()=>`Verbosity level is expected to be an integer >= 0, but got ${t}`),ie.checkForDuplicate(e),ie.constructors[t]==null&&(ie.constructors[t]=[]),ie.constructors[t].push(e)}static checkForDuplicate(t){for(const e in ie.constructors)ie.constructors[+e].forEach(r=>{if(r===t)throw new I("Duplicate callback constructor.")})}static clear(){ie.constructors={}}static createCallbacks(t){const e=[];for(const s in ie.constructors){const r=+s;t>=r&&e.push(...ie.constructors[r])}return e.map(s=>new s)}}ie.constructors={};function Tc(n,t,e,s,r,o,i,a,l){const u=new Tw,c=[new _w,...ie.createCallbacks(t)];n!=null&&c.push(...n),c.push(u);const h=new kw(c);return h.setParams({epochs:e,initialEpoch:s,samples:r,steps:o,batchSize:i,verbose:t,doValidation:a,metrics:l}),{callbackList:h,history:u}}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2940,7 +2940,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Nc(n,t={},e=!1){return gs(n,se.getMap().classNameMap,t,"layer",e)}/**
+ */function Nc(n,t={},e=!1){return bs(n,se.getMap().classNameMap,t,"layer",e)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2948,7 +2948,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Dc(n,t){return k(()=>{const e=N(.5,Pl(t)),s=uu(ds(t,e),n.dtype);return St(cn(n,s),-1)})}function Rc(n,t){return k(()=>uu(cn(er(n,-1),er(t,-1)),"float32"))}function Dw(n,t){return k(()=>ot(et(lr(cn(n,1),cn(t,1))),"float32"))}function Rw(n,t){return k(()=>ot(et(lr(cn(n,0),cn(t,1))),"float32"))}function Pw(n,t){return k(()=>{const e=Dw(n,t),s=Rw(n,t),r=O(e,s);return ot(hn(ds(r,0),X(e,r),0),"float32")})}function Lw(n,t){return Tr(n,t)}function Mw(n,t){return n.rank===t.rank&&(n=hr(n,[n.rank-1])),t=er(t,-1),t.dtype!==n.dtype&&(t=ot(t,n.dtype)),ot(cn(n,t),"float32")}const Ow=kr,Bw=kr,Fw=Ai,zw=Ai,Uw=Ei,Ww=Ei,Pc=Es,Gw=xc,Lc=_r,Dr={binaryAccuracy:Dc,categoricalAccuracy:Rc,precision:Pw,categoricalCrossentropy:Pc,sparseCategoricalCrossentropy:Lc,mse:Ow,MSE:Bw,mae:Fw,MAE:zw,mape:Uw,MAPE:Ww,cosine:Gw};function Vw(n){if(typeof n=="string"&&n in Dr)return Dr[n];if(typeof n!="string"&&n!=null)return n;throw new I(`Unknown metric ${n}`)}function Rr(n){if(Ae(n!==null,`Unknown LossOrMetricFn ${n}`),typeof n=="string")return n;{let t;for(const e of Object.keys(Nr))if(Nr[e]===n){t=e;break}if(t!==void 0)return t;for(const e of Object.keys(Dr))if(Dr[e]===n){t=e;break}return t!==void 0?t:n.name}}/**
+ */function Dc(n,t){return k(()=>{const e=N(.5,Pl(t)),s=uu(ps(t,e),n.dtype);return St(hn(n,s),-1)})}function Rc(n,t){return k(()=>uu(hn(nr(n,-1),nr(t,-1)),"float32"))}function Dw(n,t){return k(()=>ot(et(ur(hn(n,1),hn(t,1))),"float32"))}function Rw(n,t){return k(()=>ot(et(ur(hn(n,0),hn(t,1))),"float32"))}function Pw(n,t){return k(()=>{const e=Dw(n,t),s=Rw(n,t),r=O(e,s);return ot(fn(ps(r,0),X(e,r),0),"float32")})}function Lw(n,t){return Nr(n,t)}function Mw(n,t){return n.rank===t.rank&&(n=fr(n,[n.rank-1])),t=nr(t,-1),t.dtype!==n.dtype&&(t=ot(t,n.dtype)),ot(hn(n,t),"float32")}const Ow=_r,Bw=_r,Fw=Ai,zw=Ai,Uw=Ei,Ww=Ei,Pc=Cs,Gw=xc,Lc=Tr,Rr={binaryAccuracy:Dc,categoricalAccuracy:Rc,precision:Pw,categoricalCrossentropy:Pc,sparseCategoricalCrossentropy:Lc,mse:Ow,MSE:Bw,mae:Fw,MAE:zw,mape:Uw,MAPE:Ww,cosine:Gw};function Vw(n){if(typeof n=="string"&&n in Rr)return Rr[n];if(typeof n!="string"&&n!=null)return n;throw new I(`Unknown metric ${n}`)}function Pr(n){if(Ae(n!==null,`Unknown LossOrMetricFn ${n}`),typeof n=="string")return n;{let t;for(const e of Object.keys(Dr))if(Dr[e]===n){t=e;break}if(t!==void 0)return t;for(const e of Object.keys(Rr))if(Rr[e]===n){t=e;break}return t!==void 0?t:n.name}}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2972,7 +2972,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function jw(n,t,e,s=console.log){const r=Kw(n),o=["Layer (type)","Input Shape","Output shape","Param #"];r?(t=t||90,e=e||[.32,.61,.89,1]):(t=t||115,e=e||[.24,.48,.7,.8,1]),e[e.length-1]<=1&&(e=e.map(c=>Math.floor(t*c)));let i;if(!r){o.push("Receives inputs"),i=[];for(const c in n.nodesByDepth)i.push(...n.nodesByDepth[c])}s("_".repeat(t)),Pr(o,e,s),s("=".repeat(t));const a=n.layers;for(let c=0;c<a.length;++c)r?Yw(a[c],e,s):Xw(a[c],e,i,s),s((c===a.length-1?"=":"_").repeat(t));n.checkTrainableWeightsConsistency();const l=Hw(n),u=Ir(n.nonTrainableWeights);s(`Total params: ${l+u}`),s(`Trainable params: ${l}`),s(`Non-trainable params: ${u}`),s("_".repeat(t))}function Hw(n){let t;return n.collectedTrainableWeights!=null?t=Ir(n.collectedTrainableWeights):t=Ir(n.trainableWeights),t}function Kw(n){let t=!0;const e=[],s=[];for(const r in n.nodesByDepth)e.push(n.nodesByDepth[r]);for(const r of e){if(r.length>1||r.length===1&&r[0].inboundLayers.length>1){t=!1;break}s.push(...r)}if(t)for(const r of n.layers){let o=!1;for(const i of r.inboundNodes)if(s.indexOf(i)!==-1)if(o){t=!1;break}else o=!0;if(!t)break}return t}function Pr(n,t,e=console.log){let s="";for(let r=0;r<n.length;++r)r>0&&(s=s.slice(0,s.length-1)+" "),s+=n[r],s=s.slice(0,t[r]),s+=" ".repeat(t[r]-s.length);e(s)}function Yw(n,t,e){let s,r;try{r=n.inboundNodes.map(l=>JSON.stringify(l.inputShapes)).join(",")}catch{r="multiple"}try{s=JSON.stringify(n.outputShape)}catch{s="multiple"}const o=n.name,i=n.getClassName(),a=[`${o} (${i})`,r,s,n.countParams().toString()];Pr(a,t,e)}function Xw(n,t,e,s){let r,o;try{o=n.inboundNodes.map(h=>JSON.stringify(h.inputShapes)).join(",")}catch{o="multiple"}try{r=JSON.stringify(n.outputShape)}catch{r="multiple"}const i=[];for(const h of n.inboundNodes)if(!(e!=null&&e.length>0&&e.indexOf(h)===-1))for(let f=0;f<h.inboundLayers.length;++f){const d=h.inboundLayers[f].name,p=h.nodeIndices[f],g=h.tensorIndices[f];i.push(`${d}[${p}][${g}]`)}const a=n.name,l=n.getClassName(),u=i.length===0?"":i[0],c=[`${a} (${l})`,o,r,n.countParams().toString(),u];Pr(c,t,s);for(let h=1;h<i.length;++h)Pr(["","","","",i[h]],t,s)}/**
+ */function jw(n,t,e,s=console.log){const r=Kw(n),o=["Layer (type)","Input Shape","Output shape","Param #"];r?(t=t||90,e=e||[.32,.61,.89,1]):(t=t||115,e=e||[.24,.48,.7,.8,1]),e[e.length-1]<=1&&(e=e.map(c=>Math.floor(t*c)));let i;if(!r){o.push("Receives inputs"),i=[];for(const c in n.nodesByDepth)i.push(...n.nodesByDepth[c])}s("_".repeat(t)),Lr(o,e,s),s("=".repeat(t));const a=n.layers;for(let c=0;c<a.length;++c)r?Yw(a[c],e,s):Xw(a[c],e,i,s),s((c===a.length-1?"=":"_").repeat(t));n.checkTrainableWeightsConsistency();const l=Hw(n),u=Ar(n.nonTrainableWeights);s(`Total params: ${l+u}`),s(`Trainable params: ${l}`),s(`Non-trainable params: ${u}`),s("_".repeat(t))}function Hw(n){let t;return n.collectedTrainableWeights!=null?t=Ar(n.collectedTrainableWeights):t=Ar(n.trainableWeights),t}function Kw(n){let t=!0;const e=[],s=[];for(const r in n.nodesByDepth)e.push(n.nodesByDepth[r]);for(const r of e){if(r.length>1||r.length===1&&r[0].inboundLayers.length>1){t=!1;break}s.push(...r)}if(t)for(const r of n.layers){let o=!1;for(const i of r.inboundNodes)if(s.indexOf(i)!==-1)if(o){t=!1;break}else o=!0;if(!t)break}return t}function Lr(n,t,e=console.log){let s="";for(let r=0;r<n.length;++r)r>0&&(s=s.slice(0,s.length-1)+" "),s+=n[r],s=s.slice(0,t[r]),s+=" ".repeat(t[r]-s.length);e(s)}function Yw(n,t,e){let s,r;try{r=n.inboundNodes.map(l=>JSON.stringify(l.inputShapes)).join(",")}catch{r="multiple"}try{s=JSON.stringify(n.outputShape)}catch{s="multiple"}const o=n.name,i=n.getClassName(),a=[`${o} (${i})`,r,s,n.countParams().toString()];Lr(a,t,e)}function Xw(n,t,e,s){let r,o;try{o=n.inboundNodes.map(h=>JSON.stringify(h.inputShapes)).join(",")}catch{o="multiple"}try{r=JSON.stringify(n.outputShape)}catch{r="multiple"}const i=[];for(const h of n.inboundNodes)if(!(e!=null&&e.length>0&&e.indexOf(h)===-1))for(let f=0;f<h.inboundLayers.length;++f){const d=h.inboundLayers[f].name,p=h.nodeIndices[f],g=h.tensorIndices[f];i.push(`${d}[${p}][${g}]`)}const a=n.name,l=n.getClassName(),u=i.length===0?"":i[0],c=[`${a} (${l})`,o,r,n.countParams().toString(),u];Lr(c,t,s);for(let h=1;h<i.length;++h)Lr(["","","","",i[h]],t,s)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -2980,7 +2980,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Bc(n,t,e){return(n==="inboundNodes"||n==="outputLayers"||n==="inputLayers")&&t===0&&typeof e=="string"}function Ti(n,t){if(n===null)return null;if(typeof n=="string")return pn(n);if(typeof n=="number"||typeof n=="boolean")return n;if(n instanceof Array){const e=[],s=n.length;for(let r=0;r<s;++r){const o=n[r];Bc(t,r,o)?e.push(o):e.push(Ti(o,t))}return e}else{const e={};for(const s of Object.keys(n)){const r=n[s];if(s==="name"&&typeof r=="string")e[s]=r;else{const o=pn(s);e[o]=Ti(r,o)}}return e}}function Ni(n,t){if(n==null)return null;if(typeof n=="string")return Le(n);if(typeof n=="number"||typeof n=="boolean")return n;if(n instanceof Array){const e=[],s=n.length;for(let r=0;r<s;++r){const o=n[r];Bc(t,r,o)?e.push(o):e.push(Ni(o,t))}return e}else{const e={};for(const s of Object.keys(n)){const r=n[s],o=Le(s);(s==="name"||s==="className")&&typeof r=="string"?e[o]=r:e[o]=Ni(r,s)}return e}}/** @license See the LICENSE file. */const Fc="4.20.0";/**
+ */function Bc(n,t,e){return(n==="inboundNodes"||n==="outputLayers"||n==="inputLayers")&&t===0&&typeof e=="string"}function Ti(n,t){if(n===null)return null;if(typeof n=="string")return mn(n);if(typeof n=="number"||typeof n=="boolean")return n;if(n instanceof Array){const e=[],s=n.length;for(let r=0;r<s;++r){const o=n[r];Bc(t,r,o)?e.push(o):e.push(Ti(o,t))}return e}else{const e={};for(const s of Object.keys(n)){const r=n[s];if(s==="name"&&typeof r=="string")e[s]=r;else{const o=mn(s);e[o]=Ti(r,o)}}return e}}function Ni(n,t){if(n==null)return null;if(typeof n=="string")return Le(n);if(typeof n=="number"||typeof n=="boolean")return n;if(n instanceof Array){const e=[],s=n.length;for(let r=0;r<s;++r){const o=n[r];Bc(t,r,o)?e.push(o):e.push(Ni(o,t))}return e}else{const e={};for(const s of Object.keys(n)){const r=n[s],o=Le(s);(s==="name"||s==="className")&&typeof r=="string"?e[o]=r:e[o]=Ni(r,s)}return e}}/** @license See the LICENSE file. */const Fc="4.20.0";/**
  * @license
  * Copyright 2022 Google LLC
  *
@@ -2996,7 +2996,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */class _s extends pe{constructor(t){if(super({dtype:t.dtype,name:t.name!=null?t.name:di("input").toString()}),t.batchSize==null&&(t.batchSize=null),t.sparse==null&&(t.sparse=!1),this.trainable=!1,this.built=!0,this.sparse=t.sparse,t.inputShape!=null&&t.batchInputShape!=null)throw new I("Only provide the inputShape OR batchInputShape argument to inputLayer, not both at the same time.");let e=t.batchInputShape;if(e==null){if(t.inputShape==null)throw new I("An InputLayer should be passed either a `batchInputShape` or an `inputShape`.");e=[t.batchSize].concat(t.inputShape)}else if(t.batchSize!=null)throw new I("Cannot specify batchSize if batchInputShape is specified when creating an InputLayer.");const s=t.dtype||"float32";this.batchInputShape=e,this.dtype=s,this.inputSpec=[{shape:e}];const r=new bn(this.dtype,this.batchInputShape,this,[],{},this.name);r.nodeIndex=0,r.tensorIndex=0,new xi({outboundLayer:this,inboundLayers:[],nodeIndices:[],tensorIndices:[],inputTensors:[r],outputTensors:[r],inputMasks:[null],outputMasks:[null],inputShapes:[e],outputShapes:[e]})}apply(t,e){throw new I(`Cannot pass any input to an InputLayer's apply() method. InputLayer name: ${this.name}`)}dispose(){return{refCountAfterDispose:this._refCount,numDisposedVariables:0}}getConfig(){return{batchInputShape:this.batchInputShape,dtype:this.dtype,sparse:this.sparse,name:this.name}}}_s.className="InputLayer",M(_s);function Jw(n){if(n.batchShape==null&&n.shape==null)throw new Error("Please provide to Input either a `shape` or a `batchShape` argument. Note that `shape` does not include the batch dimension.");if(n.batchShape!=null&&n.shape!=null)throw new I("Please provide either a `shape` or `batchShape` argument to Input, but not both.");let t=n.batchShape;n.shape!=null&&t==null&&(t=[null].concat(n.shape));let e=n.dtype;return e==null&&(e="float32"),new _s({batchInputShape:t,name:n.name,dtype:e,sparse:n.sparse}).inboundNodes[0].outputTensors[0]}/**
+ */class Ts extends pe{constructor(t){if(super({dtype:t.dtype,name:t.name!=null?t.name:di("input").toString()}),t.batchSize==null&&(t.batchSize=null),t.sparse==null&&(t.sparse=!1),this.trainable=!1,this.built=!0,this.sparse=t.sparse,t.inputShape!=null&&t.batchInputShape!=null)throw new I("Only provide the inputShape OR batchInputShape argument to inputLayer, not both at the same time.");let e=t.batchInputShape;if(e==null){if(t.inputShape==null)throw new I("An InputLayer should be passed either a `batchInputShape` or an `inputShape`.");e=[t.batchSize].concat(t.inputShape)}else if(t.batchSize!=null)throw new I("Cannot specify batchSize if batchInputShape is specified when creating an InputLayer.");const s=t.dtype||"float32";this.batchInputShape=e,this.dtype=s,this.inputSpec=[{shape:e}];const r=new yn(this.dtype,this.batchInputShape,this,[],{},this.name);r.nodeIndex=0,r.tensorIndex=0,new xi({outboundLayer:this,inboundLayers:[],nodeIndices:[],tensorIndices:[],inputTensors:[r],outputTensors:[r],inputMasks:[null],outputMasks:[null],inputShapes:[e],outputShapes:[e]})}apply(t,e){throw new I(`Cannot pass any input to an InputLayer's apply() method. InputLayer name: ${this.name}`)}dispose(){return{refCountAfterDispose:this._refCount,numDisposedVariables:0}}getConfig(){return{batchInputShape:this.batchInputShape,dtype:this.dtype,sparse:this.sparse,name:this.name}}}Ts.className="InputLayer",M(Ts);function Jw(n){if(n.batchShape==null&&n.shape==null)throw new Error("Please provide to Input either a `shape` or a `batchShape` argument. Note that `shape` does not include the batch dimension.");if(n.batchShape!=null&&n.shape!=null)throw new I("Please provide either a `shape` or `batchShape` argument to Input, but not both.");let t=n.batchShape;n.shape!=null&&t==null&&(t=[null].concat(n.shape));let e=n.dtype;return e==null&&(e="float32"),new Ts({batchInputShape:t,name:n.name,dtype:e,sparse:n.sparse}).inboundNodes[0].outputTensors[0]}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -3004,7 +3004,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Zw(n,t){if(n.dtype==null||n.dtype===t.dtype)return t;try{return ot(t,n.dtype)}catch{throw new I(`The dtype of the feed (${t.dtype}) can not be cast to the dtype of the key '${n.name}' (${n.dtype}).`)}}class Ke{constructor(t){if(this.id2Value={},this.id2Mask={},this.name2Id={},t instanceof Ke)for(const e in t.id2Value)this.id2Value[e]=t.id2Value[e],e in t.id2Mask&&(this.id2Mask[e]=t.id2Mask[e]);else{if(t==null)return;for(const e of t)this.add(e.key,e.value)}}add(t,e,s){if(this.id2Value[t.id]==null)this.id2Value[t.id]=Zw(t,e),this.name2Id[t.name]=t.id,s!=null&&(this.id2Mask[t.id]=s);else throw new I(`Duplicate key: name=${t.name}, id=${t.id}`);return this}addFeed(t){this.add(t.key,t.value)}hasKey(t){return this.id2Value[t.id]!=null}names(){return Object.keys(this.name2Id)}getValue(t){if(t instanceof bn){if(this.id2Value[t.id]==null)throw new I(`Nonexistent key: ${t.name}`);return this.id2Value[t.id]}else{const e=this.name2Id[t];if(e==null)throw new I(`Feed dict has no SymbolicTensor name: ${t}`);return this.id2Value[e]}}getMask(t){if(t instanceof bn){if(this.id2Value[t.id]==null)throw new I(`Nonexistent key: ${t.name}`);return this.id2Mask[t.id]}else{const e=this.name2Id[t];if(e==null)throw new I(`Feed dict has no SymbolicTensor name: ${t}`);return this.id2Mask[e]}}disposeMasks(){this.id2Mask!=null&&ut(this.id2Mask)}}const Uc=new zc,Wc=new zc;function Ts(n,t,e,s){const r=e==null?!1:e.training,o=Array.isArray(n),i=o?n:[n],a=i.map(p=>p.name),l=[],u=t.names();for(const p of a)u.indexOf(p)!==-1?l.push(t.getValue(p)):l.push(null);const c=a.join(",")+"|"+t.names().sort().join(",");let h=Uc.get(c),f;if(h==null){const p=Qw(i,t);h=p.sorted,f=p.recipientCounts,Uc.put(c,h),Wc.put(c,f)}f={},r||Object.assign(f,Wc.get(c));const d=new Ke(t);for(let p=0;p<h.length;++p){const g=h[p],m=g.sourceLayer;if(m instanceof _s)continue;const b=[],y=[],S=[];let x=!1;for(const T of g.inputs){const P=d.getValue(T),B=d.getMask(T);b.push(P),y.push(B),B!=null&&(x=!0),r||(f[T.name]--,f[T.name]===0&&!t.hasKey(T)&&a.indexOf(T.name)===-1&&!P.isDisposed&&T.sourceLayer.stateful!==!0&&S.push(P))}x&&(e=e||{},e.mask=y[0]);const $=nt(m.apply(b,e));let E=null;m.supportsMasking&&(E=m.computeMask(b,y));const D=e1(g),_=Array.isArray(D)?D:[D];for(let T=0;T<_.length;++T){d.hasKey(_[T])||d.add(_[T],$[T],Array.isArray(E)?E[0]:E);const P=a.indexOf(_[T].name);P!==-1&&(l[P]=$[T])}r||ut(S)}return d.disposeMasks(),o?l:l[0]}function Qw(n,t){w(n!=null&&n.length>0,()=>"Expected at least one fetch, got none");let e=[],s={};if(n.length===1){const r=Gc(n[0],t);e=r.sorted,s=r.recipientMap}else{const r=new Set;for(const o of n){const{sorted:i,recipientMap:a}=Gc(o,t);for(const l of i)r.has(l.name)||(e.push(l),r.add(l.name));for(const l in a)s[l]==null&&(s[l]=new Set),a[l].forEach(u=>s[l].add(u))}}return{sorted:e,recipientCounts:t1(s)}}function t1(n){const t={};for(const e in n)t[e]=n[e].size;return t}function Gc(n,t){const e=new Set,s=[],r={};for(const a of t.names())e.add(a);const o=[],i=[];for(o.push(n);o.length>0;){const a=o[o.length-1];if(e.has(a.name)){o.pop();continue}const l=i[i.length-1]===o.length-1;if(a.inputs.length===0||l)o.pop(),s.push(a),e.add(a.name),l&&i.pop();else{i.push(o.length-1);for(const u of a.inputs)r[u.name]==null&&(r[u.name]=new Set),r[u.name].add(a.name),!e.has(u.name)&&o.push(u)}}return{sorted:s,recipientMap:r}}function e1(n){let t;if(n.sourceLayer.inboundNodes.length===1)t=n.sourceLayer.output;else{let e=null;for(let s=0;s<n.sourceLayer.inboundNodes.length;++s)for(const r of n.sourceLayer.inboundNodes[s].outputTensors)if(r.id===n.id){e=s;break}t=n.sourceLayer.getOutputAt(e)}return t}/**
+ */function Zw(n,t){if(n.dtype==null||n.dtype===t.dtype)return t;try{return ot(t,n.dtype)}catch{throw new I(`The dtype of the feed (${t.dtype}) can not be cast to the dtype of the key '${n.name}' (${n.dtype}).`)}}class Ke{constructor(t){if(this.id2Value={},this.id2Mask={},this.name2Id={},t instanceof Ke)for(const e in t.id2Value)this.id2Value[e]=t.id2Value[e],e in t.id2Mask&&(this.id2Mask[e]=t.id2Mask[e]);else{if(t==null)return;for(const e of t)this.add(e.key,e.value)}}add(t,e,s){if(this.id2Value[t.id]==null)this.id2Value[t.id]=Zw(t,e),this.name2Id[t.name]=t.id,s!=null&&(this.id2Mask[t.id]=s);else throw new I(`Duplicate key: name=${t.name}, id=${t.id}`);return this}addFeed(t){this.add(t.key,t.value)}hasKey(t){return this.id2Value[t.id]!=null}names(){return Object.keys(this.name2Id)}getValue(t){if(t instanceof yn){if(this.id2Value[t.id]==null)throw new I(`Nonexistent key: ${t.name}`);return this.id2Value[t.id]}else{const e=this.name2Id[t];if(e==null)throw new I(`Feed dict has no SymbolicTensor name: ${t}`);return this.id2Value[e]}}getMask(t){if(t instanceof yn){if(this.id2Value[t.id]==null)throw new I(`Nonexistent key: ${t.name}`);return this.id2Mask[t.id]}else{const e=this.name2Id[t];if(e==null)throw new I(`Feed dict has no SymbolicTensor name: ${t}`);return this.id2Mask[e]}}disposeMasks(){this.id2Mask!=null&&ut(this.id2Mask)}}const Uc=new zc,Wc=new zc;function Ns(n,t,e,s){const r=e==null?!1:e.training,o=Array.isArray(n),i=o?n:[n],a=i.map(p=>p.name),l=[],u=t.names();for(const p of a)u.indexOf(p)!==-1?l.push(t.getValue(p)):l.push(null);const c=a.join(",")+"|"+t.names().sort().join(",");let h=Uc.get(c),f;if(h==null){const p=Qw(i,t);h=p.sorted,f=p.recipientCounts,Uc.put(c,h),Wc.put(c,f)}f={},r||Object.assign(f,Wc.get(c));const d=new Ke(t);for(let p=0;p<h.length;++p){const g=h[p],m=g.sourceLayer;if(m instanceof Ts)continue;const b=[],y=[],S=[];let x=!1;for(const T of g.inputs){const P=d.getValue(T),B=d.getMask(T);b.push(P),y.push(B),B!=null&&(x=!0),r||(f[T.name]--,f[T.name]===0&&!t.hasKey(T)&&a.indexOf(T.name)===-1&&!P.isDisposed&&T.sourceLayer.stateful!==!0&&S.push(P))}x&&(e=e||{},e.mask=y[0]);const $=nt(m.apply(b,e));let E=null;m.supportsMasking&&(E=m.computeMask(b,y));const D=e1(g),_=Array.isArray(D)?D:[D];for(let T=0;T<_.length;++T){d.hasKey(_[T])||d.add(_[T],$[T],Array.isArray(E)?E[0]:E);const P=a.indexOf(_[T].name);P!==-1&&(l[P]=$[T])}r||ut(S)}return d.disposeMasks(),o?l:l[0]}function Qw(n,t){w(n!=null&&n.length>0,()=>"Expected at least one fetch, got none");let e=[],s={};if(n.length===1){const r=Gc(n[0],t);e=r.sorted,s=r.recipientMap}else{const r=new Set;for(const o of n){const{sorted:i,recipientMap:a}=Gc(o,t);for(const l of i)r.has(l.name)||(e.push(l),r.add(l.name));for(const l in a)s[l]==null&&(s[l]=new Set),a[l].forEach(u=>s[l].add(u))}}return{sorted:e,recipientCounts:t1(s)}}function t1(n){const t={};for(const e in n)t[e]=n[e].size;return t}function Gc(n,t){const e=new Set,s=[],r={};for(const a of t.names())e.add(a);const o=[],i=[];for(o.push(n);o.length>0;){const a=o[o.length-1];if(e.has(a.name)){o.pop();continue}const l=i[i.length-1]===o.length-1;if(a.inputs.length===0||l)o.pop(),s.push(a),e.add(a.name),l&&i.pop();else{i.push(o.length-1);for(const u of a.inputs)r[u.name]==null&&(r[u.name]=new Set),r[u.name].add(a.name),!e.has(u.name)&&o.push(u)}}return{sorted:s,recipientMap:r}}function e1(n){let t;if(n.sourceLayer.inboundNodes.length===1)t=n.sourceLayer.output;else{let e=null;for(let s=0;s<n.sourceLayer.inboundNodes.length;++s)for(const r of n.sourceLayer.inboundNodes[s].outputTensors)if(r.id===n.id){e=s;break}t=n.sourceLayer.getOutputAt(e)}return t}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -3012,7 +3012,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */const n1=n=>{const t=Object.keys(n);if(t.length===0)return!1;const e=t[0].split("/");return!isNaN(parseInt(e[e.length-1],10))};class me extends pe{constructor(t){if(super({}),this.containerNodes=new Set,this.name=t.name,this.name==null){const y=this.getClassName().toLowerCase();this.name=di(y)}if(this.supportsMasking=!1,this.trainable_=!0,Array.isArray(t.inputs)?this.inputs=t.inputs.slice():this.inputs=[t.inputs],Array.isArray(t.outputs)?this.outputs=t.outputs.slice():this.outputs=[t.outputs],mn(this.inputs).length!==this.inputs.length)throw new I(`The list of inputs passed to the model is redundant. All inputs should only appear once. Found: ${this.inputs.map(y=>y.name)}`);mn(this.outputs).length!==this.outputs.length&&console.warn(`The list of outputs passed to the model is redundant. All outputs should only appear once. Found: ${this.outputs.map(y=>y.name)}`),this.inputLayers=[],this.inputLayersNodeIndices=[],this.inputLayersTensorIndices=[],this.outputLayers=[],this.outputLayersNodeIndices=[],this.outputLayersTensorIndices=[],this.layers=[],this.internalContainerRefs=[];for(const y of this.outputs){const S=y.sourceLayer,x=y.nodeIndex,$=y.tensorIndex;this.outputLayers.push(S),this.outputLayersNodeIndices.push(x),this.outputLayersTensorIndices.push($)}for(const y of this.inputs){const S=y.sourceLayer,x=y.nodeIndex,$=y.tensorIndex;Ae(x===0,"input layer has >1 nodes"),Ae($===0,"input layer has >1 tensors"),this.inputLayers.push(S),this.inputLayersNodeIndices.push(x),this.inputLayersTensorIndices.push($)}this.inputNames=[],this.outputNames=[],this.feedInputShapes=[],this.feedInputNames=[],this.feedOutputNames=[];for(let y=0;y<this.inputLayers.length;y++){const S=this.inputLayers[y];if(!(S instanceof _s))throw new TypeError(`Input layers to a LayersModel must be InputLayer objects. Received inputs: ${t.inputs}. Input ${y} (0-based) originates from layer type ${S.getClassName()}.`);this.inputNames.push(S.name),this.feedInputShapes.push(S.batchInputShape),this.feedInputNames.push(S.name)}for(const y of this.outputLayers)this.outputNames.push(y.name);this.internalInputShapes=this.inputs.map(y=>y.shape),this.internalOutputShapes=this.outputs.map(y=>y.shape);const e={},s={},r={},o={},i={},a=[],l=(y,S,x,$,E,D)=>{($==null||E==null||D==null)&&($=y.sourceLayer,E=y.nodeIndex,D=y.tensorIndex);const _=$.inboundNodes[E];if(x.indexOf(_)!==-1)throw new He(`The tensor ${y.name} at layer "${$.name}" is part of a cycle.`);if(S.indexOf(_)!==-1)return;this.containerNodes.add(me.nodeKey($,E)),$.id in i||(i[$.id]=Object.keys(i).length),x.indexOf(_)===-1&&x.push(_);const T=_.inboundLayers.length;for(let P=0;P<T;P++){const B=_.inputTensors[P],Y=_.inboundLayers[P],j=_.nodeIndices[P],F=_.tensorIndices[P];l(B,S,x,Y,j,F)}for(S.push(_);x.indexOf(_)>=0;)x.splice(x.indexOf(_),1);a.push(_)},u=[],c=[];for(const y of this.outputs)l(y,u,c);const h=a.slice().reverse();for(const y of h){s[y.id]=y,y.id in e||(e[y.id]=0);let S=e[y.id];const x=r[y.outboundLayer.id]==null?0:r[y.outboundLayer.id];S=Math.max(S,x),r[y.outboundLayer.id]=S,o[y.outboundLayer.id]=y.outboundLayer,e[y.id]=S;for(let $=0;$<y.inboundLayers.length;$++){const E=y.inboundLayers[$],D=y.nodeIndices[$],_=E.inboundNodes[D],T=e[_.id]==null?0:e[_.id];e[_.id]=Math.max(S+1,T),s[_.id]=_}}const f={};for(const y in e){const S=e[y];S in f||(f[S]=[]),f[S].push(s[y])}const d={};for(const y in r){const S=r[y];S in d||(d[S]=[]),d[S].push(o[y])}let p=Object.keys(d).map(y=>parseInt(y,10)).sort(mr);this.layers=[];for(const y of p){const S=d[y];S.sort((x,$)=>{const E=i[x.id],D=i[$.id];return E<D?-1:E>D?1:0});for(const x of S)x instanceof me&&this.internalContainerRefs.push(x),this.layers.push(x)}this.layersByDepth=d,p=Object.keys(f).map(y=>parseInt(y,10)).sort(mr);const g=this.inputs.slice(),m=[];for(const y of p)for(const S of f[y]){const x=S.outboundLayer;if(x!=null){for(const $ of S.inputTensors)if(g.indexOf($)===-1)throw new He(`Graph disconnected: cannot obtain value for tensor ${$} at layer "${x.name}". The following previous layers were accessed without issue: ${m}`);for(const $ of S.outputTensors)g.push($);m.push(x.name)}}this.nodesByDepth=f;const b=this.layers.map(y=>y.name);for(const y of b){const S=b.filter(x=>x===y).length;if(S!==1)throw new He(`The name "${y}" is used ${S} times in the model. All layer names should be unique. Layer names: `+JSON.stringify(b))}this.outboundNodes=[],this.inboundNodes=[],new xi({outboundLayer:this,inboundLayers:[],nodeIndices:[],tensorIndices:[],inputTensors:this.inputs,outputTensors:this.outputs,inputMasks:this.inputs.map(y=>null),outputMasks:this.outputs.map(y=>null),inputShapes:this.inputs.map(y=>y.shape),outputShapes:this.outputs.map(y=>y.shape)}),this.built=!0,this._refCount=1}assertNotDisposed(){if(this._refCount===0)throw new Error(`Container '${this.name}' is already disposed.`)}dispose(){this.assertNotDisposed();const t={refCountAfterDispose:null,numDisposedVariables:0};if(--this._refCount===0){for(const e of this.layers)t.numDisposedVariables+=e.dispose().numDisposedVariables;for(const e of this.internalContainerRefs)t.numDisposedVariables+=e.dispose().numDisposedVariables}return t.refCountAfterDispose=this._refCount,t}get trainable(){return this.trainable_}set trainable(t){this.layers.forEach(e=>{e._trainableWeights.forEach(s=>s.trainable=t)}),this.trainable_=t}get trainableWeights(){if(this._trainableWeights.length>0)throw new I("Container instance unexpectedly contains _trainableWeights.The trainable weights of a Container are a union of the trainable weights of its consituent Layers. Its own _trainableWeights must remain an empty Array.");if(!this.trainable)return[];let t=[];for(const e of this.layers)t=t.concat(e.trainableWeights);return t}get nonTrainableWeights(){const t=[];for(const e of this.layers)t.push(...e.nonTrainableWeights);if(!this.trainable){const e=[];for(const s of this.layers)e.push(...s.trainableWeights);return e.concat(t)}return t}get weights(){return this.trainableWeights.concat(this.nonTrainableWeights)}loadWeights(t,e=!0){const s={};let r=0;const o=n1(t);o&&this.parseWeights(t);for(const a of this.layers)for(const[l,u]of a.weights.entries()){const c=o?`${u.name.split("/").slice(0,-1).join("/")+"/"}${l}`:u.originalName;if(s[c]!=null)throw new I(`Duplicate weight name: ${c}`);s[c]=u,r++}const i=[];for(const a in t){let l=a;if(s[a]==null){const u=a.split("/");l=u.slice(0,-2).concat([u[u.length-1]]).join("/")}if(s[l]!=null)i.push([s[l],t[a]]);else if(e)throw new I(`Provided weight data has no target variable: ${a}`);delete s[l]}if(e){const a=[];for(const l in s)a.push(l);if(a.length>0)throw new I(`${a.length} of ${r} weights are not set: ${a}`)}ju(i)}parseWeights(t){for(const e in Object.keys(t)){const s=e.split("/"),r=["vars","layer_checkpoint_dependencies"],o=s.map(i=>i.startsWith("_")?i.slice(1):i).filter(i=>!r.includes(i)).join("/");o!==e&&(t[o]=t[e],delete t[e])}}updatedConfig(){const t=this.getConfig(),e={};return e.className=this.getClassName(),e.config=t,e.kerasVersion=`tfjs-layers ${Fc}`,e.backend="TensorFlow.js",e}toJSON(t,e=!0){const s=Ni(this.updatedConfig());return e?JSON.stringify(s):s}call(t,e){return k(()=>{t=nt(t);const s=new Ke;for(let r=0;r<this.inputs.length;++r)s.add(this.inputs[r],t[r]);return Ts(this.outputs,s,e)})}computeMask(t,e){return k(()=>{t=nt(t);let s;return e==null?s=pr(null,t.length):s=nt(e),this.runInternalGraph(t,s)[1]})}computeOutputShape(t){const e=vr(t);if(e.length!==this.inputLayers.length)throw new I(`Invalid inputShape argument ${t}: model has ${this.inputLayers.length} tensor inputs.`);const s={};for(let a=0;a<e.length;a++){const l=this.inputLayers[a],u=e[a],c=l.name+"_0_0";s[c]=u}const r=Object.keys(this.nodesByDepth).map(a=>parseInt(a,10)).sort(mr);if(r.length>1)for(const a of r){const l=this.nodesByDepth[a];for(const u of l){const c=u.outboundLayer;if(this.inputLayers.map(g=>g.id).indexOf(c.id)!==-1)continue;const h=[];for(let g=0;g<u.inboundLayers.length;g++){const m=u.inboundLayers[g],b=u.nodeIndices[g],y=u.tensorIndices[g],S=`${m.name}_${b}_${y}`,x=s[S];h.push(x)}const f=c.computeOutputShape(Ut(h)),d=vr(f),p=c.inboundNodes.indexOf(u);for(let g=0;g<d.length;g++){const m=`${c.name}_${p}_${g}`;s[m]=d[g]}}}const o=[],i=[];for(let a=0;a<this.outputLayers.length;a++){const l=this.outputLayers[a],u=this.outputLayersNodeIndices[a],c=this.outputLayersTensorIndices[a],h=`${l.name}_${u}_${c}`;i.push(h)}for(let a=0;a<i.length;a++){const l=i[a];Ae(l in s),o.push(s[l])}return Ut(o)}runInternalGraph(t,e){e==null&&(e=pr(null,t.length));const s={};for(let l=0;l<this.inputs.length;++l){const u=this.inputs[l],c=t[l],h=e[l];s[u.id]=[c,h]}const r=Object.keys(this.nodesByDepth).map(l=>parseInt(l,10)).sort(mr);for(const l of r){const u=this.nodesByDepth[l];for(const c of u){const h=c.outboundLayer,f=c.inputTensors,d=c.outputTensors,p=new Array;for(const g of f)g.id in s&&p.push(s[g.id]);if(p.length===f.length){let g={},m,b,y,S;if(c.callArgs!=null&&(g=c.callArgs),p.length===1){const[x,$]=p[0];g.mask==null&&(g.mask=$),y=nt(h.call(x,g)),S=nt(h.computeMask(x,$)),m=[x],b=[$]}else m=p.map(x=>x[0]),b=p.map(x=>x[1]),g.mask==null&&(g.mask=b),y=nt(h.call(m,g)),S=nt(h.computeMask(m,b));if(h.activityRegularizer)throw new Z("LayersModel invocation with concrete Tensor value(s) in the presence of activity regularizer(s) is not supported yet.");for(let x=0;x<d.length;++x){const $=d[x],E=y[x],D=S[x];s[$.id]=[E,D]}}}}const o=[],i=[],a=[];for(const l of this.outputs){Ae(l.id in s,`Could not compute output ${l.name} : ${l.id}`);const[u,c]=s[l.id];a.push(u.shape),o.push(u),i.push(c)}return[o,i,a]}buildNodeConversionMap(t){const e={};let s;for(const r of this.layers){s=r instanceof me?1:0;for(let o=0;o<r.inboundNodes.length;o++){const i=me.nodeKey(r,o);this.containerNodes.has(i)&&(e[i]=s,s+=1)}}return e}getLayer(t,e){if(e!=null)return this.findLayer(e);if(t==null)throw new I("Provide either a layer name or layer index");if(typeof t=="number")return this.findLayer(t);for(const s of this.layers)if(s.name===t)return s;throw new I(`No such layer: ${t}`)}findLayer(t){if(this.layers.length<=t)throw new I(`Was asked to retrieve layer at index ${t}, but model only has ${this.layers.length} layer(s).`);return this.layers[t]}calculateLosses(){return k(()=>{const t=[];for(const e of this.layers)for(let s=0;s<e.inboundNodes.length;++s){const r=me.nodeKey(e,s);this.containerNodes.has(r)&&t.push(...e.calculateLosses())}return t})}getConfig(){const t={name:this.name},e=this.buildNodeConversionMap(this.layers),s=[];for(const i of this.layers){const a=i.getClassName(),l=i.getConfig(),u=[];for(let h=0;h<i.inboundNodes.length;h++){const f=i.inboundNodes[h],d=me.nodeKey(i,h);let p={};if(this.containerNodes.has(d)){if(f.callArgs)try{JSON.stringify(f.callArgs),p=f.callArgs}catch{console.warn(`Layer ${i.name} was passed non-serializable keyword arguments: ${f.callArgs}. They will not be included in the serialized model (and thus will be missing at deserialization time).`),p={}}if(f.inboundLayers.length>0){const g=[];for(let m=0;m<f.inboundLayers.length;m++){const b=f.inboundLayers[m],y=f.nodeIndices[m],S=f.tensorIndices[m],x=me.nodeKey(b,y);let $=e[x];$==null&&($=0),g.push([b.name,$,S,p])}u.push(g)}}}const c={};c.name=i.name,c.className=a,c.config=l,c.inboundNodes=u,s.push(c)}t.layers=s;const r=[];for(let i=0;i<this.inputLayers.length;i++){const a=this.inputLayers[i],l=this.inputLayersNodeIndices[i],u=me.nodeKey(a,l);if(!this.containerNodes.has(u))continue;let c=e[u];c==null&&(c=0);const h=this.inputLayersTensorIndices[i];r.push([a.name,c,h])}t.inputLayers=r;const o=[];for(let i=0;i<this.outputLayers.length;i++){const a=this.outputLayers[i],l=this.outputLayersNodeIndices[i],u=me.nodeKey(a,l);if(!this.containerNodes.has(u))continue;let c=e[u];c==null&&(c=0);const h=this.outputLayersTensorIndices[i];o.push([a.name,c,h])}return t.outputLayers=o,t}static fromConfig(t,e,s={},r=!1){const o={},i={};function a(m,b){m.name in i?i[m.name].push(b):i[m.name]=[b]}function l(m,b){const y=[];let S;for(const x of b){const $=x[0],E=x[1],D=x[2];if(S=x[3]==null?{}:x[3],!($ in o)){a(m,b);return}const _=o[$];if(_.inboundNodes.length<=E){a(m,b);return}const T=_.inboundNodes[E];y.push(T.outputTensors[D])}y.length>0&&m.apply(Ut(y),S)}function u(m){const b=m.name,y=Nc(m,e.customObjects!=null?e.customObjects:{});y.setFastWeightInitDuringBuild(r),o[b]=y,m.inboundNodes.forEach(x=>{if(!(x instanceof Array))throw new I(`Corrupted configuration, expected array for nodeData: ${x}`);a(y,x)})}const c=e.name,h=e.layers;for(const m of h)u(m);for(;!Fy(i);)for(const m of h){const b=o[m.name];if(b.name in i){const y=i[b.name];delete i[b.name];for(const S of y)l(b,S)}}const f=[],d=[],p=e.inputLayers;for(const m of p){const b=m[0],y=m[1],S=m[2];Ae(b in o);const $=o[b].inboundNodes[y].outputTensors;f.push($[S])}const g=e.outputLayers;for(const m of g){const b=m[0],y=m[1],S=m[2];Ae(b in o);const $=o[b].inboundNodes[y].outputTensors;d.push($[S])}return new t({inputs:f,outputs:d,name:c})}get stateful(){if(this._stateful)throw new I("Container instance unexpectedly has _stateful = true. The statefulness of a Container is determined by the Layers it contains. Its _stateful property must remain the default false.");for(const t of this.layers)if(t.stateful)return!0;return!1}resetStates(){k(()=>{this.layers.forEach(t=>{t.stateful&&t.resetStates()})})}}/**
+ */const n1=n=>{const t=Object.keys(n);if(t.length===0)return!1;const e=t[0].split("/");return!isNaN(parseInt(e[e.length-1],10))};class me extends pe{constructor(t){if(super({}),this.containerNodes=new Set,this.name=t.name,this.name==null){const y=this.getClassName().toLowerCase();this.name=di(y)}if(this.supportsMasking=!1,this.trainable_=!0,Array.isArray(t.inputs)?this.inputs=t.inputs.slice():this.inputs=[t.inputs],Array.isArray(t.outputs)?this.outputs=t.outputs.slice():this.outputs=[t.outputs],gn(this.inputs).length!==this.inputs.length)throw new I(`The list of inputs passed to the model is redundant. All inputs should only appear once. Found: ${this.inputs.map(y=>y.name)}`);gn(this.outputs).length!==this.outputs.length&&console.warn(`The list of outputs passed to the model is redundant. All outputs should only appear once. Found: ${this.outputs.map(y=>y.name)}`),this.inputLayers=[],this.inputLayersNodeIndices=[],this.inputLayersTensorIndices=[],this.outputLayers=[],this.outputLayersNodeIndices=[],this.outputLayersTensorIndices=[],this.layers=[],this.internalContainerRefs=[];for(const y of this.outputs){const S=y.sourceLayer,x=y.nodeIndex,$=y.tensorIndex;this.outputLayers.push(S),this.outputLayersNodeIndices.push(x),this.outputLayersTensorIndices.push($)}for(const y of this.inputs){const S=y.sourceLayer,x=y.nodeIndex,$=y.tensorIndex;Ae(x===0,"input layer has >1 nodes"),Ae($===0,"input layer has >1 tensors"),this.inputLayers.push(S),this.inputLayersNodeIndices.push(x),this.inputLayersTensorIndices.push($)}this.inputNames=[],this.outputNames=[],this.feedInputShapes=[],this.feedInputNames=[],this.feedOutputNames=[];for(let y=0;y<this.inputLayers.length;y++){const S=this.inputLayers[y];if(!(S instanceof Ts))throw new TypeError(`Input layers to a LayersModel must be InputLayer objects. Received inputs: ${t.inputs}. Input ${y} (0-based) originates from layer type ${S.getClassName()}.`);this.inputNames.push(S.name),this.feedInputShapes.push(S.batchInputShape),this.feedInputNames.push(S.name)}for(const y of this.outputLayers)this.outputNames.push(y.name);this.internalInputShapes=this.inputs.map(y=>y.shape),this.internalOutputShapes=this.outputs.map(y=>y.shape);const e={},s={},r={},o={},i={},a=[],l=(y,S,x,$,E,D)=>{($==null||E==null||D==null)&&($=y.sourceLayer,E=y.nodeIndex,D=y.tensorIndex);const _=$.inboundNodes[E];if(x.indexOf(_)!==-1)throw new He(`The tensor ${y.name} at layer "${$.name}" is part of a cycle.`);if(S.indexOf(_)!==-1)return;this.containerNodes.add(me.nodeKey($,E)),$.id in i||(i[$.id]=Object.keys(i).length),x.indexOf(_)===-1&&x.push(_);const T=_.inboundLayers.length;for(let P=0;P<T;P++){const B=_.inputTensors[P],Y=_.inboundLayers[P],j=_.nodeIndices[P],F=_.tensorIndices[P];l(B,S,x,Y,j,F)}for(S.push(_);x.indexOf(_)>=0;)x.splice(x.indexOf(_),1);a.push(_)},u=[],c=[];for(const y of this.outputs)l(y,u,c);const h=a.slice().reverse();for(const y of h){s[y.id]=y,y.id in e||(e[y.id]=0);let S=e[y.id];const x=r[y.outboundLayer.id]==null?0:r[y.outboundLayer.id];S=Math.max(S,x),r[y.outboundLayer.id]=S,o[y.outboundLayer.id]=y.outboundLayer,e[y.id]=S;for(let $=0;$<y.inboundLayers.length;$++){const E=y.inboundLayers[$],D=y.nodeIndices[$],_=E.inboundNodes[D],T=e[_.id]==null?0:e[_.id];e[_.id]=Math.max(S+1,T),s[_.id]=_}}const f={};for(const y in e){const S=e[y];S in f||(f[S]=[]),f[S].push(s[y])}const d={};for(const y in r){const S=r[y];S in d||(d[S]=[]),d[S].push(o[y])}let p=Object.keys(d).map(y=>parseInt(y,10)).sort(gr);this.layers=[];for(const y of p){const S=d[y];S.sort((x,$)=>{const E=i[x.id],D=i[$.id];return E<D?-1:E>D?1:0});for(const x of S)x instanceof me&&this.internalContainerRefs.push(x),this.layers.push(x)}this.layersByDepth=d,p=Object.keys(f).map(y=>parseInt(y,10)).sort(gr);const g=this.inputs.slice(),m=[];for(const y of p)for(const S of f[y]){const x=S.outboundLayer;if(x!=null){for(const $ of S.inputTensors)if(g.indexOf($)===-1)throw new He(`Graph disconnected: cannot obtain value for tensor ${$} at layer "${x.name}". The following previous layers were accessed without issue: ${m}`);for(const $ of S.outputTensors)g.push($);m.push(x.name)}}this.nodesByDepth=f;const b=this.layers.map(y=>y.name);for(const y of b){const S=b.filter(x=>x===y).length;if(S!==1)throw new He(`The name "${y}" is used ${S} times in the model. All layer names should be unique. Layer names: `+JSON.stringify(b))}this.outboundNodes=[],this.inboundNodes=[],new xi({outboundLayer:this,inboundLayers:[],nodeIndices:[],tensorIndices:[],inputTensors:this.inputs,outputTensors:this.outputs,inputMasks:this.inputs.map(y=>null),outputMasks:this.outputs.map(y=>null),inputShapes:this.inputs.map(y=>y.shape),outputShapes:this.outputs.map(y=>y.shape)}),this.built=!0,this._refCount=1}assertNotDisposed(){if(this._refCount===0)throw new Error(`Container '${this.name}' is already disposed.`)}dispose(){this.assertNotDisposed();const t={refCountAfterDispose:null,numDisposedVariables:0};if(--this._refCount===0){for(const e of this.layers)t.numDisposedVariables+=e.dispose().numDisposedVariables;for(const e of this.internalContainerRefs)t.numDisposedVariables+=e.dispose().numDisposedVariables}return t.refCountAfterDispose=this._refCount,t}get trainable(){return this.trainable_}set trainable(t){this.layers.forEach(e=>{e._trainableWeights.forEach(s=>s.trainable=t)}),this.trainable_=t}get trainableWeights(){if(this._trainableWeights.length>0)throw new I("Container instance unexpectedly contains _trainableWeights.The trainable weights of a Container are a union of the trainable weights of its consituent Layers. Its own _trainableWeights must remain an empty Array.");if(!this.trainable)return[];let t=[];for(const e of this.layers)t=t.concat(e.trainableWeights);return t}get nonTrainableWeights(){const t=[];for(const e of this.layers)t.push(...e.nonTrainableWeights);if(!this.trainable){const e=[];for(const s of this.layers)e.push(...s.trainableWeights);return e.concat(t)}return t}get weights(){return this.trainableWeights.concat(this.nonTrainableWeights)}loadWeights(t,e=!0){const s={};let r=0;const o=n1(t);o&&this.parseWeights(t);for(const a of this.layers)for(const[l,u]of a.weights.entries()){const c=o?`${u.name.split("/").slice(0,-1).join("/")+"/"}${l}`:u.originalName;if(s[c]!=null)throw new I(`Duplicate weight name: ${c}`);s[c]=u,r++}const i=[];for(const a in t){let l=a;if(s[a]==null){const u=a.split("/");l=u.slice(0,-2).concat([u[u.length-1]]).join("/")}if(s[l]!=null)i.push([s[l],t[a]]);else if(e)throw new I(`Provided weight data has no target variable: ${a}`);delete s[l]}if(e){const a=[];for(const l in s)a.push(l);if(a.length>0)throw new I(`${a.length} of ${r} weights are not set: ${a}`)}ju(i)}parseWeights(t){for(const e in Object.keys(t)){const s=e.split("/"),r=["vars","layer_checkpoint_dependencies"],o=s.map(i=>i.startsWith("_")?i.slice(1):i).filter(i=>!r.includes(i)).join("/");o!==e&&(t[o]=t[e],delete t[e])}}updatedConfig(){const t=this.getConfig(),e={};return e.className=this.getClassName(),e.config=t,e.kerasVersion=`tfjs-layers ${Fc}`,e.backend="TensorFlow.js",e}toJSON(t,e=!0){const s=Ni(this.updatedConfig());return e?JSON.stringify(s):s}call(t,e){return k(()=>{t=nt(t);const s=new Ke;for(let r=0;r<this.inputs.length;++r)s.add(this.inputs[r],t[r]);return Ns(this.outputs,s,e)})}computeMask(t,e){return k(()=>{t=nt(t);let s;return e==null?s=mr(null,t.length):s=nt(e),this.runInternalGraph(t,s)[1]})}computeOutputShape(t){const e=Ir(t);if(e.length!==this.inputLayers.length)throw new I(`Invalid inputShape argument ${t}: model has ${this.inputLayers.length} tensor inputs.`);const s={};for(let a=0;a<e.length;a++){const l=this.inputLayers[a],u=e[a],c=l.name+"_0_0";s[c]=u}const r=Object.keys(this.nodesByDepth).map(a=>parseInt(a,10)).sort(gr);if(r.length>1)for(const a of r){const l=this.nodesByDepth[a];for(const u of l){const c=u.outboundLayer;if(this.inputLayers.map(g=>g.id).indexOf(c.id)!==-1)continue;const h=[];for(let g=0;g<u.inboundLayers.length;g++){const m=u.inboundLayers[g],b=u.nodeIndices[g],y=u.tensorIndices[g],S=`${m.name}_${b}_${y}`,x=s[S];h.push(x)}const f=c.computeOutputShape(Ut(h)),d=Ir(f),p=c.inboundNodes.indexOf(u);for(let g=0;g<d.length;g++){const m=`${c.name}_${p}_${g}`;s[m]=d[g]}}}const o=[],i=[];for(let a=0;a<this.outputLayers.length;a++){const l=this.outputLayers[a],u=this.outputLayersNodeIndices[a],c=this.outputLayersTensorIndices[a],h=`${l.name}_${u}_${c}`;i.push(h)}for(let a=0;a<i.length;a++){const l=i[a];Ae(l in s),o.push(s[l])}return Ut(o)}runInternalGraph(t,e){e==null&&(e=mr(null,t.length));const s={};for(let l=0;l<this.inputs.length;++l){const u=this.inputs[l],c=t[l],h=e[l];s[u.id]=[c,h]}const r=Object.keys(this.nodesByDepth).map(l=>parseInt(l,10)).sort(gr);for(const l of r){const u=this.nodesByDepth[l];for(const c of u){const h=c.outboundLayer,f=c.inputTensors,d=c.outputTensors,p=new Array;for(const g of f)g.id in s&&p.push(s[g.id]);if(p.length===f.length){let g={},m,b,y,S;if(c.callArgs!=null&&(g=c.callArgs),p.length===1){const[x,$]=p[0];g.mask==null&&(g.mask=$),y=nt(h.call(x,g)),S=nt(h.computeMask(x,$)),m=[x],b=[$]}else m=p.map(x=>x[0]),b=p.map(x=>x[1]),g.mask==null&&(g.mask=b),y=nt(h.call(m,g)),S=nt(h.computeMask(m,b));if(h.activityRegularizer)throw new Z("LayersModel invocation with concrete Tensor value(s) in the presence of activity regularizer(s) is not supported yet.");for(let x=0;x<d.length;++x){const $=d[x],E=y[x],D=S[x];s[$.id]=[E,D]}}}}const o=[],i=[],a=[];for(const l of this.outputs){Ae(l.id in s,`Could not compute output ${l.name} : ${l.id}`);const[u,c]=s[l.id];a.push(u.shape),o.push(u),i.push(c)}return[o,i,a]}buildNodeConversionMap(t){const e={};let s;for(const r of this.layers){s=r instanceof me?1:0;for(let o=0;o<r.inboundNodes.length;o++){const i=me.nodeKey(r,o);this.containerNodes.has(i)&&(e[i]=s,s+=1)}}return e}getLayer(t,e){if(e!=null)return this.findLayer(e);if(t==null)throw new I("Provide either a layer name or layer index");if(typeof t=="number")return this.findLayer(t);for(const s of this.layers)if(s.name===t)return s;throw new I(`No such layer: ${t}`)}findLayer(t){if(this.layers.length<=t)throw new I(`Was asked to retrieve layer at index ${t}, but model only has ${this.layers.length} layer(s).`);return this.layers[t]}calculateLosses(){return k(()=>{const t=[];for(const e of this.layers)for(let s=0;s<e.inboundNodes.length;++s){const r=me.nodeKey(e,s);this.containerNodes.has(r)&&t.push(...e.calculateLosses())}return t})}getConfig(){const t={name:this.name},e=this.buildNodeConversionMap(this.layers),s=[];for(const i of this.layers){const a=i.getClassName(),l=i.getConfig(),u=[];for(let h=0;h<i.inboundNodes.length;h++){const f=i.inboundNodes[h],d=me.nodeKey(i,h);let p={};if(this.containerNodes.has(d)){if(f.callArgs)try{JSON.stringify(f.callArgs),p=f.callArgs}catch{console.warn(`Layer ${i.name} was passed non-serializable keyword arguments: ${f.callArgs}. They will not be included in the serialized model (and thus will be missing at deserialization time).`),p={}}if(f.inboundLayers.length>0){const g=[];for(let m=0;m<f.inboundLayers.length;m++){const b=f.inboundLayers[m],y=f.nodeIndices[m],S=f.tensorIndices[m],x=me.nodeKey(b,y);let $=e[x];$==null&&($=0),g.push([b.name,$,S,p])}u.push(g)}}}const c={};c.name=i.name,c.className=a,c.config=l,c.inboundNodes=u,s.push(c)}t.layers=s;const r=[];for(let i=0;i<this.inputLayers.length;i++){const a=this.inputLayers[i],l=this.inputLayersNodeIndices[i],u=me.nodeKey(a,l);if(!this.containerNodes.has(u))continue;let c=e[u];c==null&&(c=0);const h=this.inputLayersTensorIndices[i];r.push([a.name,c,h])}t.inputLayers=r;const o=[];for(let i=0;i<this.outputLayers.length;i++){const a=this.outputLayers[i],l=this.outputLayersNodeIndices[i],u=me.nodeKey(a,l);if(!this.containerNodes.has(u))continue;let c=e[u];c==null&&(c=0);const h=this.outputLayersTensorIndices[i];o.push([a.name,c,h])}return t.outputLayers=o,t}static fromConfig(t,e,s={},r=!1){const o={},i={};function a(m,b){m.name in i?i[m.name].push(b):i[m.name]=[b]}function l(m,b){const y=[];let S;for(const x of b){const $=x[0],E=x[1],D=x[2];if(S=x[3]==null?{}:x[3],!($ in o)){a(m,b);return}const _=o[$];if(_.inboundNodes.length<=E){a(m,b);return}const T=_.inboundNodes[E];y.push(T.outputTensors[D])}y.length>0&&m.apply(Ut(y),S)}function u(m){const b=m.name,y=Nc(m,e.customObjects!=null?e.customObjects:{});y.setFastWeightInitDuringBuild(r),o[b]=y,m.inboundNodes.forEach(x=>{if(!(x instanceof Array))throw new I(`Corrupted configuration, expected array for nodeData: ${x}`);a(y,x)})}const c=e.name,h=e.layers;for(const m of h)u(m);for(;!Fy(i);)for(const m of h){const b=o[m.name];if(b.name in i){const y=i[b.name];delete i[b.name];for(const S of y)l(b,S)}}const f=[],d=[],p=e.inputLayers;for(const m of p){const b=m[0],y=m[1],S=m[2];Ae(b in o);const $=o[b].inboundNodes[y].outputTensors;f.push($[S])}const g=e.outputLayers;for(const m of g){const b=m[0],y=m[1],S=m[2];Ae(b in o);const $=o[b].inboundNodes[y].outputTensors;d.push($[S])}return new t({inputs:f,outputs:d,name:c})}get stateful(){if(this._stateful)throw new I("Container instance unexpectedly has _stateful = true. The statefulness of a Container is determined by the Layers it contains. Its _stateful property must remain the default false.");for(const t of this.layers)if(t.stateful)return!0;return!1}resetStates(){k(()=>{this.layers.forEach(t=>{t.stateful&&t.resetStates()})})}}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -3020,7 +3020,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function s1(n,t,e){const s=t.length;if(n==null||Array.isArray(n)&&n.length===0)return t.map(r=>null);if(s===1)return Array.isArray(n)&&n.length===1?n:typeof n=="object"&&t[0]in n?[n[t[0]]]:[n];if(Array.isArray(n)){if(n.length!==s)throw new Error(`Provided ${e} is an array of ${n.length} element(s), but the model has ${s} outputs. Make sure a set of weights is provided for each model output.`);return n}else if(typeof n=="object"&&Object.keys(n).length>0&&typeof n[Object.keys(n)[0]]=="object"){const r=[];return t.forEach(o=>{o in n?r.push(n[o]):r.push(null)}),r}else throw new Error(`The model has multiple (${s}) outputs, so ${e} must be either an array with ${s} elements or an object with ${t} keys. Provided ${e} not understood: ${JSON.stringify(n)}`)}function Vc(n,t){return s1(n,t,"classWeight")}async function qc(n,t,e,s){if(e!=null){const r=k(()=>{if(n.shape.length===1)return rn(n);if(n.shape.length===2){if(n.shape[1]>1)return er(n,1);if(n.shape[1]===1)return L(n,[n.shape[0]]);throw new Error(`Encountered unexpected last-dimension size (${n.shape[1]}) during handling of class weights. The size is expected to be >= 1.`)}else throw new Error(`Unexpected rank of target (y) tensor (${n.rank}) during handling of class weights. The rank is expected to be 1 or 2.`)}),o=Array.from(await r.data());ut(r);const i=[];return o.forEach(a=>{if(e[a]==null)throw new Error(`classWeight must contain all classes in the training data. The class ${a} exists in the data but not in classWeight`);i.push(e[a])}),Rt(i,"float32")}else return null}function r1(n,t){return N(n,t)}/**
+ */function s1(n,t,e){const s=t.length;if(n==null||Array.isArray(n)&&n.length===0)return t.map(r=>null);if(s===1)return Array.isArray(n)&&n.length===1?n:typeof n=="object"&&t[0]in n?[n[t[0]]]:[n];if(Array.isArray(n)){if(n.length!==s)throw new Error(`Provided ${e} is an array of ${n.length} element(s), but the model has ${s} outputs. Make sure a set of weights is provided for each model output.`);return n}else if(typeof n=="object"&&Object.keys(n).length>0&&typeof n[Object.keys(n)[0]]=="object"){const r=[];return t.forEach(o=>{o in n?r.push(n[o]):r.push(null)}),r}else throw new Error(`The model has multiple (${s}) outputs, so ${e} must be either an array with ${s} elements or an object with ${t} keys. Provided ${e} not understood: ${JSON.stringify(n)}`)}function Vc(n,t){return s1(n,t,"classWeight")}async function qc(n,t,e,s){if(e!=null){const r=k(()=>{if(n.shape.length===1)return on(n);if(n.shape.length===2){if(n.shape[1]>1)return nr(n,1);if(n.shape[1]===1)return L(n,[n.shape[0]]);throw new Error(`Encountered unexpected last-dimension size (${n.shape[1]}) during handling of class weights. The size is expected to be >= 1.`)}else throw new Error(`Unexpected rank of target (y) tensor (${n.rank}) during handling of class weights. The rank is expected to be 1 or 2.`)}),o=Array.from(await r.data());ut(r);const i=[];return o.forEach(a=>{if(e[a]==null)throw new Error(`classWeight must contain all classes in the training data. The class ${a} exists in the data but not in classWeight`);i.push(e[a])}),Rt(i,"float32")}else return null}function r1(n,t){return N(n,t)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -3028,7 +3028,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */const o1=32;function jc(n,t){let e,s;const r=t;e=r.xs,s=r.ys,w(e!=null&&s!=null,()=>`A Dataset iterator for fitDataset() is expected to generate objects of the form \`{xs: xVal, ys: yVal}\`, where the two values may be \`tf.Tensor\`, an array of Tensors, or a map of string to Tensor.  The provided Dataset instead generates ${t}`);const o=Hc("input",n.inputNames,e),i=Hc("output",n.outputNames,s),a=o[0].shape[0];w(o.length===n.inputs.length,()=>`LayersModel has ${n.inputs.length} inputs, but the dataset provides ${o.length} inputs.  (Expected input keys: ${JSON.stringify(n.inputNames)})`),w(i.length===n.outputs.length,()=>`LayersModel has ${n.outputs.length} outputs, but the dataset provides ${i.length} outputs.  (Expected output keys: ${JSON.stringify(n.outputNames)})`);for(let l=0;l<o.length;l++)w(o[l].shape[0]===a,()=>`Batch size mismatch: input ${n.inputNames[l]} has ${o[l].shape[0]}; expected  ${a} based on input ${n.inputNames[0]}.`);for(let l=0;l<i.length;l++)w(i[l].shape[0]===a,()=>`Batch size mismatch: output ${n.outputNames[l]} has ${i[l].shape[0]}; expected  ${a} based on input ${n.inputNames[0]}.`);return{xs:o,ys:i}}function Hc(n,t,e){if(e instanceof At)return[e];if(Array.isArray(e))return w(e.length===t.length,()=>`Received an array of ${e.length} Tensors, but expected ${t.length} to match the ${n} keys ${t}.`),e;{const s=[];for(const r of t){if(e[r]==null)throw new I(`The feature data generated by the dataset lacks the required ${n} key '${r}'.`);s.push(e[r])}return s}}function i1(n){if(n.length===3)throw new Z("Validation with sample weights is not implemented yet.");return{xs:n[0],ys:n[1]}}async function a1(n,t,e){const s=e.batchesPerEpoch!=null;if(w(n.optimizer!=null,()=>"You must compile a model before training/testing. Use LayersModel.compile(modelCompileConfig)."),w(e!=null,()=>"For fitDataset(), the 2nd argument (config) is required, but it is not provided in this call."),w(e.epochs!=null&&e.epochs>0&&Number.isInteger(e.epochs),()=>`For fitDataset(), config.epochs is expected to be a positive integer, but got ${e.epochs}`),w(!s||e.batchesPerEpoch>0&&Number.isInteger(e.batchesPerEpoch),()=>`For fitDataset(), config.batchesPerEpoch is expected to be a positive integer if specified, but got ${e.batchesPerEpoch}`),w(e.validationSplit==null,()=>"`validationSplit` is not supported by `fitDataset()`. Use validationData instead."),n.isTraining)throw new Error("Cannot start training because another fit() call is ongoing.");n.isTraining=!0;try{const r=e.validationData!=null;let o,i;if(r)if(Kc(e.validationData))w(e.validationBatches==null||e.validationBatches>0&&Number.isInteger(e.validationBatches),()=>`For fitDataset() with dataset-based validation, config.validationBatches is expected not to be provided, or to be a positive integer, but got ${e.validationBatches}`);else{const m=i1(e.validationData);o=m.xs,i=m.ys}const a=n.makeTrainFunction(),l=n.getDedupedMetricsNames();let u;r?u=l.slice().concat(l.map(m=>"val_"+m)):u=l.slice();const c=_c(e.callbacks,e.yieldEvery),h=e.verbose==null?1:e.verbose,{callbackList:f,history:d}=Tc(c,h,e.epochs,null,null,l1(t,e),null,r,u);f.setModel(n),n.history=d,await f.onTrainBegin(),n.stopTraining_=!1;let p=e.initialEpoch==null?0:e.initialEpoch,g=await t.iterator();for(;p<e.epochs;){const m={};await f.onEpochBegin(p);let b=0,y=0;for(s||(g=await t.iterator());!s||b<e.batchesPerEpoch;){const S=await g.next();if(s&&S.done){console.warn(`You provided \`batchesPerEpoch\` as ${e.batchesPerEpoch}, but your dataset iterator ran out of data after ${b} batches; interrupting training. Make sure that your dataset can generate at least \`batchesPerEpoch * epochs\` batches (in this case, ${e.batchesPerEpoch*e.epochs} batches). You may need to use the repeat() function when building your dataset.`);break}if(S.value!=null){const{xs:x,ys:$}=jc(n,S.value),E={};E.batch=y,E.size=x[0].shape[0],await f.onBatchBegin(y,E);const D=[];if(e.classWeight!=null){const P=Vc(e.classWeight,n.outputNames);for(let B=0;B<P.length;++B)D.push(await qc($[B],null,P[B]))}const _=x.concat($).concat(D),T=a(_);ut(_);for(let P=0;P<l.length;++P){const B=l[P],Y=T[P];E[B]=Y,Ln(Y)}await f.onBatchEnd(y,E),Cc(E),y++,b++}if(s?b>=e.batchesPerEpoch:S.done){if(r){let x;Kc(e.validationData)?x=nt(await n.evaluateDataset(e.validationData,{batches:e.validationBatches})):x=nt(n.evaluate(o,i,{batchSize:e.validationBatchSize==null?o1:e.validationBatchSize,verbose:0}));for(let $=0;$<n.metricsNames.length;++$)m[`val_${n.metricsNames[$]}`]=x[$]}break}if(n.stopTraining_)break}if(await f.onEpochEnd(p,m),p++,n.stopTraining_)break}return await f.onTrainEnd(),await n.history.syncData(),n.history}finally{n.isTraining=!1}}function l1(n,t){let e=null;return t.batchesPerEpoch!=null?e=t.batchesPerEpoch:Number.isFinite(n.size)&&(e=n.size),e}function Kc(n){return typeof n.iterator=="function"}function u1(n){return typeof n.next=="function"}async function c1(n,t,e){e=e||{};const s=e.batches!=null,r=n.testFunction;let o=[];if(e.verbose>0)throw new Z("Verbose mode is not implemented yet.");w(!s||e.batches>0&&Number.isInteger(e.batches),()=>`Test loop expects \`batches\` to be a positive integer, but received ${JSON.stringify(e.batches)}`);const i=u1(t)?t:await t.iterator();let a=0,l=0;for(;!s||l<e.batches;){const u=await i.next();if(o=k(()=>{if(u.value){const{xs:c,ys:h}=jc(n,u.value),f=c.concat(h),d=k(()=>r(f));if(ut(f),l===0)for(let g=0;g<d.length;++g)o.push(Ht(0));const p=f[0].shape[0];for(let g=0;g<d.length;++g){const m=d[g],b=o[g];o[g]=k(()=>O(o[g],N(p,m))),l>0&&ut(b)}ut(d),a+=p,++l}return o}),u.done){s&&console.warn(`Your dataset iterator ran out of data during evaluateDataset(). Interrupting evalution. Make sure that your dataset can generate at least \`batches\` batches (in this case, ${e.batches} batches). You may need to use the repeat() function when building your dataset.`);break}}for(let u=0;u<o.length;++u){const c=o[u];o[u]=X(o[u],a),ut(c)}return Ut(o)}/**
+ */const o1=32;function jc(n,t){let e,s;const r=t;e=r.xs,s=r.ys,w(e!=null&&s!=null,()=>`A Dataset iterator for fitDataset() is expected to generate objects of the form \`{xs: xVal, ys: yVal}\`, where the two values may be \`tf.Tensor\`, an array of Tensors, or a map of string to Tensor.  The provided Dataset instead generates ${t}`);const o=Hc("input",n.inputNames,e),i=Hc("output",n.outputNames,s),a=o[0].shape[0];w(o.length===n.inputs.length,()=>`LayersModel has ${n.inputs.length} inputs, but the dataset provides ${o.length} inputs.  (Expected input keys: ${JSON.stringify(n.inputNames)})`),w(i.length===n.outputs.length,()=>`LayersModel has ${n.outputs.length} outputs, but the dataset provides ${i.length} outputs.  (Expected output keys: ${JSON.stringify(n.outputNames)})`);for(let l=0;l<o.length;l++)w(o[l].shape[0]===a,()=>`Batch size mismatch: input ${n.inputNames[l]} has ${o[l].shape[0]}; expected  ${a} based on input ${n.inputNames[0]}.`);for(let l=0;l<i.length;l++)w(i[l].shape[0]===a,()=>`Batch size mismatch: output ${n.outputNames[l]} has ${i[l].shape[0]}; expected  ${a} based on input ${n.inputNames[0]}.`);return{xs:o,ys:i}}function Hc(n,t,e){if(e instanceof At)return[e];if(Array.isArray(e))return w(e.length===t.length,()=>`Received an array of ${e.length} Tensors, but expected ${t.length} to match the ${n} keys ${t}.`),e;{const s=[];for(const r of t){if(e[r]==null)throw new I(`The feature data generated by the dataset lacks the required ${n} key '${r}'.`);s.push(e[r])}return s}}function i1(n){if(n.length===3)throw new Z("Validation with sample weights is not implemented yet.");return{xs:n[0],ys:n[1]}}async function a1(n,t,e){const s=e.batchesPerEpoch!=null;if(w(n.optimizer!=null,()=>"You must compile a model before training/testing. Use LayersModel.compile(modelCompileConfig)."),w(e!=null,()=>"For fitDataset(), the 2nd argument (config) is required, but it is not provided in this call."),w(e.epochs!=null&&e.epochs>0&&Number.isInteger(e.epochs),()=>`For fitDataset(), config.epochs is expected to be a positive integer, but got ${e.epochs}`),w(!s||e.batchesPerEpoch>0&&Number.isInteger(e.batchesPerEpoch),()=>`For fitDataset(), config.batchesPerEpoch is expected to be a positive integer if specified, but got ${e.batchesPerEpoch}`),w(e.validationSplit==null,()=>"`validationSplit` is not supported by `fitDataset()`. Use validationData instead."),n.isTraining)throw new Error("Cannot start training because another fit() call is ongoing.");n.isTraining=!0;try{const r=e.validationData!=null;let o,i;if(r)if(Kc(e.validationData))w(e.validationBatches==null||e.validationBatches>0&&Number.isInteger(e.validationBatches),()=>`For fitDataset() with dataset-based validation, config.validationBatches is expected not to be provided, or to be a positive integer, but got ${e.validationBatches}`);else{const m=i1(e.validationData);o=m.xs,i=m.ys}const a=n.makeTrainFunction(),l=n.getDedupedMetricsNames();let u;r?u=l.slice().concat(l.map(m=>"val_"+m)):u=l.slice();const c=_c(e.callbacks,e.yieldEvery),h=e.verbose==null?1:e.verbose,{callbackList:f,history:d}=Tc(c,h,e.epochs,null,null,l1(t,e),null,r,u);f.setModel(n),n.history=d,await f.onTrainBegin(),n.stopTraining_=!1;let p=e.initialEpoch==null?0:e.initialEpoch,g=await t.iterator();for(;p<e.epochs;){const m={};await f.onEpochBegin(p);let b=0,y=0;for(s||(g=await t.iterator());!s||b<e.batchesPerEpoch;){const S=await g.next();if(s&&S.done){console.warn(`You provided \`batchesPerEpoch\` as ${e.batchesPerEpoch}, but your dataset iterator ran out of data after ${b} batches; interrupting training. Make sure that your dataset can generate at least \`batchesPerEpoch * epochs\` batches (in this case, ${e.batchesPerEpoch*e.epochs} batches). You may need to use the repeat() function when building your dataset.`);break}if(S.value!=null){const{xs:x,ys:$}=jc(n,S.value),E={};E.batch=y,E.size=x[0].shape[0],await f.onBatchBegin(y,E);const D=[];if(e.classWeight!=null){const P=Vc(e.classWeight,n.outputNames);for(let B=0;B<P.length;++B)D.push(await qc($[B],null,P[B]))}const _=x.concat($).concat(D),T=a(_);ut(_);for(let P=0;P<l.length;++P){const B=l[P],Y=T[P];E[B]=Y,Ln(Y)}await f.onBatchEnd(y,E),Cc(E),y++,b++}if(s?b>=e.batchesPerEpoch:S.done){if(r){let x;Kc(e.validationData)?x=nt(await n.evaluateDataset(e.validationData,{batches:e.validationBatches})):x=nt(n.evaluate(o,i,{batchSize:e.validationBatchSize==null?o1:e.validationBatchSize,verbose:0}));for(let $=0;$<n.metricsNames.length;++$)m[`val_${n.metricsNames[$]}`]=x[$]}break}if(n.stopTraining_)break}if(await f.onEpochEnd(p,m),p++,n.stopTraining_)break}return await f.onTrainEnd(),await n.history.syncData(),n.history}finally{n.isTraining=!1}}function l1(n,t){let e=null;return t.batchesPerEpoch!=null?e=t.batchesPerEpoch:Number.isFinite(n.size)&&(e=n.size),e}function Kc(n){return typeof n.iterator=="function"}function u1(n){return typeof n.next=="function"}async function c1(n,t,e){e=e||{};const s=e.batches!=null,r=n.testFunction;let o=[];if(e.verbose>0)throw new Z("Verbose mode is not implemented yet.");w(!s||e.batches>0&&Number.isInteger(e.batches),()=>`Test loop expects \`batches\` to be a positive integer, but received ${JSON.stringify(e.batches)}`);const i=u1(t)?t:await t.iterator();let a=0,l=0;for(;!s||l<e.batches;){const u=await i.next();if(o=k(()=>{if(u.value){const{xs:c,ys:h}=jc(n,u.value),f=c.concat(h),d=k(()=>r(f));if(ut(f),l===0)for(let g=0;g<d.length;++g)o.push(Yt(0));const p=f[0].shape[0];for(let g=0;g<d.length;++g){const m=d[g],b=o[g];o[g]=k(()=>O(o[g],N(p,m))),l>0&&ut(b)}ut(d),a+=p,++l}return o}),u.done){s&&console.warn(`Your dataset iterator ran out of data during evaluateDataset(). Interrupting evalution. Make sure that your dataset can generate at least \`batches\` batches (in this case, ${e.batches} batches). You may need to use the repeat() function when building your dataset.`);break}}for(let u=0;u<o.length;++u){const c=o[u];o[u]=X(o[u],a),ut(c)}return Ut(o)}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -3036,7 +3036,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function Di(n){w(n>0&&Number.isInteger(n),()=>`batchSize is required to be a positive integer, but got ${n}`)}function Ns(n,t,e){return n==null?[null]:Array.isArray(n)?n.map(s=>gn(s,t,e-t)):gn(n,t,e-t)}function Ri(n,t){return k(()=>n==null?null:Array.isArray(n)?n.map(e=>Ri(e,t)):Ky(n,t.dtype==="int32"?t:ot(t,"int32")))}function Pi(n,t){const e=[];let s=0,r=null;for(;s<n;)r=s+t,r>=n&&(r=n),e.push([s,r]),s=r;return e}function Yc(n){const t=[];n instanceof At&&(n=[n]);for(let e=0;e<n.length;++e){const s=n[e];if(s.rank===1)t.push(ui(s,1));else{if(s.rank===0)throw new Error("Expected tensor to be at least 1D, but received a 0D tensor (scalar).");t.push(s)}}return t}function ge(n,t){if(n==null)return;const e=[];if(t instanceof At)e.push(t.id);else if(Array.isArray(t))t.forEach(r=>e.push(r.id));else if(t!=null)for(const r in t){const o=t[r];e.push(o.id)}const s=[];if(n instanceof At)e.indexOf(n.id)===-1&&s.push(n);else if(Array.isArray(n))n.forEach(r=>{e.indexOf(r.id)===-1&&s.push(r)});else if(n!=null)for(const r in n){const o=n[r];e.indexOf(o.id)===-1&&s.push(o)}s.forEach(r=>{r.isDisposed||r.dispose()})}/**
+ */function Di(n){w(n>0&&Number.isInteger(n),()=>`batchSize is required to be a positive integer, but got ${n}`)}function Ds(n,t,e){return n==null?[null]:Array.isArray(n)?n.map(s=>bn(s,t,e-t)):bn(n,t,e-t)}function Ri(n,t){return k(()=>n==null?null:Array.isArray(n)?n.map(e=>Ri(e,t)):Ky(n,t.dtype==="int32"?t:ot(t,"int32")))}function Pi(n,t){const e=[];let s=0,r=null;for(;s<n;)r=s+t,r>=n&&(r=n),e.push([s,r]),s=r;return e}function Yc(n){const t=[];n instanceof At&&(n=[n]);for(let e=0;e<n.length;++e){const s=n[e];if(s.rank===1)t.push(ui(s,1));else{if(s.rank===0)throw new Error("Expected tensor to be at least 1D, but received a 0D tensor (scalar).");t.push(s)}}return t}function ge(n,t){if(n==null)return;const e=[];if(t instanceof At)e.push(t.id);else if(Array.isArray(t))t.forEach(r=>e.push(r.id));else if(t!=null)for(const r in t){const o=t[r];e.push(o.id)}const s=[];if(n instanceof At)e.indexOf(n.id)===-1&&s.push(n);else if(Array.isArray(n))n.forEach(r=>{e.indexOf(r.id)===-1&&s.push(r)});else if(n!=null)for(const r in n){const o=n[r];e.indexOf(o.id)===-1&&s.push(o)}s.forEach(r=>{r.isDisposed||r.dispose()})}/**
  * @license
  * Copyright 2018 Google LLC
  *
@@ -3044,7 +3044,7 @@
  * license that can be found in the LICENSE file or at
  * https://opensource.org/licenses/MIT.
  * =============================================================================
- */function h1(n){return n instanceof At}function Li(n){return Array.isArray(n)}function Xc(n){return!h1(n)&&!Li(n)}function Jc(n,t,e,s=!0,r=""){if(t==null||t.length===0){if(n!=null){let i=!1;if(Li(n)&&n.length>0)i=!0;else if(Xc(n)){for(const a in n)if(n.hasOwnProperty(a)){i=!0;break}}else i=!0;if(i)throw new I(`Error when checking model ${r} expected no data, but got ${n}`)}return[]}if(n==null)return t.map(i=>null);let o;if(Xc(n)){n=n,o=[];for(const i of t){if(n[i]==null)throw new I(`No data provided for "${i}". Need data for each key in: ${t}`);o.push(n[i])}}else if(Li(n)){if(n=n,n.length!==t.length)throw new I(`Error when checking model ${r}: the Array of Tensors that you are passing to your model is not the size the model expected. Expected to see ${t.length} Tensor(s), but instead got the following list of Tensor(s): ${n}`);o=n}else{if(n=n,t.length>1)throw new I(`The model ${r} expects ${t.length} Tensor(s), but only received one Tensor. Found: Tensor with shape ${n.shape}`);o=[n]}if(o=Yc(o),e!=null)for(let i=0;i<t.length;++i){if(e[i]==null)continue;const a=o[i];if(a.shape.length!==e[i].length)throw new I(`Error when checking ${r}: expected ${t[i]} to have ${e[i].length} dimension(s). but got array with shape ${a.shape}`);for(let l=0;l<e[i].length;++l){if(l===0&&!s)continue;const u=a.shape[l],c=e[i][l];if(c!=null&&c>=0&&u!==c)throw new I(`${r} expected a batch of elements where each example has shape [${e[i].slice(1,e[i].length)}] (i.e.,tensor shape [*,${e[i].slice(1,e[i].length)}]) but the ${r} received an input with ${a.shape[0]} examples, each with shape [${a.shape.slice(1,a.shape.length)}] (tensor shape [${a.shape}])`)}}return o}function f1(n,t,e){const s=mn(n.map(o=>o.shape[0]));s.sort();const r=mn(t.map(o=>o.shape[0]));if(r.sort(),s.length>1)throw new I(`All input Tensors (x) should have the same number of samples. Got array shapes: ${JSON.stringify(n.map(o=>o.shape))}`);if(r.length>1)throw new I(`All target Tensors (y) should have the same number of samples. Got array shapes: ${JSON.stringify(t.map(o=>o.shape))}`);if(s.length>0&&r.length>0&&!Zt(s,r))throw new I(`Input Tensors should have the same number of samples as target Tensors. Found ${s[0]} input sample(s) and ${r[0]} target sample(s).`)}function d1(n,t,e){const s=[kr,Tr,Es];for(let r=0;r<n.length;++r){const o=n[r],i=t[r],a=e[r];if(i!=null){if(i===Es&&o.shape[o.shape.length-1]===1)throw new I(`You are passing a target array of shape ${o.shape} while using a loss 'categorical_crossentropy'. 'categorical_crossentropy'expects targets to be binary matrices (1s and 0s) of shape [samples, classes].`);if(s.indexOf(i)!==-1){const l=o.shape.slice(1),u=a.slice(1);for(let c=0;c<l.length;++c){const h=l[c],f=u[c];if(f!=null&&h!==f)throw new I(`A target Tensor with shape ${o.shape} was passed for an output of shape ${a}, while using a loss function that expects targets to have the same shape as the output.`)}}}}}function Zc(n,t,e,s=!0,r=""){let o;if(Array.isArray(n)){if(n.length!==t.length)throw new I(`Error when checking model ${r}: the Array of Tensors that you are passing to your model is not the size the the model expected. Expected to see ${t.length} Tensor(s), but instead got ${n.length} Tensors(s).`);o=n}else{if(t.length>1)throw new I(`The model expects ${t.length} ${r} Tensors, but only received one Tensor. Found: array with shape ${JSON.stringify(n.shape)}.`);o=[n]}if(e!=null)for(let i=0;i<t.length;++i){if(e[i]==null)continue;const a=o[i];if(a.shape.length!==e[i].length)throw new I(`Error when checking ${r}: expected ${t[i]} to have ${e[i].length} dimension(s), but got array with shape ${JSON.stringify(a.shape)}`);for(let l=0;l<e[i].length;++l){if(l===0&&!s)continue;const u=a.shape[l],c=e[i][l];if(c!=null&&c!==u)throw new I(`Error when checking ${r}: expected ${t[i]} to have shape ${JSON.stringify(e[i])} but got array with shape ${JSON.stringify(a.shape)}.`)}}}function p1(n,t){if(n==null||Array.isArray(n)&&n.length===0)return t.map(s=>[]);let e;if(typeof n=="string"||typeof n=="function")e=[n];else if(Array.isArray(n)||typeof n=="object")e=n;else throw new TypeError(`Type of metrics argument not understood. Expected an string,function, Array, or Object, found: ${n}`);if(Array.isArray(e))return t.map(s=>e);{const s=[];for(const r of t){let o=e.hasOwnProperty(r)?e[r]:[];Array.isArray(o)||(o=[o]),s.push(o)}return s}}const m1="layers-model";class Lr extends me{constructor(t){super(t),this.isTraining=!1}summary(t,e,s=console.log){if(!this.built)throw new I("This model has never been called, thus its weights have not been created yet. So no summary can be displayed. Build the model first (e.g., by calling it on some test data).");jw(this,t,e,s)}compile(t){if(t.loss==null&&(t.loss=[]),this.loss=t.loss,typeof t.optimizer=="string")this.optimizer_=qw(t.optimizer),this.isOptimizerOwned=!0;else{if(!(t.optimizer instanceof qe))throw new I("User-defined optimizer must be an instance of tf.Optimizer.");this.optimizer_=t.optimizer,this.isOptimizerOwned=!1}let e=[];if(!Array.isArray(t.loss)&&typeof t.loss!="string"&&typeof t.loss!="function"){t.loss=t.loss;for(const i in t.loss)if(this.outputNames.indexOf(i)===-1)throw new I(`Unknown entry in loss dictionary: "${i}". Only expected the following keys: ${this.outputNames}`);for(const i of this.outputNames)t.loss[i]==null&&console.warn(`Output "${i}" is missing from loss dictionary. We assume this was done on purpose, and we will not be expecting data to be passed to ${i} during training`),e.push(Ci(t.loss[i]))}else if(Array.isArray(t.loss)){if(t.loss.length!==this.outputs.length)throw new I(`When passing an Array as loss, it should have one entry per model output. The model has ${this.outputs.length} output(s), but you passed loss=${t.loss}.`);e=t.loss.map(a=>Ci(a))}else{const i=Ci(t.loss);this.outputs.forEach(a=>{e.push(i)})}this.lossFunctions=e,this.feedOutputNames=[],this.feedOutputShapes=[],this.feedLossFns=[];for(let i=0;i<this.outputs.length;++i){const a=this.internalOutputShapes[i],l=this.outputNames[i];this.feedOutputNames.push(l),this.feedOutputShapes.push(a),this.feedLossFns.push(this.lossFunctions[i])}const s=[];this.metrics=t.metrics,this.metricsNames=["loss"],this.metricsTensors=[],gr("loss",()=>{for(let i=0;i<this.outputs.length;++i){if(s.indexOf(i)!==-1)continue;const a=this.lossFunctions[i];this.outputs.length>1&&(this.metricsTensors.push([a,i]),this.metricsNames.push(this.outputNames[i]+"_loss"))}});const r=p1(t.metrics,this.outputNames),o=(i,a,l)=>{this.outputNames.length>1&&(a=this.outputNames[i]+"_"+a),this.metricsNames.push(a),this.metricsTensors.push([l,i])};gr("metric",()=>{for(let i=0;i<this.outputs.length;++i){if(s.indexOf(i)!==-1)continue;const a=r[i];(u=>{const c="";let h,f,d;for(const p of u){if(typeof p=="string"&&["accuracy","acc","crossentropy","ce"].indexOf(p)!==-1){const m=this.internalOutputShapes[i];m[m.length-1]===1||this.lossFunctions[i]===Tr?["accuracy","acc"].indexOf(p)!==-1?f=Dc:["crossentropy","ce"].indexOf(p)!==-1&&(f=Lw):this.lossFunctions[i]===_r?["accuracy","acc"].indexOf(p)!==-1?f=Mw:["crossentropy","ce"].indexOf(p)!==-1&&(f=Lc):["accuracy","acc"].indexOf(p)!==-1?f=Rc:["crossentropy","ce"].indexOf(p)!==-1&&(f=Pc);let b;["accuracy","acc"].indexOf(p)!==-1?b="acc":["crossentropy","ce"].indexOf(p)!==-1&&(b="ce"),d=f,h=c+b}else d=Vw(p),h=c+Rr(p);let g;gr(h,()=>{g=d}),o(i,h,g)}})(a)}}),this.collectedTrainableWeights=this.trainableWeights}checkTrainableWeightsConsistency(){this.collectedTrainableWeights!=null&&this.trainableWeights.length!==this.collectedTrainableWeights.length&&console.warn("Discrepancy between trainableweights and collected trainable weights. Did you set `model.trainable` without calling `model.compile()` afterwards?")}evaluate(t,e,s={}){const r=s.batchSize==null?32:s.batchSize;Di(r);const i=this.standardizeUserDataXY(t,e,!0,r);try{const a=i[0].concat(i[1]);this.makeTestFunction();const l=this.testFunction,u=this.testLoop(l,a,r,s.verbose,s.steps);return Ut(u)}finally{ge(i[0],t),ge(i[1],e)}}async evaluateDataset(t,e){return this.makeTestFunction(),c1(this,t,e)}checkNumSamples(t,e,s,r="steps"){let o;if(s!=null){if(o=null,e!=null)throw new I(`If ${r} is set, batchSize must be null or undefined.Got batchSize = ${e}`)}else if(t!=null)Array.isArray(t)?o=t[0].shape[0]:o=t.shape[0];else throw new I(`Either the input data should have a defined shape, or ${r} shoud be specified.`);return o}execute(t,e){if(Array.isArray(e)&&e.length===0)throw new I("`outputs` is an empty Array, which is not allowed.");const s=Array.isArray(e),r=s?e:[e],o=this.retrieveSymbolicTensors(r),i=new Ke;if(t instanceof At&&(t=[t]),Array.isArray(t)){if(t.length!==this.inputs.length)throw new I(`The number of inputs provided (${t.length}) does not match the number of inputs of this model (${this.inputs.length}).`);for(let l=0;l<this.inputs.length;++l)i.add(this.inputs[l],t[l])}else for(const l of this.inputs){const u=t[l.name];if(u==null)throw new I(`No value is provided for the model's input ${l.name}`);i.add(l,u)}const a=Ts(o,i);return s?a:a[0]}retrieveSymbolicTensors(t){const e=pr(null,t.length);let s=t.length;for(const r of this.layers){const o=Array.isArray(r.output)?r.output:[r.output],i=o.map(a=>a.name);for(let a=0;a<t.length;++a){const l=i.indexOf(t[a]);if(l!==-1&&(e[a]=o[l],s--),s===0)break}if(s===0)break}if(s>0){const r=[];throw e.forEach((o,i)=>{o==null&&r.push(t[i])}),new I(`Cannot find SymbolicTensors for output name(s): ${JSON.stringify(r)}`)}return e}predictLoop(t,e=32,s=!1){return k(()=>{const r=this.checkNumSamples(t);if(s)throw new Z("Verbose predictLoop() is not implemented yet.");const o=Pi(r,e),i=this.outputs.map(a=>[]);for(let a=0;a<o.length;++a)k(()=>{const u=o[a][0],c=o[a][1],h=Ns(t,u,c),f=[];if(Array.isArray(h))for(let p=0;p<h.length;++p)f.push({key:this.inputs[p],value:h[p]});else f.push({key:this.inputs[0],value:h});const d=new Ke(f);return Ts(this.outputs,d)}).forEach((u,c)=>i[c].push(u));return Ut(i.map(a=>on(a,0)))})}predict(t,e={}){const s=Yc(t);Zc(s,this.inputNames,this.feedInputShapes,!1);try{const r=e.batchSize==null?32:e.batchSize;return Di(r),this.predictLoop(s,r)}finally{ge(s,t)}}predictOnBatch(t){Zc(t,this.inputNames,this.feedInputShapes,!0);const e=(Array.isArray(t)?t[0]:t).shape[0];return this.predictLoop(t,e)}standardizeUserDataXY(t,e,s=!0,r){if(this.optimizer_==null)throw new He("You must compile a model before training/testing. Use LayersModel.compile(modelCompileArgs).");const o=[];for(let i=0;i<this.feedOutputShapes.length;++i){const a=this.feedOutputShapes[i];this.feedLossFns[i]===_r?o.push(a.slice(0,a.length-1).concat([1])):o.push(a)}if(t=Jc(t,this.feedInputNames,this.feedInputShapes,!1,"input"),e=Jc(e,this.feedOutputNames,o,!1,"target"),f1(t,e),d1(e,this.feedLossFns,this.feedOutputShapes),this.stateful&&r!=null&&r>0&&t[0].shape[0]%r!==0)throw new I(`In a stateful network, you should only pass inputs with a number of samples that is divisible by the batch size ${r}. Found: ${t[0].shape[0]} sample(s).`);return[t,e]}async standardizeUserData(t,e,s,r,o=!0,i){const[a,l]=this.standardizeUserDataXY(t,e,o,i);if(s!=null)throw new Error("sample weight is not supported yet.");let u=null;if(r!=null){const c=Vc(r,this.outputNames);u=[];for(let h=0;h<c.length;++h)u.push(await qc(l[h],null,c[h]))}return[a,l,u]}testLoop(t,e,s,r=0,o){return k(()=>{const i=this.checkNumSamples(e,s,o,"steps"),a=[];if(r>0)throw new Z("Verbose mode is not implemented yet.");if(o!=null)throw new Z("steps mode in testLoop() is not implemented yet");{const l=Pi(i,s),u=Rt(br(0,i));for(let c=0;c<l.length;++c){const h=l[c][0],f=l[c][1],d=gn(u,h,f-h),p=Ri(e,d),g=t(p);if(c===0)for(let m=0;m<g.length;++m)a.push(Ht(0));for(let m=0;m<g.length;++m){const b=g[m];a[m]=O(a[m],N(f-h,b))}}for(let c=0;c<a.length;++c)a[c]=X(a[c],i)}return a})}getDedupedMetricsNames(){const t=this.metricsNames,e=[];for(let s=0;s<t.length;++s){const r=t[s];let o=r;if(eu(t,r)>1){const i=eu(t.slice(0,s),r);o+=`_${i}`}e.push(o)}return e}makeTrainFunction(){return t=>{const e=[],s=t.slice(0,this.inputs.length),r=t.slice(this.inputs.length,this.inputs.length+this.outputs.length),o=t.slice(this.inputs.length+this.outputs.length,this.inputs.length+this.outputs.length*2),i=[],a=()=>{const h=[];for(let g=0;g<this.inputs.length;++g)h.push({key:this.inputs[g],value:s[g]});const f=new Ke(h),d=Ts(this.outputs,f,{training:!0});let p;for(let g=0;g<this.lossFunctions.length;++g){const m=this.lossFunctions[g];let b=m(r[g],d[g]);o[g]!=null&&(b=r1(b,o[g]));const y=St(b);e.push(y),g===0?p=b:p=O(p,b)}for(let g=0;g<this.metricsTensors.length;++g){let m;if(this.outputs.length>1&&g<this.outputs.length)m=e[g];else{const b=this.metricsTensors[g][0],y=this.metricsTensors[g][1];m=St(b(r[y],d[y]))}Ln(m),i.push(m)}return p=St(p),this.calculateLosses().forEach(g=>{p=O(p,g)}),p},l=this.collectedTrainableWeights.map(h=>h.read());return[this.optimizer_.minimize(a,!0,l)].concat(i)}}makeTestFunction(){this.testFunction=t=>k(()=>{const e=[];let s;const r=t.slice(0,this.inputs.length),o=t.slice(this.inputs.length,this.inputs.length+this.outputs.length),i=[];for(let u=0;u<this.inputs.length;++u)i.push({key:this.inputs[u],value:r[u]});const a=new Ke(i),l=Ts(this.outputs,a);for(let u=0;u<this.lossFunctions.length;++u){const c=this.lossFunctions[u],h=St(c(o[u],l[u]));u===0?s=h:s=O(s,h),e.push(s)}for(let u=0;u<this.metricsTensors.length;++u){const c=this.metricsTensors[u][0],h=this.metricsTensors[u][1],f=St(c(o[h],l[h]));e.push(f)}return e})}async fit(t,e,s={}){if(this.isTraining)throw new Error("Cannot start training because another fit() call is ongoing.");this.isTraining=!0;let r,o,i,a,l,u,c,h,f;try{const d=s.batchSize==null?32:s.batchSize;Di(d);const g=await this.standardizeUserData(t,e,s.sampleWeight,s.classWeight,!1,d);r=g[0],o=g[1],f=g[2];let m=!1,b;if(s.validationData!=null&&s.validationData.length>0){if(m=!0,s.validationData.length===2)l=s.validationData[0],u=s.validationData[1];else throw s.validationData.length===3?new Z("validationData including sample weights is not supported yet."):new I(`When passing validation data, it must contain 2 (valX, valY) or 3 (valX, valY, valSampleWeight) items; ${s.validationData} is invalid.`);const P=await this.standardizeUserData(l,u,null,null,!0,d);c=P[0],h=P[1],b=c.concat(h)}else if(s.validationSplit!=null&&s.validationSplit>0&&s.validationSplit<1){m=!0;const T=Math.floor(r[0].shape[0]*(1-s.validationSplit)),P=r[0].shape[0];c=Ns(r,T,P),i=r,r=Ns(r,0,T),h=Ns(o,T,P),a=o,o=Ns(o,0,T),b=c.concat(h)}else s.validationSteps!=null&&(m=!0);const y=r.concat(o).concat(f);this.checkTrainableWeightsConsistency();const S=this.makeTrainFunction(),x=this.getDedupedMetricsNames();let $,E;m?(this.makeTestFunction(),$=this.testFunction,E=x.slice().concat(x.map(T=>"val_"+T))):($=null,b=[],E=x.slice());const D=_c(s.callbacks,s.yieldEvery);return await this.fitLoop(S,y,x,d,s.epochs,s.verbose,D,$,b,s.shuffle,E,s.initialEpoch,null,null)}finally{this.isTraining=!1,ge(r,t),ge(o,e),ge(i,t),ge(a,e),ge(c,l),ge(h,u),f!=null&&ut(f)}}async fitLoop(t,e,s,r,o,i,a,l,u,c,h,f,d,p){r==null&&(r=32),o==null&&(o=1),c==null&&(c=!0),f==null&&(f=0);let g=!1;if(l!=null&&u!=null&&(g=!0),p!=null&&(g=!0,d==null))throw new I("Can only use `validationSteps` when doing step-wise training, i.e., `stepsPerEpoch` must be set.");const m=this.checkNumSamples(e,r,d,"steps_per_epoch");let b;m!=null&&(b=br(0,m)),i==null&&(i=1);const{callbackList:y,history:S}=Tc(a,i,o,f,m,d,r,g,h);y.setModel(this),this.history=S,await y.onTrainBegin(),this.stopTraining_=!1;for(let x=f;x<o;++x){await y.onEpochBegin(x);const $={};if(d!=null)throw new Z("stepsPerEpoch mode is not implemented yet.");{if(c==="batch")throw new Z("batch shuffling is not implemneted yet");c&&If(b);const E=Rt(b),D=Pi(m,r);for(let _=0;_<D.length;++_){const T={};if(await y.onBatchBegin(_,T),k(()=>{const P=D[_][0],B=D[_][1],Y=gn(E,P,B-P);T.batch=_,T.size=B-P;const j=Ri(e,Y),F=t(j);for(let G=0;G<s.length;++G){const q=s[G],De=F[G];T[q]=De,Ln(De)}if(_===D.length-1&&g){const G=this.testLoop(l,u,r);for(let q=0;q<s.length;++q){const De=s[q],Xt=G[q];Ln(Xt),$["val_"+De]=Xt}}}),await y.onBatchEnd(_,T),Cc(T),this.stopTraining_)break}E.dispose()}if(await y.onEpochEnd(x,$),this.stopTraining_)break}return await y.onTrainEnd(),await this.history.syncData(),this.history}async fitDataset(t,e){return a1(this,t,e)}async trainOnBatch(t,e){const s=await this.standardizeUserData(t,e),r=s[0],o=s[1],a=this.makeTrainFunction()(r.concat(o)),l=[];for(const u of a){const c=await u.data();l.push(c[0])}return ut(a),ge(s[0],t),ge(s[1],e),Ut(l)}getNamedWeights(t){const e=[],s=t!=null&&t.trainableOnly,r=s?this.trainableWeights:this.weights,o=this.getWeights(s);for(let i=0;i<r.length;++i)s&&!r[i].trainable||e.push({name:r[i].originalName,tensor:o[i]});return e}set stopTraining(t){this.stopTraining_=t}get stopTraining(){return this.stopTraining_}get optimizer(){return this.optimizer_}set optimizer(t){this.optimizer_!==t&&(this.optimizer_=t,this.isOptimizerOwned=!1)}dispose(){const t=super.dispose();if(t.refCountAfterDispose===0&&this.optimizer!=null&&this.isOptimizerOwned){const e=fl().numTensors;this.optimizer_.dispose(),t.numDisposedVariables+=e-fl().numTensors}return t}getLossIdentifiers(){let t;if(typeof this.loss=="string")t=Le(this.loss);else if(Array.isArray(this.loss)){for(const e of this.loss)if(typeof e!="string")throw new Error("Serialization of non-string loss is not supported.");t=this.loss.map(e=>Le(e))}else{const e=Object.keys(this.loss);t={};const s=this.loss;for(const r of e)if(typeof s[r]=="string")t[r]=Le(s[r]);else throw new Error("Serialization of non-string loss is not supported.")}return t}getMetricIdentifiers(){if(typeof this.metrics=="string"||typeof this.metrics=="function")return[Le(Rr(this.metrics))];if(Array.isArray(this.metrics))return this.metrics.map(t=>Le(Rr(t)));{const t={};for(const e in this.metrics)t[e]=Le(Rr(this.metrics[e]));return t}}getTrainingConfig(){return{loss:this.getLossIdentifiers(),metrics:this.getMetricIdentifiers(),optimizer_config:{class_name:this.optimizer.getClassName(),config:this.optimizer.getConfig()}}}loadTrainingConfig(t){if(t.weighted_metrics!=null)throw new Error("Loading weight_metrics is not supported yet.");if(t.loss_weights!=null)throw new Error("Loading loss_weights is not supported yet.");if(t.sample_weight_mode!=null)throw new Error("Loading sample_weight_mode is not supported yet.");const e=Ti(t.optimizer_config),s=Nc(e);let r;if(typeof t.loss=="string")r=pn(t.loss);else if(Array.isArray(t.loss))r=t.loss.map(i=>pn(i));else if(t.loss!=null){r={};for(const i in t.loss)r[i]=pn(t.loss[i])}let o;if(Array.isArray(t.metrics))o=t.metrics.map(i=>pn(i));else if(t.metrics!=null){o={};for(const i in t.metrics)o[i]=pn(t.metrics[i])}this.compile({loss:r,metrics:o,optimizer:s})}async save(t,e){if(typeof t=="string"){const u=nm(t);if(u.length===0)throw new I(`Cannot find any save handlers for URL '${t}'`);if(u.length>1)throw new I(`Found more than one (${u.length}) save handlers for URL '${t}'`);t=u[0]}if(t.save==null)throw new I("LayersModel.save() cannot proceed because the IOHandler provided does not have the `save` attribute defined.");const s=await pl(this.getNamedWeights(e)),a={modelTopology:this.toJSON(null,!1),format:m1,generatedBy:`TensorFlow.js tfjs-layers v${Fc}`,convertedBy:null};if((e==null?!1:e.includeOptimizer)&&this.optimizer!=null){a.trainingConfig=this.getTrainingConfig();const u="optimizer",{data:c,specs:h}=await pl(await this.optimizer.getWeights(),u);s.specs.push(...h),s.data=em([s.data,c])}return this.userDefinedMetadata!=null&&(Oc(this.userDefinedMetadata,this.name,!0),a.userDefinedMetadata=this.userDefinedMetadata),a.weightData=s.data,a.weightSpecs=s.specs,t.save(a)}setUserDefinedMetadata(t){Oc(t,this.name),this.userDefinedMetadata=t}getUserDefinedMetadata(){return this.userDefinedMetadata}}Lr.className="Model",M(Lr);class Qc extends Lr{}Qc.className="Functional",M(Qc);const g1="This is not an object",b1="This is not a Float16Array object",th="This constructor is not a subclass of Float16Array",eh="The constructor property value is not an object",y1="Species constructor didn't return TypedArray object",w1="Derived constructor created TypedArray object which was too small length",Ds="Attempting to access detached ArrayBuffer",Mi="Cannot convert undefined or null to object",Oi="Cannot mix BigInt and other types, use explicit conversions",nh="@@iterator property is not callable",sh="Reduce of empty array with no initial value",x1="The comparison function must be either a function or undefined",Bi="Offset is out of bounds";function ct(n){return(t,...e)=>Vt(n,t,e)}function Xn(n,t){return ct(Jn(n,t).get)}const{apply:Vt,construct:Rs,defineProperty:rh,get:Fi,getOwnPropertyDescriptor:Jn,getPrototypeOf:Ps,has:zi,ownKeys:oh,set:ih,setPrototypeOf:ah}=Reflect,S1=Proxy,{EPSILON:$1,MAX_SAFE_INTEGER:lh,isFinite:uh,isNaN:Zn}=Number,{iterator:_e,species:v1,toStringTag:Ui,for:I1}=Symbol,Qn=Object,{create:Mr,defineProperty:Ls,freeze:A1,is:ch}=Qn,Wi=Qn.prototype,E1=Wi.__lookupGetter__?ct(Wi.__lookupGetter__):(n,t)=>{if(n==null)throw ht(Mi);let e=Qn(n);do{const s=Jn(e,t);if(s!==void 0)return Oe(s,"get")?s.get:void 0}while((e=Ps(e))!==null)},Oe=Qn.hasOwn||ct(Wi.hasOwnProperty),hh=Array,fh=hh.isArray,Or=hh.prototype,C1=ct(Or.join),k1=ct(Or.push),_1=ct(Or.toLocaleString),Gi=Or[_e],T1=ct(Gi),{abs:N1,trunc:dh}=Math,Br=ArrayBuffer,D1=Br.isView,ph=Br.prototype,R1=ct(ph.slice),P1=Xn(ph,"byteLength"),Vi=typeof SharedArrayBuffer<"u"?SharedArrayBuffer:null,L1=Vi&&Xn(Vi.prototype,"byteLength"),qi=Ps(Uint8Array),M1=qi.from,$t=qi.prototype,O1=$t[_e],B1=ct($t.keys),F1=ct($t.values),z1=ct($t.entries),U1=ct($t.set),mh=ct($t.reverse),W1=ct($t.fill),G1=ct($t.copyWithin),gh=ct($t.sort),Ms=ct($t.slice),V1=ct($t.subarray),vt=Xn($t,"buffer"),Sn=Xn($t,"byteOffset"),Q=Xn($t,"length"),bh=Xn($t,Ui),q1=Uint8Array,Kt=Uint16Array,yh=(...n)=>Vt(M1,Kt,n),ji=Uint32Array,j1=Float32Array,$n=Ps([][_e]()),Fr=ct($n.next),H1=ct(function*(){}().next),K1=Ps($n),ht=TypeError,Hi=RangeError,wh=WeakSet,xh=wh.prototype,Y1=ct(xh.add),X1=ct(xh.has),zr=WeakMap,Ki=zr.prototype,Ur=ct(Ki.get),J1=ct(Ki.has),Yi=ct(Ki.set),Sh=new zr,Z1=Mr(null,{next:{value:function(){const t=Ur(Sh,this);return Fr(t)}},[_e]:{value:function(){return this}}});function Wr(n){if(n[_e]===Gi&&$n.next===Fr)return n;const t=Mr(Z1);return Yi(Sh,t,T1(n)),t}const $h=new zr,vh=Mr(K1,{next:{value:function(){const t=Ur($h,this);return H1(t)},writable:!0,configurable:!0}});for(const n of oh($n))n!=="next"&&Ls(vh,n,Jn($n,n));function Ih(n){const t=Mr(vh);return Yi($h,t,n),t}function Gr(n){return n!==null&&typeof n=="object"||typeof n=="function"}function Ah(n){return n!==null&&typeof n=="object"}function Vr(n){return bh(n)!==void 0}function Xi(n){const t=bh(n);return t==="BigInt64Array"||t==="BigUint64Array"}function Q1(n){try{return fh(n)?!1:(P1(n),!0)}catch{return!1}}function Eh(n){if(Vi===null)return!1;try{return L1(n),!0}catch{return!1}}function tx(n){return Q1(n)||Eh(n)}function Ch(n){return fh(n)?n[_e]===Gi&&$n.next===Fr:!1}function ex(n){return Vr(n)?n[_e]===O1&&$n.next===Fr:!1}function qr(n){if(typeof n!="string")return!1;const t=+n;return n!==t+""||!uh(t)?!1:t===dh(t)}const jr=I1("__Float16Array__");function nx(n){if(!Ah(n))return!1;const t=Ps(n);if(!Ah(t))return!1;const e=t.constructor;if(e===void 0)return!1;if(!Gr(e))throw ht(eh);return zi(e,jr)}const Ji=1/$1;function sx(n){return n+Ji-Ji}const kh=6103515625e-14,rx=65504,_h=.0009765625,Th=_h*kh,ox=_h*Ji;function ix(n){const t=+n;if(!uh(t)||t===0)return t;const e=t>0?1:-1,s=N1(t);if(s<kh)return e*sx(s/Th)*Th;const r=(1+ox)*s,o=r-(r-s);return o>rx||Zn(o)?e*(1/0):e*o}const Nh=new Br(4),Dh=new j1(Nh),Rh=new ji(Nh),be=new Kt(512),ye=new q1(512);for(let n=0;n<256;++n){const t=n-127;t<-24?(be[n]=0,be[n|256]=32768,ye[n]=24,ye[n|256]=24):t<-14?(be[n]=1024>>-t-14,be[n|256]=1024>>-t-14|32768,ye[n]=-t-1,ye[n|256]=-t-1):t<=15?(be[n]=t+15<<10,be[n|256]=t+15<<10|32768,ye[n]=13,ye[n|256]=13):t<128?(be[n]=31744,be[n|256]=64512,ye[n]=24,ye[n|256]=24):(be[n]=31744,be[n|256]=64512,ye[n]=13,ye[n|256]=13)}function Te(n){Dh[0]=ix(n);const t=Rh[0],e=t>>23&511;return be[e]+((t&8388607)>>ye[e])}const Zi=new ji(2048);for(let n=1;n<1024;++n){let t=n<<13,e=0;for(;!(t&8388608);)t<<=1,e-=8388608;t&=-8388609,e+=947912704,Zi[n]=t|e}for(let n=1024;n<2048;++n)Zi[n]=939524096+(n-1024<<13);const ts=new ji(64);for(let n=1;n<31;++n)ts[n]=n<<23;ts[31]=1199570944,ts[32]=2147483648;for(let n=33;n<63;++n)ts[n]=2147483648+(n-32<<23);ts[63]=3347054592;const Ph=new Kt(64);for(let n=1;n<64;++n)n!==32&&(Ph[n]=1024);function st(n){const t=n>>10;return Rh[0]=Zi[Ph[t]+(n&1023)]+ts[t],Dh[0]}function Be(n){const t=+n;return Zn(t)||t===0?0:dh(t)}function Qi(n){const t=Be(n);return t<0?0:t<lh?t:lh}function Hr(n,t){if(!Gr(n))throw ht(g1);const e=n.constructor;if(e===void 0)return t;if(!Gr(e))throw ht(eh);const s=e[v1];return s??t}function Os(n){if(Eh(n))return!1;try{return R1(n,0,0),!1}catch{}return!0}function Lh(n,t){const e=Zn(n),s=Zn(t);if(e&&s)return 0;if(e)return 1;if(s||n<t)return-1;if(n>t)return 1;if(n===0&&t===0){const r=ch(n,0),o=ch(t,0);if(!r&&o)return-1;if(r&&!o)return 1}return 0}const ta=2,Kr=new zr;function es(n){return J1(Kr,n)||!D1(n)&&nx(n)}function tt(n){if(!es(n))throw ht(b1)}function Yr(n,t){const e=es(n),s=Vr(n);if(!e&&!s)throw ht(y1);if(typeof t=="number"){let r;if(e){const o=V(n);r=Q(o)}else r=Q(n);if(r<t)throw ht(w1)}if(Xi(n))throw ht(Oi)}function V(n){const t=Ur(Kr,n);if(t!==void 0){const r=vt(t);if(Os(r))throw ht(Ds);return t}const e=n.buffer;if(Os(e))throw ht(Ds);const s=Rs(ft,[e,n.byteOffset,n.length],n.constructor);return Ur(Kr,s)}function Mh(n){const t=Q(n),e=[];for(let s=0;s<t;++s)e[s]=st(n[s]);return e}const Oh=new wh;for(const n of oh($t)){if(n===Ui)continue;const t=Jn($t,n);Oe(t,"get")&&typeof t.get=="function"&&Y1(Oh,t.get)}const ax=A1({get(n,t,e){return qr(t)&&Oe(n,t)?st(Fi(n,t)):X1(Oh,E1(n,t))?Fi(n,t):Fi(n,t,e)},set(n,t,e,s){return qr(t)&&Oe(n,t)?ih(n,t,Te(e)):ih(n,t,e,s)},getOwnPropertyDescriptor(n,t){if(qr(t)&&Oe(n,t)){const e=Jn(n,t);return e.value=st(e.value),e}return Jn(n,t)},defineProperty(n,t,e){return qr(t)&&Oe(n,t)&&Oe(e,"value")&&(e.value=Te(e.value)),rh(n,t,e)}});class ft{constructor(t,e,s){let r;if(es(t))r=Rs(Kt,[V(t)],new.target);else if(Gr(t)&&!tx(t)){let i,a;if(Vr(t)){i=t,a=Q(t);const l=vt(t);if(Os(l))throw ht(Ds);if(Xi(t))throw ht(Oi);const u=new Br(a*ta);r=Rs(Kt,[u],new.target)}else{const l=t[_e];if(l!=null&&typeof l!="function")throw ht(nh);l!=null?Ch(t)?(i=t,a=t.length):(i=[...t],a=i.length):(i=t,a=Qi(i.length)),r=Rs(Kt,[a],new.target)}for(let l=0;l<a;++l)r[l]=Te(i[l])}else r=Rs(Kt,arguments,new.target);const o=new S1(r,ax);return Yi(Kr,o,r),o}static from(t,...e){const s=this;if(!zi(s,jr))throw ht(th);if(s===ft){if(es(t)&&e.length===0){const c=V(t),h=new Kt(vt(c),Sn(c),Q(c));return new ft(vt(Ms(h)))}if(e.length===0)return new ft(vt(yh(t,Te)));const l=e[0],u=e[1];return new ft(vt(yh(t,function(c,...h){return Te(Vt(l,this,[c,...Wr(h)]))},u)))}let r,o;const i=t[_e];if(i!=null&&typeof i!="function")throw ht(nh);if(i!=null)Ch(t)?(r=t,o=t.length):ex(t)?(r=t,o=Q(t)):(r=[...t],o=r.length);else{if(t==null)throw ht(Mi);r=Qn(t),o=Qi(r.length)}const a=new s(o);if(e.length===0)for(let l=0;l<o;++l)a[l]=r[l];else{const l=e[0],u=e[1];for(let c=0;c<o;++c)a[c]=Vt(l,u,[r[c],c])}return a}static of(...t){const e=this;if(!zi(e,jr))throw ht(th);const s=t.length;if(e===ft){const o=new ft(s),i=V(o);for(let a=0;a<s;++a)i[a]=Te(t[a]);return o}const r=new e(s);for(let o=0;o<s;++o)r[o]=t[o];return r}keys(){tt(this);const t=V(this);return B1(t)}values(){tt(this);const t=V(this);return Ih(function*(){for(const e of F1(t))yield st(e)}())}entries(){tt(this);const t=V(this);return Ih(function*(){for(const[e,s]of z1(t))yield[e,st(s)]}())}at(t){tt(this);const e=V(this),s=Q(e),r=Be(t),o=r>=0?r:s+r;if(!(o<0||o>=s))return st(e[o])}with(t,e){tt(this);const s=V(this),r=Q(s),o=Be(t),i=o>=0?o:r+o,a=+e;if(i<0||i>=r)throw Hi(Bi);const l=new Kt(vt(s),Sn(s),Q(s)),u=new ft(vt(Ms(l))),c=V(u);return c[i]=Te(a),u}map(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0],i=Hr(s,ft);if(i===ft){const l=new ft(r),u=V(l);for(let c=0;c<r;++c){const h=st(s[c]);u[c]=Te(Vt(t,o,[h,c,this]))}return l}const a=new i(r);Yr(a,r);for(let l=0;l<r;++l){const u=st(s[l]);a[l]=Vt(t,o,[u,l,this])}return a}filter(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0],i=[];for(let u=0;u<r;++u){const c=st(s[u]);Vt(t,o,[c,u,this])&&k1(i,c)}const a=Hr(s,ft),l=new a(i);return Yr(l),l}reduce(t,...e){tt(this);const s=V(this),r=Q(s);if(r===0&&e.length===0)throw ht(sh);let o,i;e.length===0?(o=st(s[0]),i=1):(o=e[0],i=0);for(let a=i;a<r;++a)o=t(o,st(s[a]),a,this);return o}reduceRight(t,...e){tt(this);const s=V(this),r=Q(s);if(r===0&&e.length===0)throw ht(sh);let o,i;e.length===0?(o=st(s[r-1]),i=r-2):(o=e[0],i=r-1);for(let a=i;a>=0;--a)o=t(o,st(s[a]),a,this);return o}forEach(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i)Vt(t,o,[st(s[i]),i,this])}find(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i){const a=st(s[i]);if(Vt(t,o,[a,i,this]))return a}}findIndex(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i){const a=st(s[i]);if(Vt(t,o,[a,i,this]))return i}return-1}findLast(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=r-1;i>=0;--i){const a=st(s[i]);if(Vt(t,o,[a,i,this]))return a}}findLastIndex(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=r-1;i>=0;--i){const a=st(s[i]);if(Vt(t,o,[a,i,this]))return i}return-1}every(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i)if(!Vt(t,o,[st(s[i]),i,this]))return!1;return!0}some(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i)if(Vt(t,o,[st(s[i]),i,this]))return!0;return!1}set(t,...e){tt(this);const s=V(this),r=Be(e[0]);if(r<0)throw Hi(Bi);if(t==null)throw ht(Mi);if(Xi(t))throw ht(Oi);if(es(t))return U1(V(this),V(t),r);if(Vr(t)){const l=vt(t);if(Os(l))throw ht(Ds)}const o=Q(s),i=Qn(t),a=Qi(i.length);if(r===1/0||a+r>o)throw Hi(Bi);for(let l=0;l<a;++l)s[l+r]=Te(i[l])}reverse(){tt(this);const t=V(this);return mh(t),this}toReversed(){tt(this);const t=V(this),e=new Kt(vt(t),Sn(t),Q(t)),s=new ft(vt(Ms(e))),r=V(s);return mh(r),s}fill(t,...e){tt(this);const s=V(this);return W1(s,Te(t),...Wr(e)),this}copyWithin(t,e,...s){tt(this);const r=V(this);return G1(r,t,e,...Wr(s)),this}sort(t){tt(this);const e=V(this),s=t!==void 0?t:Lh;return gh(e,(r,o)=>s(st(r),st(o))),this}toSorted(t){tt(this);const e=V(this);if(t!==void 0&&typeof t!="function")throw new ht(x1);const s=t!==void 0?t:Lh,r=new Kt(vt(e),Sn(e),Q(e)),o=new ft(vt(Ms(r))),i=V(o);return gh(i,(a,l)=>s(st(a),st(l))),o}slice(t,e){tt(this);const s=V(this),r=Hr(s,ft);if(r===ft){const p=new Kt(vt(s),Sn(s),Q(s));return new ft(vt(Ms(p,t,e)))}const o=Q(s),i=Be(t),a=e===void 0?o:Be(e);let l;i===-1/0?l=0:i<0?l=o+i>0?o+i:0:l=o<i?o:i;let u;a===-1/0?u=0:a<0?u=o+a>0?o+a:0:u=o<a?o:a;const c=u-l>0?u-l:0,h=new r(c);if(Yr(h,c),c===0)return h;const f=vt(s);if(Os(f))throw ht(Ds);let d=0;for(;l<u;)h[d]=st(s[l]),++l,++d;return h}subarray(t,e){tt(this);const s=V(this),r=Hr(s,ft),o=new Kt(vt(s),Sn(s),Q(s)),i=V1(o,t,e),a=new r(vt(i),Sn(i),Q(i));return Yr(a),a}indexOf(t,...e){tt(this);const s=V(this),r=Q(s);let o=Be(e[0]);if(o===1/0)return-1;o<0&&(o+=r,o<0&&(o=0));for(let i=o;i<r;++i)if(Oe(s,i)&&st(s[i])===t)return i;return-1}lastIndexOf(t,...e){tt(this);const s=V(this),r=Q(s);let o=e.length>=1?Be(e[0]):r-1;if(o===-1/0)return-1;o>=0?o=o<r-1?o:r-1:o+=r;for(let i=o;i>=0;--i)if(Oe(s,i)&&st(s[i])===t)return i;return-1}includes(t,...e){tt(this);const s=V(this),r=Q(s);let o=Be(e[0]);if(o===1/0)return!1;o<0&&(o+=r,o<0&&(o=0));const i=Zn(t);for(let a=o;a<r;++a){const l=st(s[a]);if(i&&Zn(l)||l===t)return!0}return!1}join(t){tt(this);const e=V(this),s=Mh(e);return C1(s,t)}toLocaleString(...t){tt(this);const e=V(this),s=Mh(e);return _1(s,...Wr(t))}get[Ui](){if(es(this))return"Float16Array"}}Ls(ft,"BYTES_PER_ELEMENT",{value:ta}),Ls(ft,jr,{}),ah(ft,qi);const Xr=ft.prototype;Ls(Xr,"BYTES_PER_ELEMENT",{value:ta}),Ls(Xr,_e,{value:Xr.values,writable:!0,configurable:!0}),ah(Xr,$t);function lx(n,t){return n.channels===t.channels}const Jr=8;class Bh{constructor(t,e,s){K(this,"_label");K(this,"_device");K(this,"_outputBuffers",{});K(this,"_pipeline");K(this,"_bindGroups",[]);K(this,"_needsUpdatePipeline",!0);K(this,"_inputs",[]);K(this,"_outputs",[]);K(this,"_uniforms",[]);K(this,"_uniformBuffers",{});K(this,"_width",10);K(this,"_height",10);K(this,"_execWidth");K(this,"_execHeight");K(this,"_csCode","");K(this,"_csMain");K(this,"_csDefine");this._label=t,this._device=e,this._csMain=s.csMain,this._csDefine=s.csDefine,this._inputs=s.inputs,this._outputs=s.outputs,this._uniforms=s.uniforms,s.uniforms.forEach(r=>{this._uniformBuffers[r.label]=e.createBuffer({label:this._label,size:r.data.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST}),this._device.queue.writeBuffer(this._uniformBuffers[r.label],0,r.data)})}setSize(t,e){t=Math.ceil(t),e=Math.ceil(e);const s=t!==this._width||e!==this._height;this._width=t,this._height=e,s&&(this._resizeOutputBuffers(),this._needsUpdatePipeline=!0)}setExecuteSize(t,e){t=Math.ceil(t),e=Math.ceil(e),this._execWidth=t,this._execHeight=e}setOutputParams(t){this._updateOutputBuffers(t),this._needsUpdatePipeline=!0}setUniform(t,e){const s=this._uniformBuffers[t];this._device.queue.writeBuffer(s,0,e)}getOutputBuffer(t){return this._outputBuffers[t].buffer}dispose(){Object.keys(this._uniformBuffers).forEach(t=>{this._uniformBuffers[t].destroy()}),Object.keys(this._outputBuffers).forEach(t=>{this._outputBuffers[t].texture.destroy()})}_createBuffer(t){const e=this._width*this._height*4*4;return this._device.createBuffer({label:this._label,size:Math.max(e,80),usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST|GPUBufferUsage.COPY_SRC})}_resizeOutputBuffers(){const t=this._outputBuffers;for(const e in t){const{buffer:s,params:r}=t[e];s.destroy(),t[e].buffer=this._createBuffer(r)}}_updateOutputBuffers(t){var s,r;const e=this._outputBuffers;for(const o in t){const i=t[o];if(!lx(i,((s=e[o])==null?void 0:s.params)||{})){(r=e[o])==null||r.buffer.destroy();const a=this._createBuffer(i);e[o]={buffer:a,params:i}}}}_updatePipeline(t){if(!this._needsUpdatePipeline)return;this._needsUpdatePipeline=!1;const e=this._device,s=this._getFullCs(t);s!==this._csCode&&(this._csCode=s,this._pipeline=e.createComputePipeline({label:this._label,layout:"auto",compute:{module:e.createShaderModule({label:this._label,code:s}),entryPoint:"main"}}),this._updateBindGroups())}_getFullCs(t){const e=this._inputs,s=e.length>0;return`
+ */function h1(n){return n instanceof At}function Li(n){return Array.isArray(n)}function Xc(n){return!h1(n)&&!Li(n)}function Jc(n,t,e,s=!0,r=""){if(t==null||t.length===0){if(n!=null){let i=!1;if(Li(n)&&n.length>0)i=!0;else if(Xc(n)){for(const a in n)if(n.hasOwnProperty(a)){i=!0;break}}else i=!0;if(i)throw new I(`Error when checking model ${r} expected no data, but got ${n}`)}return[]}if(n==null)return t.map(i=>null);let o;if(Xc(n)){n=n,o=[];for(const i of t){if(n[i]==null)throw new I(`No data provided for "${i}". Need data for each key in: ${t}`);o.push(n[i])}}else if(Li(n)){if(n=n,n.length!==t.length)throw new I(`Error when checking model ${r}: the Array of Tensors that you are passing to your model is not the size the model expected. Expected to see ${t.length} Tensor(s), but instead got the following list of Tensor(s): ${n}`);o=n}else{if(n=n,t.length>1)throw new I(`The model ${r} expects ${t.length} Tensor(s), but only received one Tensor. Found: Tensor with shape ${n.shape}`);o=[n]}if(o=Yc(o),e!=null)for(let i=0;i<t.length;++i){if(e[i]==null)continue;const a=o[i];if(a.shape.length!==e[i].length)throw new I(`Error when checking ${r}: expected ${t[i]} to have ${e[i].length} dimension(s). but got array with shape ${a.shape}`);for(let l=0;l<e[i].length;++l){if(l===0&&!s)continue;const u=a.shape[l],c=e[i][l];if(c!=null&&c>=0&&u!==c)throw new I(`${r} expected a batch of elements where each example has shape [${e[i].slice(1,e[i].length)}] (i.e.,tensor shape [*,${e[i].slice(1,e[i].length)}]) but the ${r} received an input with ${a.shape[0]} examples, each with shape [${a.shape.slice(1,a.shape.length)}] (tensor shape [${a.shape}])`)}}return o}function f1(n,t,e){const s=gn(n.map(o=>o.shape[0]));s.sort();const r=gn(t.map(o=>o.shape[0]));if(r.sort(),s.length>1)throw new I(`All input Tensors (x) should have the same number of samples. Got array shapes: ${JSON.stringify(n.map(o=>o.shape))}`);if(r.length>1)throw new I(`All target Tensors (y) should have the same number of samples. Got array shapes: ${JSON.stringify(t.map(o=>o.shape))}`);if(s.length>0&&r.length>0&&!Zt(s,r))throw new I(`Input Tensors should have the same number of samples as target Tensors. Found ${s[0]} input sample(s) and ${r[0]} target sample(s).`)}function d1(n,t,e){const s=[_r,Nr,Cs];for(let r=0;r<n.length;++r){const o=n[r],i=t[r],a=e[r];if(i!=null){if(i===Cs&&o.shape[o.shape.length-1]===1)throw new I(`You are passing a target array of shape ${o.shape} while using a loss 'categorical_crossentropy'. 'categorical_crossentropy'expects targets to be binary matrices (1s and 0s) of shape [samples, classes].`);if(s.indexOf(i)!==-1){const l=o.shape.slice(1),u=a.slice(1);for(let c=0;c<l.length;++c){const h=l[c],f=u[c];if(f!=null&&h!==f)throw new I(`A target Tensor with shape ${o.shape} was passed for an output of shape ${a}, while using a loss function that expects targets to have the same shape as the output.`)}}}}}function Zc(n,t,e,s=!0,r=""){let o;if(Array.isArray(n)){if(n.length!==t.length)throw new I(`Error when checking model ${r}: the Array of Tensors that you are passing to your model is not the size the the model expected. Expected to see ${t.length} Tensor(s), but instead got ${n.length} Tensors(s).`);o=n}else{if(t.length>1)throw new I(`The model expects ${t.length} ${r} Tensors, but only received one Tensor. Found: array with shape ${JSON.stringify(n.shape)}.`);o=[n]}if(e!=null)for(let i=0;i<t.length;++i){if(e[i]==null)continue;const a=o[i];if(a.shape.length!==e[i].length)throw new I(`Error when checking ${r}: expected ${t[i]} to have ${e[i].length} dimension(s), but got array with shape ${JSON.stringify(a.shape)}`);for(let l=0;l<e[i].length;++l){if(l===0&&!s)continue;const u=a.shape[l],c=e[i][l];if(c!=null&&c!==u)throw new I(`Error when checking ${r}: expected ${t[i]} to have shape ${JSON.stringify(e[i])} but got array with shape ${JSON.stringify(a.shape)}.`)}}}function p1(n,t){if(n==null||Array.isArray(n)&&n.length===0)return t.map(s=>[]);let e;if(typeof n=="string"||typeof n=="function")e=[n];else if(Array.isArray(n)||typeof n=="object")e=n;else throw new TypeError(`Type of metrics argument not understood. Expected an string,function, Array, or Object, found: ${n}`);if(Array.isArray(e))return t.map(s=>e);{const s=[];for(const r of t){let o=e.hasOwnProperty(r)?e[r]:[];Array.isArray(o)||(o=[o]),s.push(o)}return s}}const m1="layers-model";class Mr extends me{constructor(t){super(t),this.isTraining=!1}summary(t,e,s=console.log){if(!this.built)throw new I("This model has never been called, thus its weights have not been created yet. So no summary can be displayed. Build the model first (e.g., by calling it on some test data).");jw(this,t,e,s)}compile(t){if(t.loss==null&&(t.loss=[]),this.loss=t.loss,typeof t.optimizer=="string")this.optimizer_=qw(t.optimizer),this.isOptimizerOwned=!0;else{if(!(t.optimizer instanceof qe))throw new I("User-defined optimizer must be an instance of tf.Optimizer.");this.optimizer_=t.optimizer,this.isOptimizerOwned=!1}let e=[];if(!Array.isArray(t.loss)&&typeof t.loss!="string"&&typeof t.loss!="function"){t.loss=t.loss;for(const i in t.loss)if(this.outputNames.indexOf(i)===-1)throw new I(`Unknown entry in loss dictionary: "${i}". Only expected the following keys: ${this.outputNames}`);for(const i of this.outputNames)t.loss[i]==null&&console.warn(`Output "${i}" is missing from loss dictionary. We assume this was done on purpose, and we will not be expecting data to be passed to ${i} during training`),e.push(Ci(t.loss[i]))}else if(Array.isArray(t.loss)){if(t.loss.length!==this.outputs.length)throw new I(`When passing an Array as loss, it should have one entry per model output. The model has ${this.outputs.length} output(s), but you passed loss=${t.loss}.`);e=t.loss.map(a=>Ci(a))}else{const i=Ci(t.loss);this.outputs.forEach(a=>{e.push(i)})}this.lossFunctions=e,this.feedOutputNames=[],this.feedOutputShapes=[],this.feedLossFns=[];for(let i=0;i<this.outputs.length;++i){const a=this.internalOutputShapes[i],l=this.outputNames[i];this.feedOutputNames.push(l),this.feedOutputShapes.push(a),this.feedLossFns.push(this.lossFunctions[i])}const s=[];this.metrics=t.metrics,this.metricsNames=["loss"],this.metricsTensors=[],br("loss",()=>{for(let i=0;i<this.outputs.length;++i){if(s.indexOf(i)!==-1)continue;const a=this.lossFunctions[i];this.outputs.length>1&&(this.metricsTensors.push([a,i]),this.metricsNames.push(this.outputNames[i]+"_loss"))}});const r=p1(t.metrics,this.outputNames),o=(i,a,l)=>{this.outputNames.length>1&&(a=this.outputNames[i]+"_"+a),this.metricsNames.push(a),this.metricsTensors.push([l,i])};br("metric",()=>{for(let i=0;i<this.outputs.length;++i){if(s.indexOf(i)!==-1)continue;const a=r[i];(u=>{const c="";let h,f,d;for(const p of u){if(typeof p=="string"&&["accuracy","acc","crossentropy","ce"].indexOf(p)!==-1){const m=this.internalOutputShapes[i];m[m.length-1]===1||this.lossFunctions[i]===Nr?["accuracy","acc"].indexOf(p)!==-1?f=Dc:["crossentropy","ce"].indexOf(p)!==-1&&(f=Lw):this.lossFunctions[i]===Tr?["accuracy","acc"].indexOf(p)!==-1?f=Mw:["crossentropy","ce"].indexOf(p)!==-1&&(f=Lc):["accuracy","acc"].indexOf(p)!==-1?f=Rc:["crossentropy","ce"].indexOf(p)!==-1&&(f=Pc);let b;["accuracy","acc"].indexOf(p)!==-1?b="acc":["crossentropy","ce"].indexOf(p)!==-1&&(b="ce"),d=f,h=c+b}else d=Vw(p),h=c+Pr(p);let g;br(h,()=>{g=d}),o(i,h,g)}})(a)}}),this.collectedTrainableWeights=this.trainableWeights}checkTrainableWeightsConsistency(){this.collectedTrainableWeights!=null&&this.trainableWeights.length!==this.collectedTrainableWeights.length&&console.warn("Discrepancy between trainableweights and collected trainable weights. Did you set `model.trainable` without calling `model.compile()` afterwards?")}evaluate(t,e,s={}){const r=s.batchSize==null?32:s.batchSize;Di(r);const i=this.standardizeUserDataXY(t,e,!0,r);try{const a=i[0].concat(i[1]);this.makeTestFunction();const l=this.testFunction,u=this.testLoop(l,a,r,s.verbose,s.steps);return Ut(u)}finally{ge(i[0],t),ge(i[1],e)}}async evaluateDataset(t,e){return this.makeTestFunction(),c1(this,t,e)}checkNumSamples(t,e,s,r="steps"){let o;if(s!=null){if(o=null,e!=null)throw new I(`If ${r} is set, batchSize must be null or undefined.Got batchSize = ${e}`)}else if(t!=null)Array.isArray(t)?o=t[0].shape[0]:o=t.shape[0];else throw new I(`Either the input data should have a defined shape, or ${r} shoud be specified.`);return o}execute(t,e){if(Array.isArray(e)&&e.length===0)throw new I("`outputs` is an empty Array, which is not allowed.");const s=Array.isArray(e),r=s?e:[e],o=this.retrieveSymbolicTensors(r),i=new Ke;if(t instanceof At&&(t=[t]),Array.isArray(t)){if(t.length!==this.inputs.length)throw new I(`The number of inputs provided (${t.length}) does not match the number of inputs of this model (${this.inputs.length}).`);for(let l=0;l<this.inputs.length;++l)i.add(this.inputs[l],t[l])}else for(const l of this.inputs){const u=t[l.name];if(u==null)throw new I(`No value is provided for the model's input ${l.name}`);i.add(l,u)}const a=Ns(o,i);return s?a:a[0]}retrieveSymbolicTensors(t){const e=mr(null,t.length);let s=t.length;for(const r of this.layers){const o=Array.isArray(r.output)?r.output:[r.output],i=o.map(a=>a.name);for(let a=0;a<t.length;++a){const l=i.indexOf(t[a]);if(l!==-1&&(e[a]=o[l],s--),s===0)break}if(s===0)break}if(s>0){const r=[];throw e.forEach((o,i)=>{o==null&&r.push(t[i])}),new I(`Cannot find SymbolicTensors for output name(s): ${JSON.stringify(r)}`)}return e}predictLoop(t,e=32,s=!1){return k(()=>{const r=this.checkNumSamples(t);if(s)throw new Z("Verbose predictLoop() is not implemented yet.");const o=Pi(r,e),i=this.outputs.map(a=>[]);for(let a=0;a<o.length;++a)k(()=>{const u=o[a][0],c=o[a][1],h=Ds(t,u,c),f=[];if(Array.isArray(h))for(let p=0;p<h.length;++p)f.push({key:this.inputs[p],value:h[p]});else f.push({key:this.inputs[0],value:h});const d=new Ke(f);return Ns(this.outputs,d)}).forEach((u,c)=>i[c].push(u));return Ut(i.map(a=>an(a,0)))})}predict(t,e={}){const s=Yc(t);Zc(s,this.inputNames,this.feedInputShapes,!1);try{const r=e.batchSize==null?32:e.batchSize;return Di(r),this.predictLoop(s,r)}finally{ge(s,t)}}predictOnBatch(t){Zc(t,this.inputNames,this.feedInputShapes,!0);const e=(Array.isArray(t)?t[0]:t).shape[0];return this.predictLoop(t,e)}standardizeUserDataXY(t,e,s=!0,r){if(this.optimizer_==null)throw new He("You must compile a model before training/testing. Use LayersModel.compile(modelCompileArgs).");const o=[];for(let i=0;i<this.feedOutputShapes.length;++i){const a=this.feedOutputShapes[i];this.feedLossFns[i]===Tr?o.push(a.slice(0,a.length-1).concat([1])):o.push(a)}if(t=Jc(t,this.feedInputNames,this.feedInputShapes,!1,"input"),e=Jc(e,this.feedOutputNames,o,!1,"target"),f1(t,e),d1(e,this.feedLossFns,this.feedOutputShapes),this.stateful&&r!=null&&r>0&&t[0].shape[0]%r!==0)throw new I(`In a stateful network, you should only pass inputs with a number of samples that is divisible by the batch size ${r}. Found: ${t[0].shape[0]} sample(s).`);return[t,e]}async standardizeUserData(t,e,s,r,o=!0,i){const[a,l]=this.standardizeUserDataXY(t,e,o,i);if(s!=null)throw new Error("sample weight is not supported yet.");let u=null;if(r!=null){const c=Vc(r,this.outputNames);u=[];for(let h=0;h<c.length;++h)u.push(await qc(l[h],null,c[h]))}return[a,l,u]}testLoop(t,e,s,r=0,o){return k(()=>{const i=this.checkNumSamples(e,s,o,"steps"),a=[];if(r>0)throw new Z("Verbose mode is not implemented yet.");if(o!=null)throw new Z("steps mode in testLoop() is not implemented yet");{const l=Pi(i,s),u=Rt(yr(0,i));for(let c=0;c<l.length;++c){const h=l[c][0],f=l[c][1],d=bn(u,h,f-h),p=Ri(e,d),g=t(p);if(c===0)for(let m=0;m<g.length;++m)a.push(Yt(0));for(let m=0;m<g.length;++m){const b=g[m];a[m]=O(a[m],N(f-h,b))}}for(let c=0;c<a.length;++c)a[c]=X(a[c],i)}return a})}getDedupedMetricsNames(){const t=this.metricsNames,e=[];for(let s=0;s<t.length;++s){const r=t[s];let o=r;if(eu(t,r)>1){const i=eu(t.slice(0,s),r);o+=`_${i}`}e.push(o)}return e}makeTrainFunction(){return t=>{const e=[],s=t.slice(0,this.inputs.length),r=t.slice(this.inputs.length,this.inputs.length+this.outputs.length),o=t.slice(this.inputs.length+this.outputs.length,this.inputs.length+this.outputs.length*2),i=[],a=()=>{const h=[];for(let g=0;g<this.inputs.length;++g)h.push({key:this.inputs[g],value:s[g]});const f=new Ke(h),d=Ns(this.outputs,f,{training:!0});let p;for(let g=0;g<this.lossFunctions.length;++g){const m=this.lossFunctions[g];let b=m(r[g],d[g]);o[g]!=null&&(b=r1(b,o[g]));const y=St(b);e.push(y),g===0?p=b:p=O(p,b)}for(let g=0;g<this.metricsTensors.length;++g){let m;if(this.outputs.length>1&&g<this.outputs.length)m=e[g];else{const b=this.metricsTensors[g][0],y=this.metricsTensors[g][1];m=St(b(r[y],d[y]))}Ln(m),i.push(m)}return p=St(p),this.calculateLosses().forEach(g=>{p=O(p,g)}),p},l=this.collectedTrainableWeights.map(h=>h.read());return[this.optimizer_.minimize(a,!0,l)].concat(i)}}makeTestFunction(){this.testFunction=t=>k(()=>{const e=[];let s;const r=t.slice(0,this.inputs.length),o=t.slice(this.inputs.length,this.inputs.length+this.outputs.length),i=[];for(let u=0;u<this.inputs.length;++u)i.push({key:this.inputs[u],value:r[u]});const a=new Ke(i),l=Ns(this.outputs,a);for(let u=0;u<this.lossFunctions.length;++u){const c=this.lossFunctions[u],h=St(c(o[u],l[u]));u===0?s=h:s=O(s,h),e.push(s)}for(let u=0;u<this.metricsTensors.length;++u){const c=this.metricsTensors[u][0],h=this.metricsTensors[u][1],f=St(c(o[h],l[h]));e.push(f)}return e})}async fit(t,e,s={}){if(this.isTraining)throw new Error("Cannot start training because another fit() call is ongoing.");this.isTraining=!0;let r,o,i,a,l,u,c,h,f;try{const d=s.batchSize==null?32:s.batchSize;Di(d);const g=await this.standardizeUserData(t,e,s.sampleWeight,s.classWeight,!1,d);r=g[0],o=g[1],f=g[2];let m=!1,b;if(s.validationData!=null&&s.validationData.length>0){if(m=!0,s.validationData.length===2)l=s.validationData[0],u=s.validationData[1];else throw s.validationData.length===3?new Z("validationData including sample weights is not supported yet."):new I(`When passing validation data, it must contain 2 (valX, valY) or 3 (valX, valY, valSampleWeight) items; ${s.validationData} is invalid.`);const P=await this.standardizeUserData(l,u,null,null,!0,d);c=P[0],h=P[1],b=c.concat(h)}else if(s.validationSplit!=null&&s.validationSplit>0&&s.validationSplit<1){m=!0;const T=Math.floor(r[0].shape[0]*(1-s.validationSplit)),P=r[0].shape[0];c=Ds(r,T,P),i=r,r=Ds(r,0,T),h=Ds(o,T,P),a=o,o=Ds(o,0,T),b=c.concat(h)}else s.validationSteps!=null&&(m=!0);const y=r.concat(o).concat(f);this.checkTrainableWeightsConsistency();const S=this.makeTrainFunction(),x=this.getDedupedMetricsNames();let $,E;m?(this.makeTestFunction(),$=this.testFunction,E=x.slice().concat(x.map(T=>"val_"+T))):($=null,b=[],E=x.slice());const D=_c(s.callbacks,s.yieldEvery);return await this.fitLoop(S,y,x,d,s.epochs,s.verbose,D,$,b,s.shuffle,E,s.initialEpoch,null,null)}finally{this.isTraining=!1,ge(r,t),ge(o,e),ge(i,t),ge(a,e),ge(c,l),ge(h,u),f!=null&&ut(f)}}async fitLoop(t,e,s,r,o,i,a,l,u,c,h,f,d,p){r==null&&(r=32),o==null&&(o=1),c==null&&(c=!0),f==null&&(f=0);let g=!1;if(l!=null&&u!=null&&(g=!0),p!=null&&(g=!0,d==null))throw new I("Can only use `validationSteps` when doing step-wise training, i.e., `stepsPerEpoch` must be set.");const m=this.checkNumSamples(e,r,d,"steps_per_epoch");let b;m!=null&&(b=yr(0,m)),i==null&&(i=1);const{callbackList:y,history:S}=Tc(a,i,o,f,m,d,r,g,h);y.setModel(this),this.history=S,await y.onTrainBegin(),this.stopTraining_=!1;for(let x=f;x<o;++x){await y.onEpochBegin(x);const $={};if(d!=null)throw new Z("stepsPerEpoch mode is not implemented yet.");{if(c==="batch")throw new Z("batch shuffling is not implemneted yet");c&&If(b);const E=Rt(b),D=Pi(m,r);for(let _=0;_<D.length;++_){const T={};if(await y.onBatchBegin(_,T),k(()=>{const P=D[_][0],B=D[_][1],Y=bn(E,P,B-P);T.batch=_,T.size=B-P;const j=Ri(e,Y),F=t(j);for(let G=0;G<s.length;++G){const q=s[G],De=F[G];T[q]=De,Ln(De)}if(_===D.length-1&&g){const G=this.testLoop(l,u,r);for(let q=0;q<s.length;++q){const De=s[q],qt=G[q];Ln(qt),$["val_"+De]=qt}}}),await y.onBatchEnd(_,T),Cc(T),this.stopTraining_)break}E.dispose()}if(await y.onEpochEnd(x,$),this.stopTraining_)break}return await y.onTrainEnd(),await this.history.syncData(),this.history}async fitDataset(t,e){return a1(this,t,e)}async trainOnBatch(t,e){const s=await this.standardizeUserData(t,e),r=s[0],o=s[1],a=this.makeTrainFunction()(r.concat(o)),l=[];for(const u of a){const c=await u.data();l.push(c[0])}return ut(a),ge(s[0],t),ge(s[1],e),Ut(l)}getNamedWeights(t){const e=[],s=t!=null&&t.trainableOnly,r=s?this.trainableWeights:this.weights,o=this.getWeights(s);for(let i=0;i<r.length;++i)s&&!r[i].trainable||e.push({name:r[i].originalName,tensor:o[i]});return e}set stopTraining(t){this.stopTraining_=t}get stopTraining(){return this.stopTraining_}get optimizer(){return this.optimizer_}set optimizer(t){this.optimizer_!==t&&(this.optimizer_=t,this.isOptimizerOwned=!1)}dispose(){const t=super.dispose();if(t.refCountAfterDispose===0&&this.optimizer!=null&&this.isOptimizerOwned){const e=fl().numTensors;this.optimizer_.dispose(),t.numDisposedVariables+=e-fl().numTensors}return t}getLossIdentifiers(){let t;if(typeof this.loss=="string")t=Le(this.loss);else if(Array.isArray(this.loss)){for(const e of this.loss)if(typeof e!="string")throw new Error("Serialization of non-string loss is not supported.");t=this.loss.map(e=>Le(e))}else{const e=Object.keys(this.loss);t={};const s=this.loss;for(const r of e)if(typeof s[r]=="string")t[r]=Le(s[r]);else throw new Error("Serialization of non-string loss is not supported.")}return t}getMetricIdentifiers(){if(typeof this.metrics=="string"||typeof this.metrics=="function")return[Le(Pr(this.metrics))];if(Array.isArray(this.metrics))return this.metrics.map(t=>Le(Pr(t)));{const t={};for(const e in this.metrics)t[e]=Le(Pr(this.metrics[e]));return t}}getTrainingConfig(){return{loss:this.getLossIdentifiers(),metrics:this.getMetricIdentifiers(),optimizer_config:{class_name:this.optimizer.getClassName(),config:this.optimizer.getConfig()}}}loadTrainingConfig(t){if(t.weighted_metrics!=null)throw new Error("Loading weight_metrics is not supported yet.");if(t.loss_weights!=null)throw new Error("Loading loss_weights is not supported yet.");if(t.sample_weight_mode!=null)throw new Error("Loading sample_weight_mode is not supported yet.");const e=Ti(t.optimizer_config),s=Nc(e);let r;if(typeof t.loss=="string")r=mn(t.loss);else if(Array.isArray(t.loss))r=t.loss.map(i=>mn(i));else if(t.loss!=null){r={};for(const i in t.loss)r[i]=mn(t.loss[i])}let o;if(Array.isArray(t.metrics))o=t.metrics.map(i=>mn(i));else if(t.metrics!=null){o={};for(const i in t.metrics)o[i]=mn(t.metrics[i])}this.compile({loss:r,metrics:o,optimizer:s})}async save(t,e){if(typeof t=="string"){const u=nm(t);if(u.length===0)throw new I(`Cannot find any save handlers for URL '${t}'`);if(u.length>1)throw new I(`Found more than one (${u.length}) save handlers for URL '${t}'`);t=u[0]}if(t.save==null)throw new I("LayersModel.save() cannot proceed because the IOHandler provided does not have the `save` attribute defined.");const s=await pl(this.getNamedWeights(e)),a={modelTopology:this.toJSON(null,!1),format:m1,generatedBy:`TensorFlow.js tfjs-layers v${Fc}`,convertedBy:null};if((e==null?!1:e.includeOptimizer)&&this.optimizer!=null){a.trainingConfig=this.getTrainingConfig();const u="optimizer",{data:c,specs:h}=await pl(await this.optimizer.getWeights(),u);s.specs.push(...h),s.data=em([s.data,c])}return this.userDefinedMetadata!=null&&(Oc(this.userDefinedMetadata,this.name,!0),a.userDefinedMetadata=this.userDefinedMetadata),a.weightData=s.data,a.weightSpecs=s.specs,t.save(a)}setUserDefinedMetadata(t){Oc(t,this.name),this.userDefinedMetadata=t}getUserDefinedMetadata(){return this.userDefinedMetadata}}Mr.className="Model",M(Mr);class Qc extends Mr{}Qc.className="Functional",M(Qc);const g1="This is not an object",b1="This is not a Float16Array object",th="This constructor is not a subclass of Float16Array",eh="The constructor property value is not an object",y1="Species constructor didn't return TypedArray object",w1="Derived constructor created TypedArray object which was too small length",Rs="Attempting to access detached ArrayBuffer",Mi="Cannot convert undefined or null to object",Oi="Cannot mix BigInt and other types, use explicit conversions",nh="@@iterator property is not callable",sh="Reduce of empty array with no initial value",x1="The comparison function must be either a function or undefined",Bi="Offset is out of bounds";function ct(n){return(t,...e)=>Vt(n,t,e)}function Xn(n,t){return ct(Jn(n,t).get)}const{apply:Vt,construct:Ps,defineProperty:rh,get:Fi,getOwnPropertyDescriptor:Jn,getPrototypeOf:Ls,has:zi,ownKeys:oh,set:ih,setPrototypeOf:ah}=Reflect,S1=Proxy,{EPSILON:$1,MAX_SAFE_INTEGER:lh,isFinite:uh,isNaN:Zn}=Number,{iterator:_e,species:v1,toStringTag:Ui,for:I1}=Symbol,Qn=Object,{create:Or,defineProperty:Ms,freeze:A1,is:ch}=Qn,Wi=Qn.prototype,E1=Wi.__lookupGetter__?ct(Wi.__lookupGetter__):(n,t)=>{if(n==null)throw ht(Mi);let e=Qn(n);do{const s=Jn(e,t);if(s!==void 0)return Oe(s,"get")?s.get:void 0}while((e=Ls(e))!==null)},Oe=Qn.hasOwn||ct(Wi.hasOwnProperty),hh=Array,fh=hh.isArray,Br=hh.prototype,C1=ct(Br.join),k1=ct(Br.push),_1=ct(Br.toLocaleString),Gi=Br[_e],T1=ct(Gi),{abs:N1,trunc:dh}=Math,Fr=ArrayBuffer,D1=Fr.isView,ph=Fr.prototype,R1=ct(ph.slice),P1=Xn(ph,"byteLength"),Vi=typeof SharedArrayBuffer<"u"?SharedArrayBuffer:null,L1=Vi&&Xn(Vi.prototype,"byteLength"),qi=Ls(Uint8Array),M1=qi.from,$t=qi.prototype,O1=$t[_e],B1=ct($t.keys),F1=ct($t.values),z1=ct($t.entries),U1=ct($t.set),mh=ct($t.reverse),W1=ct($t.fill),G1=ct($t.copyWithin),gh=ct($t.sort),Os=ct($t.slice),V1=ct($t.subarray),vt=Xn($t,"buffer"),$n=Xn($t,"byteOffset"),Q=Xn($t,"length"),bh=Xn($t,Ui),q1=Uint8Array,Xt=Uint16Array,yh=(...n)=>Vt(M1,Xt,n),ji=Uint32Array,j1=Float32Array,vn=Ls([][_e]()),zr=ct(vn.next),H1=ct(function*(){}().next),K1=Ls(vn),ht=TypeError,Hi=RangeError,wh=WeakSet,xh=wh.prototype,Y1=ct(xh.add),X1=ct(xh.has),Ur=WeakMap,Ki=Ur.prototype,Wr=ct(Ki.get),J1=ct(Ki.has),Yi=ct(Ki.set),Sh=new Ur,Z1=Or(null,{next:{value:function(){const t=Wr(Sh,this);return zr(t)}},[_e]:{value:function(){return this}}});function Gr(n){if(n[_e]===Gi&&vn.next===zr)return n;const t=Or(Z1);return Yi(Sh,t,T1(n)),t}const $h=new Ur,vh=Or(K1,{next:{value:function(){const t=Wr($h,this);return H1(t)},writable:!0,configurable:!0}});for(const n of oh(vn))n!=="next"&&Ms(vh,n,Jn(vn,n));function Ih(n){const t=Or(vh);return Yi($h,t,n),t}function Vr(n){return n!==null&&typeof n=="object"||typeof n=="function"}function Ah(n){return n!==null&&typeof n=="object"}function qr(n){return bh(n)!==void 0}function Xi(n){const t=bh(n);return t==="BigInt64Array"||t==="BigUint64Array"}function Q1(n){try{return fh(n)?!1:(P1(n),!0)}catch{return!1}}function Eh(n){if(Vi===null)return!1;try{return L1(n),!0}catch{return!1}}function tx(n){return Q1(n)||Eh(n)}function Ch(n){return fh(n)?n[_e]===Gi&&vn.next===zr:!1}function ex(n){return qr(n)?n[_e]===O1&&vn.next===zr:!1}function jr(n){if(typeof n!="string")return!1;const t=+n;return n!==t+""||!uh(t)?!1:t===dh(t)}const Hr=I1("__Float16Array__");function nx(n){if(!Ah(n))return!1;const t=Ls(n);if(!Ah(t))return!1;const e=t.constructor;if(e===void 0)return!1;if(!Vr(e))throw ht(eh);return zi(e,Hr)}const Ji=1/$1;function sx(n){return n+Ji-Ji}const kh=6103515625e-14,rx=65504,_h=.0009765625,Th=_h*kh,ox=_h*Ji;function ix(n){const t=+n;if(!uh(t)||t===0)return t;const e=t>0?1:-1,s=N1(t);if(s<kh)return e*sx(s/Th)*Th;const r=(1+ox)*s,o=r-(r-s);return o>rx||Zn(o)?e*(1/0):e*o}const Nh=new Fr(4),Dh=new j1(Nh),Rh=new ji(Nh),be=new Xt(512),ye=new q1(512);for(let n=0;n<256;++n){const t=n-127;t<-24?(be[n]=0,be[n|256]=32768,ye[n]=24,ye[n|256]=24):t<-14?(be[n]=1024>>-t-14,be[n|256]=1024>>-t-14|32768,ye[n]=-t-1,ye[n|256]=-t-1):t<=15?(be[n]=t+15<<10,be[n|256]=t+15<<10|32768,ye[n]=13,ye[n|256]=13):t<128?(be[n]=31744,be[n|256]=64512,ye[n]=24,ye[n|256]=24):(be[n]=31744,be[n|256]=64512,ye[n]=13,ye[n|256]=13)}function Te(n){Dh[0]=ix(n);const t=Rh[0],e=t>>23&511;return be[e]+((t&8388607)>>ye[e])}const Zi=new ji(2048);for(let n=1;n<1024;++n){let t=n<<13,e=0;for(;!(t&8388608);)t<<=1,e-=8388608;t&=-8388609,e+=947912704,Zi[n]=t|e}for(let n=1024;n<2048;++n)Zi[n]=939524096+(n-1024<<13);const ts=new ji(64);for(let n=1;n<31;++n)ts[n]=n<<23;ts[31]=1199570944,ts[32]=2147483648;for(let n=33;n<63;++n)ts[n]=2147483648+(n-32<<23);ts[63]=3347054592;const Ph=new Xt(64);for(let n=1;n<64;++n)n!==32&&(Ph[n]=1024);function st(n){const t=n>>10;return Rh[0]=Zi[Ph[t]+(n&1023)]+ts[t],Dh[0]}function Be(n){const t=+n;return Zn(t)||t===0?0:dh(t)}function Qi(n){const t=Be(n);return t<0?0:t<lh?t:lh}function Kr(n,t){if(!Vr(n))throw ht(g1);const e=n.constructor;if(e===void 0)return t;if(!Vr(e))throw ht(eh);const s=e[v1];return s??t}function Bs(n){if(Eh(n))return!1;try{return R1(n,0,0),!1}catch{}return!0}function Lh(n,t){const e=Zn(n),s=Zn(t);if(e&&s)return 0;if(e)return 1;if(s||n<t)return-1;if(n>t)return 1;if(n===0&&t===0){const r=ch(n,0),o=ch(t,0);if(!r&&o)return-1;if(r&&!o)return 1}return 0}const ta=2,Yr=new Ur;function es(n){return J1(Yr,n)||!D1(n)&&nx(n)}function tt(n){if(!es(n))throw ht(b1)}function Xr(n,t){const e=es(n),s=qr(n);if(!e&&!s)throw ht(y1);if(typeof t=="number"){let r;if(e){const o=V(n);r=Q(o)}else r=Q(n);if(r<t)throw ht(w1)}if(Xi(n))throw ht(Oi)}function V(n){const t=Wr(Yr,n);if(t!==void 0){const r=vt(t);if(Bs(r))throw ht(Rs);return t}const e=n.buffer;if(Bs(e))throw ht(Rs);const s=Ps(ft,[e,n.byteOffset,n.length],n.constructor);return Wr(Yr,s)}function Mh(n){const t=Q(n),e=[];for(let s=0;s<t;++s)e[s]=st(n[s]);return e}const Oh=new wh;for(const n of oh($t)){if(n===Ui)continue;const t=Jn($t,n);Oe(t,"get")&&typeof t.get=="function"&&Y1(Oh,t.get)}const ax=A1({get(n,t,e){return jr(t)&&Oe(n,t)?st(Fi(n,t)):X1(Oh,E1(n,t))?Fi(n,t):Fi(n,t,e)},set(n,t,e,s){return jr(t)&&Oe(n,t)?ih(n,t,Te(e)):ih(n,t,e,s)},getOwnPropertyDescriptor(n,t){if(jr(t)&&Oe(n,t)){const e=Jn(n,t);return e.value=st(e.value),e}return Jn(n,t)},defineProperty(n,t,e){return jr(t)&&Oe(n,t)&&Oe(e,"value")&&(e.value=Te(e.value)),rh(n,t,e)}});class ft{constructor(t,e,s){let r;if(es(t))r=Ps(Xt,[V(t)],new.target);else if(Vr(t)&&!tx(t)){let i,a;if(qr(t)){i=t,a=Q(t);const l=vt(t);if(Bs(l))throw ht(Rs);if(Xi(t))throw ht(Oi);const u=new Fr(a*ta);r=Ps(Xt,[u],new.target)}else{const l=t[_e];if(l!=null&&typeof l!="function")throw ht(nh);l!=null?Ch(t)?(i=t,a=t.length):(i=[...t],a=i.length):(i=t,a=Qi(i.length)),r=Ps(Xt,[a],new.target)}for(let l=0;l<a;++l)r[l]=Te(i[l])}else r=Ps(Xt,arguments,new.target);const o=new S1(r,ax);return Yi(Yr,o,r),o}static from(t,...e){const s=this;if(!zi(s,Hr))throw ht(th);if(s===ft){if(es(t)&&e.length===0){const c=V(t),h=new Xt(vt(c),$n(c),Q(c));return new ft(vt(Os(h)))}if(e.length===0)return new ft(vt(yh(t,Te)));const l=e[0],u=e[1];return new ft(vt(yh(t,function(c,...h){return Te(Vt(l,this,[c,...Gr(h)]))},u)))}let r,o;const i=t[_e];if(i!=null&&typeof i!="function")throw ht(nh);if(i!=null)Ch(t)?(r=t,o=t.length):ex(t)?(r=t,o=Q(t)):(r=[...t],o=r.length);else{if(t==null)throw ht(Mi);r=Qn(t),o=Qi(r.length)}const a=new s(o);if(e.length===0)for(let l=0;l<o;++l)a[l]=r[l];else{const l=e[0],u=e[1];for(let c=0;c<o;++c)a[c]=Vt(l,u,[r[c],c])}return a}static of(...t){const e=this;if(!zi(e,Hr))throw ht(th);const s=t.length;if(e===ft){const o=new ft(s),i=V(o);for(let a=0;a<s;++a)i[a]=Te(t[a]);return o}const r=new e(s);for(let o=0;o<s;++o)r[o]=t[o];return r}keys(){tt(this);const t=V(this);return B1(t)}values(){tt(this);const t=V(this);return Ih(function*(){for(const e of F1(t))yield st(e)}())}entries(){tt(this);const t=V(this);return Ih(function*(){for(const[e,s]of z1(t))yield[e,st(s)]}())}at(t){tt(this);const e=V(this),s=Q(e),r=Be(t),o=r>=0?r:s+r;if(!(o<0||o>=s))return st(e[o])}with(t,e){tt(this);const s=V(this),r=Q(s),o=Be(t),i=o>=0?o:r+o,a=+e;if(i<0||i>=r)throw Hi(Bi);const l=new Xt(vt(s),$n(s),Q(s)),u=new ft(vt(Os(l))),c=V(u);return c[i]=Te(a),u}map(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0],i=Kr(s,ft);if(i===ft){const l=new ft(r),u=V(l);for(let c=0;c<r;++c){const h=st(s[c]);u[c]=Te(Vt(t,o,[h,c,this]))}return l}const a=new i(r);Xr(a,r);for(let l=0;l<r;++l){const u=st(s[l]);a[l]=Vt(t,o,[u,l,this])}return a}filter(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0],i=[];for(let u=0;u<r;++u){const c=st(s[u]);Vt(t,o,[c,u,this])&&k1(i,c)}const a=Kr(s,ft),l=new a(i);return Xr(l),l}reduce(t,...e){tt(this);const s=V(this),r=Q(s);if(r===0&&e.length===0)throw ht(sh);let o,i;e.length===0?(o=st(s[0]),i=1):(o=e[0],i=0);for(let a=i;a<r;++a)o=t(o,st(s[a]),a,this);return o}reduceRight(t,...e){tt(this);const s=V(this),r=Q(s);if(r===0&&e.length===0)throw ht(sh);let o,i;e.length===0?(o=st(s[r-1]),i=r-2):(o=e[0],i=r-1);for(let a=i;a>=0;--a)o=t(o,st(s[a]),a,this);return o}forEach(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i)Vt(t,o,[st(s[i]),i,this])}find(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i){const a=st(s[i]);if(Vt(t,o,[a,i,this]))return a}}findIndex(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i){const a=st(s[i]);if(Vt(t,o,[a,i,this]))return i}return-1}findLast(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=r-1;i>=0;--i){const a=st(s[i]);if(Vt(t,o,[a,i,this]))return a}}findLastIndex(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=r-1;i>=0;--i){const a=st(s[i]);if(Vt(t,o,[a,i,this]))return i}return-1}every(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i)if(!Vt(t,o,[st(s[i]),i,this]))return!1;return!0}some(t,...e){tt(this);const s=V(this),r=Q(s),o=e[0];for(let i=0;i<r;++i)if(Vt(t,o,[st(s[i]),i,this]))return!0;return!1}set(t,...e){tt(this);const s=V(this),r=Be(e[0]);if(r<0)throw Hi(Bi);if(t==null)throw ht(Mi);if(Xi(t))throw ht(Oi);if(es(t))return U1(V(this),V(t),r);if(qr(t)){const l=vt(t);if(Bs(l))throw ht(Rs)}const o=Q(s),i=Qn(t),a=Qi(i.length);if(r===1/0||a+r>o)throw Hi(Bi);for(let l=0;l<a;++l)s[l+r]=Te(i[l])}reverse(){tt(this);const t=V(this);return mh(t),this}toReversed(){tt(this);const t=V(this),e=new Xt(vt(t),$n(t),Q(t)),s=new ft(vt(Os(e))),r=V(s);return mh(r),s}fill(t,...e){tt(this);const s=V(this);return W1(s,Te(t),...Gr(e)),this}copyWithin(t,e,...s){tt(this);const r=V(this);return G1(r,t,e,...Gr(s)),this}sort(t){tt(this);const e=V(this),s=t!==void 0?t:Lh;return gh(e,(r,o)=>s(st(r),st(o))),this}toSorted(t){tt(this);const e=V(this);if(t!==void 0&&typeof t!="function")throw new ht(x1);const s=t!==void 0?t:Lh,r=new Xt(vt(e),$n(e),Q(e)),o=new ft(vt(Os(r))),i=V(o);return gh(i,(a,l)=>s(st(a),st(l))),o}slice(t,e){tt(this);const s=V(this),r=Kr(s,ft);if(r===ft){const p=new Xt(vt(s),$n(s),Q(s));return new ft(vt(Os(p,t,e)))}const o=Q(s),i=Be(t),a=e===void 0?o:Be(e);let l;i===-1/0?l=0:i<0?l=o+i>0?o+i:0:l=o<i?o:i;let u;a===-1/0?u=0:a<0?u=o+a>0?o+a:0:u=o<a?o:a;const c=u-l>0?u-l:0,h=new r(c);if(Xr(h,c),c===0)return h;const f=vt(s);if(Bs(f))throw ht(Rs);let d=0;for(;l<u;)h[d]=st(s[l]),++l,++d;return h}subarray(t,e){tt(this);const s=V(this),r=Kr(s,ft),o=new Xt(vt(s),$n(s),Q(s)),i=V1(o,t,e),a=new r(vt(i),$n(i),Q(i));return Xr(a),a}indexOf(t,...e){tt(this);const s=V(this),r=Q(s);let o=Be(e[0]);if(o===1/0)return-1;o<0&&(o+=r,o<0&&(o=0));for(let i=o;i<r;++i)if(Oe(s,i)&&st(s[i])===t)return i;return-1}lastIndexOf(t,...e){tt(this);const s=V(this),r=Q(s);let o=e.length>=1?Be(e[0]):r-1;if(o===-1/0)return-1;o>=0?o=o<r-1?o:r-1:o+=r;for(let i=o;i>=0;--i)if(Oe(s,i)&&st(s[i])===t)return i;return-1}includes(t,...e){tt(this);const s=V(this),r=Q(s);let o=Be(e[0]);if(o===1/0)return!1;o<0&&(o+=r,o<0&&(o=0));const i=Zn(t);for(let a=o;a<r;++a){const l=st(s[a]);if(i&&Zn(l)||l===t)return!0}return!1}join(t){tt(this);const e=V(this),s=Mh(e);return C1(s,t)}toLocaleString(...t){tt(this);const e=V(this),s=Mh(e);return _1(s,...Gr(t))}get[Ui](){if(es(this))return"Float16Array"}}Ms(ft,"BYTES_PER_ELEMENT",{value:ta}),Ms(ft,Hr,{}),ah(ft,qi);const Jr=ft.prototype;Ms(Jr,"BYTES_PER_ELEMENT",{value:ta}),Ms(Jr,_e,{value:Jr.values,writable:!0,configurable:!0}),ah(Jr,$t);function lx(n,t){return n.channels===t.channels}const Zr=8;class Bh{constructor(t,e,s){K(this,"_label");K(this,"_device");K(this,"_outputBuffers",{});K(this,"_pipeline");K(this,"_bindGroups",[]);K(this,"_needsUpdatePipeline",!0);K(this,"_inputs",[]);K(this,"_outputs",[]);K(this,"_uniforms",[]);K(this,"_uniformBuffers",{});K(this,"_width",10);K(this,"_height",10);K(this,"_execWidth");K(this,"_execHeight");K(this,"_csCode","");K(this,"_csMain");K(this,"_csDefine");this._label=t,this._device=e,this._csMain=s.csMain,this._csDefine=s.csDefine,this._inputs=s.inputs,this._outputs=s.outputs,this._uniforms=s.uniforms,s.uniforms.forEach(r=>{this._uniformBuffers[r.label]=e.createBuffer({label:this._label,size:r.data.byteLength,usage:GPUBufferUsage.UNIFORM|GPUBufferUsage.COPY_DST}),this._device.queue.writeBuffer(this._uniformBuffers[r.label],0,r.data)})}setSize(t,e){t=Math.ceil(t),e=Math.ceil(e);const s=t!==this._width||e!==this._height;this._width=t,this._height=e,s&&(this._resizeOutputBuffers(),this._needsUpdatePipeline=!0)}setExecuteSize(t,e){t=Math.ceil(t),e=Math.ceil(e),this._execWidth=t,this._execHeight=e}setOutputParams(t){this._updateOutputBuffers(t),this._needsUpdatePipeline=!0}setUniform(t,e){const s=this._uniformBuffers[t];this._device.queue.writeBuffer(s,0,e)}getOutputBuffer(t){return this._outputBuffers[t].buffer}dispose(){Object.keys(this._uniformBuffers).forEach(t=>{this._uniformBuffers[t].destroy()}),Object.keys(this._outputBuffers).forEach(t=>{this._outputBuffers[t].texture.destroy()})}_createBuffer(t){const e=this._width*this._height*4*4;return this._device.createBuffer({label:this._label,size:Math.max(e,80),usage:GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_DST|GPUBufferUsage.COPY_SRC})}_resizeOutputBuffers(){const t=this._outputBuffers;for(const e in t){const{buffer:s,params:r}=t[e];s.destroy(),t[e].buffer=this._createBuffer(r)}}_updateOutputBuffers(t){var s,r;const e=this._outputBuffers;for(const o in t){const i=t[o];if(!lx(i,((s=e[o])==null?void 0:s.params)||{})){(r=e[o])==null||r.buffer.destroy();const a=this._createBuffer(i);e[o]={buffer:a,params:i}}}}_updatePipeline(t){if(!this._needsUpdatePipeline)return;this._needsUpdatePipeline=!1;const e=this._device,s=this._getFullCs(t);s!==this._csCode&&(this._csCode=s,this._pipeline=e.createComputePipeline({label:this._label,layout:"auto",compute:{module:e.createShaderModule({label:this._label,code:s}),entryPoint:"main"}}),this._updateBindGroups())}_getFullCs(t){const e=this._inputs,s=e.length>0;return`
 ${e.sort().map((o,i)=>`@group(0) @binding(${i}) var<storage, read> in_${o}: array<vec${t[o].channels}f>;`).join(`
 `)}
 ${this._uniforms.map((o,i)=>`@group(${s?1:0}) @binding(${i}) var<uniform> ${o.label}: ${o.type};`).join(`
@@ -3053,11 +3053,11 @@ ${this._uniforms.map((o,i)=>`@group(${s?1:0}) @binding(${i}) var<uniform> ${o.la
 ${this._outputs.map((o,i)=>`@group(${s?2:1}) @binding(${i}) var<storage, read_write> out_${o}: array<vec${this._outputBuffers[o].params.channels}f>;`).join(`
 `)}
 ${this._csDefine??""}
-@compute @workgroup_size(${Jr}, ${Jr}, 1)
+@compute @workgroup_size(${Zr}, ${Zr}, 1)
 fn main(@builtin(global_invocation_id) globalId: vec3u) {
 ${this._csMain}
 }
-`}_updateBindGroups(){const t=[],e=this._device,s=this._inputs.length>0?1:0;this._uniforms.length>0&&(t[s]=e.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(s),entries:this._uniforms.map((r,o)=>({binding:o,resource:{buffer:this._uniformBuffers[r.label]}}))})),this._bindGroups=t}createPass(t,e){this._updatePipeline(e);const s=this._inputs.length>0;s&&(this._bindGroups[0]=this._device.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(0),entries:this._inputs.map((o,i)=>({binding:i,resource:{buffer:e[o].buffer}}))})),this._bindGroups[s?2:1]=this._device.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(s?2:1),entries:this._outputs.map((o,i)=>({binding:i,resource:{buffer:this._outputBuffers[o].buffer}}))});const r=t.beginComputePass();r.setPipeline(this._pipeline),this._bindGroups.forEach((o,i)=>{r.setBindGroup(i,o)}),r.dispatchWorkgroups(Math.ceil((this._execWidth??this._width)/Jr),Math.ceil((this._execHeight??this._height)/Jr),1),r.end()}}const ea=1412.83765,na=1.64593172,sa=.431384981,ra=-.00294139609,oa=.192653254,ia=.00626026094,aa=.998620152,Fh=15794576e-13,zh=.0322087631,Uh=.00223151711,Wh=.370974749;function Gh(n){return n<=Fh?n=ea*n:n<=zh?n=na*Math.pow(n,sa)+ra:n=oa*Math.log(n+ia)+aa,n}function ux(n){return n<=Uh?n=n/ea:n<=Wh?n=Math.pow((n-ra)/na,1/sa):n=Math.exp((n-aa)/oa)-ia,n}const Vh=Gh(65504),qh=1/Vh,jh=Vh;class la{constructor(t,e,s,r){this.x=t,this.y=e,this.width=s,this.height=r}}function cx({data:n,channels:t}){let e=0;for(let i=0;i<n.length;i+=t){const a=n[i],l=n[i+1],u=n[i+2],c=.212671*a+.71516*l+.072169*u;e+=Math.log2(c+1e-4)}const s=n.length/t,r=e/s;return .18/Math.pow(2,r)}function hx({data:n,channels:t,inputScale:e}){const s=new Float32Array(n.length);s.set(n);for(let r=0;r<s.length;r+=t)for(let o=0;o<3;o++){let i=s[r+o]*e;s[r+o]=Gh(i)*qh}return s}function fx({data:n,channels:t,inputScale:e}){const s=new Float32Array(n.length);s.set(n);const r=1/e;for(let o=0;o<s.length;o+=t)for(let i=0;i<3;i++){let a=s[o+i]*jh;s[o+i]=ux(a)*r}return s}const Hh=`
+`}_updateBindGroups(){const t=[],e=this._device,s=this._inputs.length>0?1:0;this._uniforms.length>0&&(t[s]=e.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(s),entries:this._uniforms.map((r,o)=>({binding:o,resource:{buffer:this._uniformBuffers[r.label]}}))})),this._bindGroups=t}createPass(t,e){this._updatePipeline(e);const s=this._inputs.length>0;s&&(this._bindGroups[0]=this._device.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(0),entries:this._inputs.map((o,i)=>({binding:i,resource:{buffer:e[o].buffer}}))})),this._bindGroups[s?2:1]=this._device.createBindGroup({label:this._label,layout:this._pipeline.getBindGroupLayout(s?2:1),entries:this._outputs.map((o,i)=>({binding:i,resource:{buffer:this._outputBuffers[o].buffer}}))});const r=t.beginComputePass();r.setPipeline(this._pipeline),this._bindGroups.forEach((o,i)=>{r.setBindGroup(i,o)}),r.dispatchWorkgroups(Math.ceil((this._execWidth??this._width)/Zr),Math.ceil((this._execHeight??this._height)/Zr),1),r.end()}}const ea=1412.83765,na=1.64593172,sa=.431384981,ra=-.00294139609,oa=.192653254,ia=.00626026094,aa=.998620152,Fh=15794576e-13,zh=.0322087631,Uh=.00223151711,Wh=.370974749;function Gh(n){return n<=Fh?n=ea*n:n<=zh?n=na*Math.pow(n,sa)+ra:n=oa*Math.log(n+ia)+aa,n}function ux(n){return n<=Uh?n=n/ea:n<=Wh?n=Math.pow((n-ra)/na,1/sa):n=Math.exp((n-aa)/oa)-ia,n}const Vh=Gh(65504),qh=1/Vh,jh=Vh;class la{constructor(t,e,s,r){this.x=t,this.y=e,this.width=s,this.height=r}}function cx({data:n,channels:t}){let e=0;for(let i=0;i<n.length;i+=t){const a=n[i],l=n[i+1],u=n[i+2],c=.212671*a+.71516*l+.072169*u;e+=Math.log2(c+1e-4)}const s=n.length/t,r=e/s;return .18/Math.pow(2,r)}function hx({data:n,channels:t,inputScale:e}){const s=new Float32Array(n.length);s.set(n);for(let r=0;r<s.length;r+=t)for(let o=0;o<3;o++){let i=s[r+o]*e;s[r+o]=Gh(i)*qh}return s}function fx({data:n,channels:t,inputScale:e}){const s=new Float32Array(n.length);s.set(n);const r=1/e;for(let o=0;o<s.length;o+=t)for(let i=0;i<3;i++){let a=s[o+i]*jh;s[o+i]=ux(a)*r}return s}const Hh=`
 const a = ${ea};
 const b = ${na};
 const c = ${sa};
@@ -3121,7 +3121,7 @@ out_color[outIdx] = vec4f(
   // Pick the alpha
   raw.a
 );
-`}),this._inputPass.setOutputParams({color:{channels:3},albedo:{channels:3},normal:{channels:3}}),this._outputPass.setOutputParams({color:{channels:4}})}setImageSize(t,e){this._inputPass.setUniform("inputSize",new Float32Array([t,e])),this._outputPass.setUniform("imageSize",new Float32Array([t,e])),this._outputPass.setSize(t,e)}setInputTile(t){const e=this._inputPass,s=new Float32Array([t.width,t.height]);e.setUniform("inputOffset",new Float32Array([t.x,t.y])),e.setUniform("outputSize",s),e.setSize(s[0],s[1]),this._outputPass.setUniform("inputSize",s)}setOutputTile(t,e){const s=this._outputPass,r=new Float32Array([t.width,t.height]),o=t.x-e.x,i=t.y-e.y;s.setUniform("outputSize",r),s.setUniform("inputOffset",new Float32Array([o,i])),s.setUniform("outputOffset",new Float32Array([t.x,t.y])),s.setExecuteSize(r[0],r[1])}forward(t,e,s){const r=this._inputPass,o=this._device.createCommandEncoder();return r.createPass(o,{color:{buffer:t,channels:4},albedo:{buffer:e,channels:4},normal:{buffer:s,channels:4}}),this._device.queue.submit([o.finish()]),{color:r.getOutputBuffer("color"),albedo:r.getOutputBuffer("albedo"),normal:r.getOutputBuffer("normal")}}inverse(t,e){const r=this._device.createCommandEncoder(),o=this._outputPass;return o.createPass(r,{color:{buffer:t,channels:4},raw:{buffer:e,channels:4}}),this._device.queue.submit([r.finish()]),o.getOutputBuffer("color")}copyInputDataToOutput(t){const e=this._device.createCommandEncoder(),r=this._outputPass.getOutputBuffer("color");e.copyBufferToBuffer(t,0,r,0,r.size),this._device.queue.submit([e.finish()])}dispose(){this._outputPass.dispose(),this._inputPass.dispose()}}function Kh(n,t){const e=n.buffer;if(t==="Float32")return new Float32Array(n.buffer);const s=new ft(e),r=new Float32Array(s.length);for(let o=0;o<r.length;++o)r[o]=s[o];return r}function px(n,t){const[e,s,r,o]=t,i=new Float32Array(n.length);for(let a=0;a<e;++a)for(let l=0;l<s;++l)for(let u=0;u<r;++u)for(let c=0;c<o;++c){const h=a*s*r*o+l*r*o+u*o+c,f=u*o*s*e+c*s*e+l*e+a;i[f]=n[h]}return i}function ua(n,t){return Math.ceil(n/t)*t}function Zr(n){return n.data instanceof GPUBuffer}const mx=174,ca=16,Qr=ua(mx/2,ca);class Yh{constructor(t,e,s={}){K(this,"_tfModel");K(this,"_device");K(this,"_tileWidth",0);K(this,"_tileHeight",0);K(this,"_tileOverlapX",0);K(this,"_tileOverlapY",0);K(this,"_aux",!1);K(this,"_hdr",!1);K(this,"_dataProcessGPU");K(this,"_maxTileSize");this._tensors=t,this._backend=e,this._aux=s.aux||!1,this._hdr=s.hdr||!1,this._maxTileSize=s.maxTileSize??512,this._device=this._backend.device}_createConv(t,e,s){const r=this._tensors.get(t+".weight"),o=this._tensors.get(t+".bias"),i=r.desc.dims,a=tr(px(Kh(r.data,r.desc.dataType),i),[i[2],i[3],i[1],i[0]],"float32"),l=Rt(Kh(o.data,o.desc.dataType),"float32");return new Yn({name:t,filters:r.desc.dims[0],kernelSize:r.desc.dims.slice(2,4),useBias:!0,activation:s,padding:"same",weights:[a,l],trainable:!1}).apply(e)}_createConcatConv(t,e,s){return this._createConv(t,new ki({trainable:!1,axis:3}).apply([e,s]),"relu")}_createPooling(t){return new Ii({name:t.name+"/pooling",poolSize:[2,2],strides:[2,2],padding:"same",trainable:!1}).apply(t)}_addUpsamplingLayer(t){return new vi({name:t.name+"/upsampling",size:[2,2],trainable:!1}).apply(t)}getDevice(){return this._device}buildModel(){const e=3+(this._aux?6:0),s=this._getTileSizeWithOverlap(),r=Jw({shape:[s.height,s.width,e],dtype:"float32"}),o=this._createConv("enc_conv0",r,"relu"),i=this._createPooling(this._createConv("enc_conv1",o,"relu")),a=this._createPooling(this._createConv("enc_conv2",i,"relu")),l=this._createPooling(this._createConv("enc_conv3",a,"relu")),u=this._createPooling(this._createConv("enc_conv4",l,"relu")),c=this._createConv("enc_conv5a",u,"relu"),h=this._addUpsamplingLayer(this._createConv("enc_conv5b",c,"relu")),f=this._createConcatConv("dec_conv4a",h,l),d=this._addUpsamplingLayer(this._createConv("dec_conv4b",f,"relu")),p=this._createConcatConv("dec_conv3a",d,a),g=this._addUpsamplingLayer(this._createConv("dec_conv3b",p,"relu")),m=this._createConcatConv("dec_conv2a",g,i),b=this._addUpsamplingLayer(this._createConv("dec_conv2b",m,"relu")),y=this._createConcatConv("dec_conv1a",b,r),S=this._createConv("dec_conv1b",y,"relu"),x=this._createConv("dec_conv0",S,"relu");this._tfModel=new Lr({inputs:[r],outputs:x})}_updateModel(t,e){const s=this._maxTileSize;let r=s,o=s,i=Qr,a=Qr;t<s+Qr*2&&(r=ua(t,ca),i=0),e<s+Qr*2&&(o=ua(e,ca),a=0),(r!==this._tileWidth||o!==this._tileHeight||i!==this._tileOverlapX||a!==this._tileOverlapY||!this._tfModel)&&(this._tileWidth=r,this._tileHeight=o,this._tileOverlapX=i,this._tileOverlapY=a,this._tfModel&&this._tfModel.dispose(),this.buildModel())}_getTileSizeWithOverlap(){return{width:this._tileWidth+2*this._tileOverlapX,height:this._tileHeight+2*this._tileOverlapY}}_processImageData(t,e,s,r){const o=t.data,i=o.length/4,a=this._aux?9:3,l=new Float32Array(i*a);if(e&&!s||s&&!e)throw new Error("Normal map and albedo map are both required");if(e&&s&&(e.width!==s.width||e.height!==s.height||t.width!==e.width||t.height!==e.height))throw new Error("Image size mismatch");const u=e==null?void 0:e.data,c=s==null?void 0:s.data;for(let h=0;h<o.length;h+=4){const f=h/4*a;for(let d=0;d<3;d++)r?l[f+d]=o[h+d]:l[f+d]=o[h+d]/255,u&&(l[f+d+3]=u[h+d]/255),c&&(l[f+d+6]=c[h+d]/255)}return l}_readTile(t,e,s,r){const o=new Float32Array(s.width*s.height*e);for(let i=0;i<s.height;i++)for(let a=0;a<s.width;a++){const l=((i+s.y)*r+(a+s.x))*e,u=(i*s.width+a)*e;for(let c=0;c<e;c++)o[u+c]=t[l+c]}return o}_writeTile(t,e,s,r,o,i){const{data:a,width:l}=t,u=s.x-e.x,c=s.y-e.y;for(let h=0;h<s.height;h++)for(let f=0;f<s.width;f++){const d=((h+c)*o+f+u)*3,p=((h+s.y)*l+(f+s.x))*4;for(let g=0;g<3;g++)i?a[p+g]=r[d+g]:a[p+g]=Math.min(Math.max(r[d+g]*255,0),255);t.data[p+3]=i?1:255}}_executeTile(t,e,s,r,o,i,a,l){const u=this._aux?9:3,c=this._tileOverlapX,h=this._tileOverlapY;let f=this._getTileSizeWithOverlap(),d={width:this._tileWidth,height:this._tileHeight},p=r>0?r*d.width-c:0,g=Math.min(p+f.width,i);p=Math.max(g-f.width,0);let m=o>0?o*d.height-h:0,b=Math.min(m+f.height,a);m=Math.max(b-f.height,0);const y=Math.min(f.width,i),S=Math.min(f.height,a),x=i<d.width||a<d.height,$=new la(p,m,y,S);let E,D=1;const _=this._device;let T=this._dataProcessGPU;if(t instanceof Float32Array){let F=this._readTile(t,u,$,i);l&&(D=cx({data:F,channels:9}),F=hx({data:F,channels:9,inputScale:D})),E=tr(F,[1,S,y,u],"float32")}else{if(!l)throw new Error("Only hdr is supported for webgpu data.");T||(T=this._dataProcessGPU=new dx(_)),T.setImageSize(i,a),T.setInputTile($),r===0&&o===0&&T.copyInputDataToOutput(t.color);const{color:F,albedo:G,normal:q}=T.forward(t.color,t.albedo,t.normal),De=[1,S,y,4];E=jp([F,G,q].map(Xt=>{const Jt=tr({buffer:Xt,zeroCopy:!0},De),An=cs(Jt,[0,0,0,0],[1,S,y,3]);return Jt.dispose(),An}),3)}if(x){const F=E;E=Mp(F,[[0,0],[0,f.height-a],[0,f.width-i],[0,0]],"reflect"),F.dispose()}const P=this._tfModel.predict(E);E.dispose();const B=Math.min(d.width,i),Y=Math.min(d.height,a),j=new la(r*B,o*Y,B,Y);if(j.width=Math.min(j.width,i-j.x),j.height=Math.min(j.height,a-j.y),t instanceof Float32Array){let F=P.dataSync();l&&(F=fx({data:F,channels:3,inputScale:D})),this._writeTile(s,$,j,F,f.width,l);for(let G=0;G<Y;G++)for(let q=0;q<B;q++){const De=(G*B+q)*4,Xt=((G+j.y)*i+(q+j.x))*4;for(let Jt=0;Jt<4;Jt++)e.data[De+Jt]=s.data[Xt+Jt]}P.dispose()}else{T.setOutputTile(j,$);const F=zp(P,[[0,0],[0,0],[0,0],[0,1]]),G=T.inverse(F.dataToGPU().buffer,t.color);return P.dispose(),F.dispose(),G}}progressiveExecute({color:t,albedo:e,normal:s,done:r,progress:o}){if(this._aux&&(!e||!s))throw new Error("Normal map and albedo map are both required");if(!this._aux&&(e||s))throw new Error("Normal map and albedo map are not required");const i=t.width,a=t.height;this._updateModel(i,a);const l=this._hdr||!1;let u;Zr(t)||(u=this._processImageData(t,e,s,l));const c=this._tileWidth,h=this._tileHeight,f=Math.ceil(a/h),d=Math.ceil(i/c);function p(S,x){return l?{data:new Float32Array(S*x*4),width:S,height:x}:new ImageData(S,x)}const g=Zr(t)?void 0:p(i,a),m=Zr(t)?void 0:p(Math.min(c,i),Math.min(h,a));let b=!1;const y=(S,x)=>{if(b)return;let $;$=this._executeTile(Zr(t)?{color:t.data,albedo:e.data,normal:s.data}:u,m,g,S,x,i,a,l);const E=g||{data:$,width:i,height:a};o==null||o(E,m,new la(S*c,x*h,c,h),S+x*d,d*f),S+1<d||x+1<f?requestAnimationFrame(()=>{S+1<d?y(S+1,x):x+1<f&&y(0,x+1)}):r(E)};return y(0,0),()=>{b=!0}}dispose(){var t,e;(t=this._tfModel)==null||t.dispose(),(e=this._dataProcessGPU)==null||e.dispose()}}/**
+`}),this._inputPass.setOutputParams({color:{channels:3},albedo:{channels:3},normal:{channels:3}}),this._outputPass.setOutputParams({color:{channels:4}})}setImageSize(t,e){this._inputPass.setUniform("inputSize",new Float32Array([t,e])),this._outputPass.setUniform("imageSize",new Float32Array([t,e])),this._outputPass.setSize(t,e)}setInputTile(t){const e=this._inputPass,s=new Float32Array([t.width,t.height]);e.setUniform("inputOffset",new Float32Array([t.x,t.y])),e.setUniform("outputSize",s),e.setSize(s[0],s[1]),this._outputPass.setUniform("inputSize",s)}setOutputTile(t,e){const s=this._outputPass,r=new Float32Array([t.width,t.height]),o=t.x-e.x,i=t.y-e.y;s.setUniform("outputSize",r),s.setUniform("inputOffset",new Float32Array([o,i])),s.setUniform("outputOffset",new Float32Array([t.x,t.y])),s.setExecuteSize(r[0],r[1])}forward(t,e,s){const r=this._inputPass,o=this._device.createCommandEncoder();return r.createPass(o,{color:{buffer:t,channels:4},albedo:{buffer:e,channels:4},normal:{buffer:s,channels:4}}),this._device.queue.submit([o.finish()]),{color:r.getOutputBuffer("color"),albedo:r.getOutputBuffer("albedo"),normal:r.getOutputBuffer("normal")}}inverse(t,e){const r=this._device.createCommandEncoder(),o=this._outputPass;return o.createPass(r,{color:{buffer:t,channels:4},raw:{buffer:e,channels:4}}),this._device.queue.submit([r.finish()]),o.getOutputBuffer("color")}copyInputDataToOutput(t){const e=this._device.createCommandEncoder(),r=this._outputPass.getOutputBuffer("color");e.copyBufferToBuffer(t,0,r,0,r.size),this._device.queue.submit([e.finish()])}dispose(){this._outputPass.dispose(),this._inputPass.dispose()}}function Kh(n,t){const e=n.buffer;if(t==="Float32")return new Float32Array(n.buffer);const s=new ft(e),r=new Float32Array(s.length);for(let o=0;o<r.length;++o)r[o]=s[o];return r}function px(n,t){const[e,s,r,o]=t,i=new Float32Array(n.length);for(let a=0;a<e;++a)for(let l=0;l<s;++l)for(let u=0;u<r;++u)for(let c=0;c<o;++c){const h=a*s*r*o+l*r*o+u*o+c,f=u*o*s*e+c*s*e+l*e+a;i[f]=n[h]}return i}function ua(n,t){return Math.ceil(n/t)*t}function Qr(n){return n.data instanceof GPUBuffer}const mx=174,ca=16,to=ua(mx/2,ca);class Yh{constructor(t,e,s={}){K(this,"_tfModel");K(this,"_device");K(this,"_tileWidth",0);K(this,"_tileHeight",0);K(this,"_tileOverlapX",0);K(this,"_tileOverlapY",0);K(this,"_aux",!1);K(this,"_hdr",!1);K(this,"_dataProcessGPU");K(this,"_maxTileSize");this._tensors=t,this._backend=e,this._aux=s.aux||!1,this._hdr=s.hdr||!1,this._maxTileSize=s.maxTileSize??512,this._device=this._backend.device}_createConv(t,e,s){const r=this._tensors.get(t+".weight"),o=this._tensors.get(t+".bias"),i=r.desc.dims,a=er(px(Kh(r.data,r.desc.dataType),i),[i[2],i[3],i[1],i[0]],"float32"),l=Rt(Kh(o.data,o.desc.dataType),"float32");return new Yn({name:t,filters:r.desc.dims[0],kernelSize:r.desc.dims.slice(2,4),useBias:!0,activation:s,padding:"same",weights:[a,l],trainable:!1}).apply(e)}_createConcatConv(t,e,s){return this._createConv(t,new ki({trainable:!1,axis:3}).apply([e,s]),"relu")}_createPooling(t){return new Ii({name:t.name+"/pooling",poolSize:[2,2],strides:[2,2],padding:"same",trainable:!1}).apply(t)}_addUpsamplingLayer(t){return new vi({name:t.name+"/upsampling",size:[2,2],trainable:!1}).apply(t)}getDevice(){return this._device}buildModel(){const e=3+(this._aux?6:0),s=this._getTileSizeWithOverlap(),r=Jw({shape:[s.height,s.width,e],dtype:"float32"}),o=this._createConv("enc_conv0",r,"relu"),i=this._createPooling(this._createConv("enc_conv1",o,"relu")),a=this._createPooling(this._createConv("enc_conv2",i,"relu")),l=this._createPooling(this._createConv("enc_conv3",a,"relu")),u=this._createPooling(this._createConv("enc_conv4",l,"relu")),c=this._createConv("enc_conv5a",u,"relu"),h=this._addUpsamplingLayer(this._createConv("enc_conv5b",c,"relu")),f=this._createConcatConv("dec_conv4a",h,l),d=this._addUpsamplingLayer(this._createConv("dec_conv4b",f,"relu")),p=this._createConcatConv("dec_conv3a",d,a),g=this._addUpsamplingLayer(this._createConv("dec_conv3b",p,"relu")),m=this._createConcatConv("dec_conv2a",g,i),b=this._addUpsamplingLayer(this._createConv("dec_conv2b",m,"relu")),y=this._createConcatConv("dec_conv1a",b,r),S=this._createConv("dec_conv1b",y,"relu"),x=this._createConv("dec_conv0",S,"relu");this._tfModel=new Mr({inputs:[r],outputs:x})}_updateModel(t,e){const s=this._maxTileSize;let r=s,o=s,i=to,a=to;t<s+to*2&&(r=ua(t,ca),i=0),e<s+to*2&&(o=ua(e,ca),a=0),(r!==this._tileWidth||o!==this._tileHeight||i!==this._tileOverlapX||a!==this._tileOverlapY||!this._tfModel)&&(this._tileWidth=r,this._tileHeight=o,this._tileOverlapX=i,this._tileOverlapY=a,this._tfModel&&this._tfModel.dispose(),this.buildModel())}_getTileSizeWithOverlap(){return{width:this._tileWidth+2*this._tileOverlapX,height:this._tileHeight+2*this._tileOverlapY}}_processImageData(t,e,s,r){const o=t.data,i=o.length/4,a=this._aux?9:3,l=new Float32Array(i*a);if(e&&!s||s&&!e)throw new Error("Normal map and albedo map are both required");if(e&&s&&(e.width!==s.width||e.height!==s.height||t.width!==e.width||t.height!==e.height))throw new Error("Image size mismatch");const u=e==null?void 0:e.data,c=s==null?void 0:s.data;for(let h=0;h<o.length;h+=4){const f=h/4*a;for(let d=0;d<3;d++)r?l[f+d]=o[h+d]:l[f+d]=o[h+d]/255,u&&(l[f+d+3]=u[h+d]/255),c&&(l[f+d+6]=c[h+d]/255)}return l}_readTile(t,e,s,r){const o=new Float32Array(s.width*s.height*e);for(let i=0;i<s.height;i++)for(let a=0;a<s.width;a++){const l=((i+s.y)*r+(a+s.x))*e,u=(i*s.width+a)*e;for(let c=0;c<e;c++)o[u+c]=t[l+c]}return o}_writeTile(t,e,s,r,o,i){const{data:a,width:l}=t,u=s.x-e.x,c=s.y-e.y;for(let h=0;h<s.height;h++)for(let f=0;f<s.width;f++){const d=((h+c)*o+f+u)*3,p=((h+s.y)*l+(f+s.x))*4;for(let g=0;g<3;g++)i?a[p+g]=r[d+g]:a[p+g]=Math.min(Math.max(r[d+g]*255,0),255);t.data[p+3]=i?1:255}}_executeTile(t,e,s,r,o,i,a,l){const u=this._aux?9:3,c=this._tileOverlapX,h=this._tileOverlapY;let f=this._getTileSizeWithOverlap(),d={width:this._tileWidth,height:this._tileHeight},p=r>0?r*d.width-c:0,g=Math.min(p+f.width,i);p=Math.max(g-f.width,0);let m=o>0?o*d.height-h:0,b=Math.min(m+f.height,a);m=Math.max(b-f.height,0);const y=Math.min(f.width,i),S=Math.min(f.height,a),x=i<d.width||a<d.height,$=new la(p,m,y,S);let E,D=1;const _=this._device;let T=this._dataProcessGPU;if(t instanceof Float32Array){let F=this._readTile(t,u,$,i);l&&(D=cx({data:F,channels:9}),F=hx({data:F,channels:9,inputScale:D})),E=er(F,[1,S,y,u],"float32")}else{if(!l)throw new Error("Only hdr is supported for webgpu data.");T||(T=this._dataProcessGPU=new dx(_)),T.setImageSize(i,a),T.setInputTile($),r===0&&o===0&&T.copyInputDataToOutput(t.color);const{color:F,albedo:G,normal:q}=T.forward(t.color,t.albedo,t.normal),De=[1,S,y,4],qt=[F,G,q].map(jt=>{const Xe=er({buffer:jt,zeroCopy:!0},De),os=hs(Xe,[0,0,0,0],[1,S,y,3]);return Xe.dispose(),os});E=jp(qt,3),qt.forEach(jt=>jt.dispose())}if(x){const F=E;E=Mp(F,[[0,0],[0,f.height-a],[0,f.width-i],[0,0]],"reflect"),F.dispose()}const P=this._tfModel.predict(E);E.dispose();const B=Math.min(d.width,i),Y=Math.min(d.height,a),j=new la(r*B,o*Y,B,Y);if(j.width=Math.min(j.width,i-j.x),j.height=Math.min(j.height,a-j.y),t instanceof Float32Array){let F=P.dataSync();l&&(F=fx({data:F,channels:3,inputScale:D})),this._writeTile(s,$,j,F,f.width,l);for(let G=0;G<Y;G++)for(let q=0;q<B;q++){const De=(G*B+q)*4,qt=((G+j.y)*i+(q+j.x))*4;for(let jt=0;jt<4;jt++)e.data[De+jt]=s.data[qt+jt]}P.dispose()}else{T.setOutputTile(j,$);const F=zp(P,[[0,0],[0,0],[0,0],[0,1]]),G=T.inverse(F.dataToGPU().buffer,t.color);return P.dispose(),F.dispose(),G}}progressiveExecute({color:t,albedo:e,normal:s,done:r,progress:o}){if(this._aux&&(!e||!s))throw new Error("Normal map and albedo map are both required");if(!this._aux&&(e||s))throw new Error("Normal map and albedo map are not required");const i=t.width,a=t.height;this._updateModel(i,a);const l=this._hdr||!1;let u;Qr(t)||(u=this._processImageData(t,e,s,l));const c=this._tileWidth,h=this._tileHeight,f=Math.ceil(a/h),d=Math.ceil(i/c);function p(S,x){return l?{data:new Float32Array(S*x*4),width:S,height:x}:new ImageData(S,x)}const g=Qr(t)?void 0:p(i,a),m=Qr(t)?void 0:p(Math.min(c,i),Math.min(h,a));let b=!1;const y=(S,x)=>{if(b)return;let $;$=this._executeTile(Qr(t)?{color:t.data,albedo:e.data,normal:s.data}:u,m,g,S,x,i,a,l);const E=g||{data:$,width:i,height:a};o==null||o(E,m,new la(S*c,x*h,c,h),S+x*d,d*f),S+1<d||x+1<f?requestAnimationFrame(()=>{S+1<d?y(S+1,x):x+1<f&&y(0,x+1)}):r(E)};return y(0,0),()=>{b=!0}}dispose(){var t,e;(t=this._tfModel)==null||t.dispose(),(e=this._dataProcessGPU)==null||e.dispose()}}/**
  * @license
  * Copyright 2019 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -3136,7 +3136,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const Yt=W();Yt.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15),Yt.registerFlag("WEBGPU_CPU_FORWARD",()=>!0),Yt.registerFlag("WEBGPU_MATMUL_PROGRAM_TYPE",()=>-1),Yt.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!0),Yt.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1),Yt.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3),Yt.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1),Yt.registerFlag("WEBGPU_IMPORT_EXTERNAL_TEXTURE",()=>!0),Yt.registerFlag("WEBGPU_USE_NAIVE_CONV2D_DEBUG",()=>!1),Yt.registerFlag("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL",()=>-1),Yt.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER",()=>!1),Yt.registerFlag("WEBGPU_PRINT_SHADER",()=>""),Yt.registerFlag("WEBGPU_ENGINE_COMPILE_ONLY",()=>!1);/**
+ */const Jt=W();Jt.registerFlag("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE",()=>15),Jt.registerFlag("WEBGPU_CPU_FORWARD",()=>!0),Jt.registerFlag("WEBGPU_MATMUL_PROGRAM_TYPE",()=>-1),Jt.registerFlag("WEBGPU_USE_NAIVE_CONV2D_TRANSPOSE",()=>!0),Jt.registerFlag("WEBGPU_USE_LOW_POWER_GPU",()=>!1),Jt.registerFlag("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD",()=>1e3),Jt.registerFlag("WEBGPU_USE_PROFILE_TOOL",()=>!1),Jt.registerFlag("WEBGPU_IMPORT_EXTERNAL_TEXTURE",()=>!0),Jt.registerFlag("WEBGPU_USE_NAIVE_CONV2D_DEBUG",()=>!1),Jt.registerFlag("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL",()=>-1),Jt.registerFlag("WEBGPU_CONV_SEPARATE_IM2COL_SHADER",()=>!1),Jt.registerFlag("WEBGPU_PRINT_SHADER",()=>""),Jt.registerFlag("WEBGPU_ENGINE_COMPILE_ONLY",()=>!1);/**
  * @license
  * Copyright 2022 Google LLC.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -3223,7 +3223,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */var to;(function(n){n[n.FROM_PIXELS=0]="FROM_PIXELS",n[n.DRAW=1]="DRAW"})(to||(to={}));const Sx=(n,t,e,s,r)=>{const o={dtype:s.dtype,shape:s.shape},i=vx(e,o,t),a=n.createShaderModule({code:i,label:t.constructor.name});let l=W().get("WEBGPU_PRINT_SHADER");if(l!==""){l=l.toLowerCase();const u=l.split(",");(l==="all"||u.some(c=>t.shaderKey.toLowerCase().includes(c)))&&(console.group(t.shaderKey),console.debug(i),console.groupEnd())}return r?n.createComputePipelineAsync({compute:{module:a,entryPoint:"_start"},label:t.constructor.name,layout:"auto"}):n.createComputePipeline({compute:{module:a,entryPoint:"_start"},label:t.constructor.name,layout:"auto"})},H=(n,t="f32")=>{switch(n){case 1:return`${t}`;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${n}-component ${t} is not supported.`)}};function Tt(n){if(n<=1)return"i32";if(n===2)return"vec2<i32>";if(n===3)return"vec3<i32>";if(n===4)return"vec4<i32>";if(n===5)return"vec5";if(n===6)return"vec6";throw Error(`GPU for rank ${n} is not yet supported`)}function vn(n){if(n===0)return"x";if(n===1)return"y";if(n===2)return"z";if(n===3)return"w";if(n===4)return"u";if(n===5)return"v";throw Error(`Index ${n} is not yet supported`)}function wt(...n){let t;switch(n.length){case 0:t=`
+ */var eo;(function(n){n[n.FROM_PIXELS=0]="FROM_PIXELS",n[n.DRAW=1]="DRAW"})(eo||(eo={}));const Sx=(n,t,e,s,r)=>{const o={dtype:s.dtype,shape:s.shape},i=vx(e,o,t),a=n.createShaderModule({code:i,label:t.constructor.name});let l=W().get("WEBGPU_PRINT_SHADER");if(l!==""){l=l.toLowerCase();const u=l.split(",");(l==="all"||u.some(c=>t.shaderKey.toLowerCase().includes(c)))&&(console.group(t.shaderKey),console.debug(i),console.groupEnd())}return r?n.createComputePipelineAsync({compute:{module:a,entryPoint:"_start"},label:t.constructor.name,layout:"auto"}):n.createComputePipeline({compute:{module:a,entryPoint:"_start"},label:t.constructor.name,layout:"auto"})},H=(n,t="f32")=>{switch(n){case 1:return`${t}`;case 2:return`vec2<${t}>`;case 3:return`vec3<${t}>`;case 4:return`vec4<${t}>`;default:throw new Error(`${n}-component ${t} is not supported.`)}};function Tt(n){if(n<=1)return"i32";if(n===2)return"vec2<i32>";if(n===3)return"vec3<i32>";if(n===4)return"vec4<i32>";if(n===5)return"vec5";if(n===6)return"vec6";throw Error(`GPU for rank ${n} is not yet supported`)}function In(n){if(n===0)return"x";if(n===1)return"y";if(n===2)return"z";if(n===3)return"w";if(n===4)return"u";if(n===5)return"v";throw Error(`Index ${n} is not yet supported`)}function wt(...n){let t;switch(n.length){case 0:t=`
         fn main()
       `;break;case 1:t=`
         fn main(${n[0]} : i32)
@@ -3258,7 +3258,7 @@ out_color[outIdx] = vec4f(
                 localIndex);
         `}
       }
-    `),e.pixelsOpType!=null){const p=e.pixelsOpType===to.FROM_PIXELS?`@group(0) @binding(0) var<storage, read_write> result: array<${ns(t.dtype,e.outputComponent)}>;`:`@group(0) @binding(1) var<storage, read> inBuf : array<${ns(n[0].dtype,e.outputComponent)}>;`,g=t.shape.length===3?"vec2<i32>":"i32";s.push(`
+    `),e.pixelsOpType!=null){const p=e.pixelsOpType===eo.FROM_PIXELS?`@group(0) @binding(0) var<storage, read_write> result: array<${ns(t.dtype,e.outputComponent)}>;`:`@group(0) @binding(1) var<storage, read> inBuf : array<${ns(n[0].dtype,e.outputComponent)}>;`,g=t.shape.length===3?"vec2<i32>":"i32";s.push(`
         struct Uniform {
           outShapeStrides : ${g},
           size            : i32,
@@ -3282,7 +3282,7 @@ out_color[outIdx] = vec4f(
       `);const u=_x(t.shape,e.dispatchLayout),c=[tf,s.join(`
 `)+Ax,ha(t.shape),u,Tx(t.shape.length)];e.atomic||c.push(Nx(t.shape,t.dtype,e.outputComponent)),e.variableNames.forEach((p,g)=>{c.push(`${ha(n[g].shape,p)}`)});const h=n.map((p,g)=>kx(p,t.shape,e.variableComponents?e.variableComponents[g]:e.outputComponent,e.dispatchLayout.x.length===t.shape.length)).join(`
 `);c.push(h),c.push(e.getUserCode());const f=nf(e);return c.push(Qh(f,e)),c.join(`
-`)}function Ix(n,t,e){let s=n.shaderKey;if(n.pixelsOpType!=null)return s;const r=[],o=[];t.forEach(c=>{r.push(c.shape),o.push(c.dtype)}),r.push(e.shape),o.push(e.dtype);const i=t.map(c=>or(c.shape,e.shape)),a=t.map(c=>Zt(c.shape,e.shape)).join("_"),l=i.map(c=>c.join("_")).join(";"),u=ef(n)?"flatDispatch":"";return s+="_"+(n.workgroupSize?n.workgroupSize.join(","):"")+r.map(c=>c.length).join(",")+o.join(",")+n.variableNames.join(",")+l+a+u,s}const tf=`
+`)}function Ix(n,t,e){let s=n.shaderKey;if(n.pixelsOpType!=null)return s;const r=[],o=[];t.forEach(c=>{r.push(c.shape),o.push(c.dtype)}),r.push(e.shape),o.push(e.dtype);const i=t.map(c=>ir(c.shape,e.shape)),a=t.map(c=>Zt(c.shape,e.shape)).join("_"),l=i.map(c=>c.join("_")).join(";"),u=ef(n)?"flatDispatch":"";return s+="_"+(n.workgroupSize?n.workgroupSize.join(","):"")+r.map(c=>c.length).join(",")+o.join(",")+n.variableNames.join(",")+l+a+u,s}const tf=`
   struct vec5 {x: i32, y: i32, z: i32, w: i32, u: i32};
   struct vec6 {x: i32, y: i32, z: i32, w: i32, u: i32, v: i32};
@@ -3336,10 +3336,10 @@ out_color[outIdx] = vec4f(
   fn isinf(val: f32) -> bool {
     return abs(val) == uniforms.INFINITY;
   }
-`;function ha(n,t=""){const e=n.length,s=t!==""?`get${t.charAt(0).toUpperCase()+t.slice(1)}CoordsFromIndex`:"getCoordsFromIndex",r=t!==""?`${t.charAt(0).toLowerCase()+t.slice(1)}ShapeStrides`:"outShapeStrides";if(e<=1)return`fn ${s}(index : i32) -> i32 { return index; }`;const o=jt(n),i=Tt(e),a=[];for(let u=0;u<e;u++)a.push(`d${u}`);if(o.length===1)return`    fn ${s}(index : i32) -> vec2<i32> {
+`;function ha(n,t=""){const e=n.length,s=t!==""?`get${t.charAt(0).toUpperCase()+t.slice(1)}CoordsFromIndex`:"getCoordsFromIndex",r=t!==""?`${t.charAt(0).toLowerCase()+t.slice(1)}ShapeStrides`:"outShapeStrides";if(e<=1)return`fn ${s}(index : i32) -> i32 { return index; }`;const o=Kt(n),i=Tt(e),a=[];for(let u=0;u<e;u++)a.push(`d${u}`);if(o.length===1)return`    fn ${s}(index : i32) -> vec2<i32> {
       let d0 = index / uniforms.${r}; let d1 = index - d0 * uniforms.${r};
       return vec2<i32>(d0, d1);
-    }`;let l;return l="var index2 = index;"+o.map((u,c)=>{const h=`let ${a[c]} = index2 / uniforms.${r}.${vn(c)}`,f=c===o.length-1?`let ${a[c+1]} = index2 - ${a[c]} * uniforms.${r}.${vn(c)}`:`index2 = index2 - ${a[c]} * uniforms.${r}.${vn(c)}`;return`${h}; ${f};`}).join(""),`
+    }`;let l;return l="var index2 = index;"+o.map((u,c)=>{const h=`let ${a[c]} = index2 / uniforms.${r}.${In(c)}`,f=c===o.length-1?`let ${a[c+1]} = index2 - ${a[c]} * uniforms.${r}.${In(c)}`:`index2 = index2 - ${a[c]} * uniforms.${r}.${In(c)}`;return`${h}; ${f};`}).join(""),`
     fn ${s}(index : i32) -> ${i} {
       ${l}
       return ${i}(${a.join(",")});
@@ -3361,7 +3361,7 @@ out_color[outIdx] = vec4f(
     fn ${i}Coords(coords : ${u}) -> ${H(e)} {
       return ${H(e)}(${r}[${l>1?"getOutputIndexFromCoords(coords)":"coords"}${e===1?"":` / ${e}`}]);
     }
-    `;const c=or(n.shape,t),h=l-a;let f="";if(a===0)return`
+    `;const c=ir(n.shape,t),h=l-a;let f="";if(a===0)return`
     fn ${i}Index(globalIndex : i32) -> ${H(e)}{
       return get${o}();
     }
@@ -3369,8 +3369,8 @@ out_color[outIdx] = vec4f(
     fn ${i}Coords(coords : ${u}) -> ${H(e)}{
       return get${o}();
     }
-  `;l<2&&c.length>=1?f="coords = 0;":f=c.map(m=>`coords.${vn(m+h)} = 0;`).join(`
-`);let d="";if(l<2&&a>0)d="coords";else if(l>1){const m=Tt(a),b=n.shape.map((y,S)=>`coords.${vn(S+h)}`).join(", ");d=`${m}(${b})`}else d="coords";const p=`uniforms.${r.charAt(0).toLowerCase()+r.slice(1)}Shape`,g=`${a}D`;return`
+  `;l<2&&c.length>=1?f="coords = 0;":f=c.map(m=>`coords.${In(m+h)} = 0;`).join(`
+`);let d="";if(l<2&&a>0)d="coords";else if(l>1){const m=Tt(a),b=n.shape.map((y,S)=>`coords.${In(S+h)}`).join(", ");d=`${m}(${b})`}else d="coords";const p=`uniforms.${r.charAt(0).toLowerCase()+r.slice(1)}Shape`,g=`${a}D`;return`
   fn ${i}Index(globalIndex : i32) -> ${H(e)} {
     var coords = getCoordsFromIndex(globalIndex);
     ${f}
@@ -3453,7 +3453,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const In=n=>{let t=1;for(let e=0;e<n.length;e++)t*=n[e];return t};function It(n,t,e=[1,1,1],s=[1,1,1]){const[r,o,i]=[Math.ceil(In(n.x.map(a=>t[a]))/(e[0]*s[0])),n.y?Math.ceil(In(n.y.map(a=>t[a]))/(e[1]*s[1])):1,n.z?Math.ceil(In(n.z.map(a=>t[a]))/(e[2]*s[2])):1];return[r,o,i]}function Rx(n,t,e,s=!1){const r=[8,8,1],o=[4,4,1];return s||(n<=8&&(o[1]=1),t<=16&&e<=16&&(r[0]=4)),{workgroupSize:r,elementsPerThread:o}}function Px(n,t,e=!1){if(e)return[8,8,1];const s=In(n.x.map(o=>t[o])),r=In(n.y.map(o=>t[o]));return s<=4?[4,16,1]:r<=4?[16,4,1]:[16,16,1]}function Lx(n,t,e=!1){if(e)return[4,4,1];const s=In(n.x.map(o=>t[o])),r=In(n.y.map(o=>t[o]));return s<=4?[1,2,1]:r<=4?[2,1,1]:[2,2,1]}function ae(n){return{x:n.map((t,e)=>e)}}function sf(n){if(n==="float32"||n==="int32"||n==="bool"||n==="string")return 4;if(n==="complex64")return 8;throw new Error(`Unknown dtype ${n}`)}function rf(){return!!(typeof globalThis<"u"&&globalThis.navigator&&globalThis.navigator.gpu)}var Ne;(function(n){n[n.MatMulReduceProgram=0]="MatMulReduceProgram",n[n.MatMulSplitKProgram=1]="MatMulSplitKProgram",n[n.MatMulSmallOutputSizeProgram=2]="MatMulSmallOutputSizeProgram",n[n.MatMulPackedProgram=3]="MatMulPackedProgram",n[n.MatMulMax=4]="MatMulMax"})(Ne||(Ne={}));/**
+ */const An=n=>{let t=1;for(let e=0;e<n.length;e++)t*=n[e];return t};function It(n,t,e=[1,1,1],s=[1,1,1]){const[r,o,i]=[Math.ceil(An(n.x.map(a=>t[a]))/(e[0]*s[0])),n.y?Math.ceil(An(n.y.map(a=>t[a]))/(e[1]*s[1])):1,n.z?Math.ceil(An(n.z.map(a=>t[a]))/(e[2]*s[2])):1];return[r,o,i]}function Rx(n,t,e,s=!1){const r=[8,8,1],o=[4,4,1];return s||(n<=8&&(o[1]=1),t<=16&&e<=16&&(r[0]=4)),{workgroupSize:r,elementsPerThread:o}}function Px(n,t,e=!1){if(e)return[8,8,1];const s=An(n.x.map(o=>t[o])),r=An(n.y.map(o=>t[o]));return s<=4?[4,16,1]:r<=4?[16,4,1]:[16,16,1]}function Lx(n,t,e=!1){if(e)return[4,4,1];const s=An(n.x.map(o=>t[o])),r=An(n.y.map(o=>t[o]));return s<=4?[1,2,1]:r<=4?[2,1,1]:[2,2,1]}function ae(n){return{x:n.map((t,e)=>e)}}function sf(n){if(n==="float32"||n==="int32"||n==="bool"||n==="string")return 4;if(n==="complex64")return 8;throw new Error(`Unknown dtype ${n}`)}function rf(){return!!(typeof globalThis<"u"&&globalThis.navigator&&globalThis.navigator.gpu)}var Ne;(function(n){n[n.MatMulReduceProgram=0]="MatMulReduceProgram",n[n.MatMulSplitKProgram=1]="MatMulSplitKProgram",n[n.MatMulSmallOutputSizeProgram=2]="MatMulSmallOutputSizeProgram",n[n.MatMulPackedProgram=3]="MatMulPackedProgram",n[n.MatMulMax=4]="MatMulMax"})(Ne||(Ne={}));/**
  * @license
  * Copyright 2019 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -3468,7 +3468,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const Mx=W().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"),Ox=(n,t)=>{const e=n.limits.maxComputeWorkgroupsPerDimension,s=t.dispatchLayout,r=t.dispatch;if(r.every(i=>i<=e))return r;w(r[0]>e&&s.y===void 0&&s.z===void 0,()=>"Dispatch size exceeds WebGPU limits in Y or Z dimension.");let o=Math.ceil(Math.sqrt(r[0]));return o>e?(o=Math.ceil(Math.cbrt(r[0])),w(o<=e,()=>"Total dispatch size exceeds WebGPU maximum."),[o,o,o]):[o,o,1]};class Bs extends ba{nextDataId(){return Bs.nextDataId++}constructor(t,e){if(super(),this.commandQueueOwnedIds=new WeakSet,this.dispatchCountInPass=0,this.disposed=!1,this.downloadWaitMs=0,this.tensorDataPendingDisposal=[],this.queryResolveBuffer=null,this.querySet=null,this.querySetCount=2,this.stagingPendingDisposal=[],this.uniformPendingDisposal=[],this.uploadWaitMs=0,this.hasReadSyncWarned=!1,this.hasTimestampQueryWarned=!1,!rf())throw new Error("WebGPU is not supported on this device");this.pipelineCache={},this.device=t,this.queue=t.queue,this.commandEncoder=null,this.computePassEncoder=null,this.adapterInfo=new gx(e),this.supportTimestampQuery=this.device.features.has("timestamp-query"),this.thresholdToIncreaseWorkgroups=this.adapterInfo.intelGPUGeneration>=12?16:8,this.bufferManager=new bx(this.device),this.textureManager=new yx(this.device),this.tensorMap=new vf(this,To()),W().getBool("WEBGPU_USE_PROFILE_TOOL")&&(this.dummyCanvas=document.createElement("canvas"),this.dummyCanvas.width=1,this.dummyCanvas.height=1,this.dummyContext=this.dummyCanvas.getContext("webgpu"),this.dummyContext.configure({device:t,format:"bgra8unorm"}),document.body.appendChild(this.dummyCanvas))}floatPrecision(){return 32}disposeData(t,e=!1){if(!this.tensorMap.has(t))return!0;const s=this.tensorMap.get(t);return e?s.refCount=0:s.refCount--,s.refCount>0?!1:(s.complexTensorInfos!=null&&(this.disposeData(s.complexTensorInfos.real.dataId),this.disposeData(s.complexTensorInfos.imag.dataId)),this.commandQueueOwnedIds.has(t)?(this.tensorDataPendingDisposal.push(t),!0):(this.releaseResource(t),this.tensorMap.delete(t),!0))}memory(){return{numBytesInGPU:this.bufferManager.numBytesUsed,numBytesAllocatedInGPU:this.bufferManager.numBytesAllocated,unreliable:!1}}releaseResource(t){const e=this.tensorMap.get(t);if(!(!e||!e.resource)){if(e.external){e.resource=null;return}e.resource instanceof GPUBuffer?this.bufferManager.releaseBuffer(e.resource):e.resource instanceof GPUTexture&&this.textureManager.releaseTexture(e.resource),e.resource=null}}refCount(t){return this.tensorMap.has(t)?this.tensorMap.get(t).refCount:0}incRef(t){const e=this.tensorMap.get(t);e.refCount++}decRef(t){if(this.tensorMap.has(t)){const e=this.tensorMap.get(t);e.refCount--}}write(t,e,s){if(s==="complex64"&&t!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");const r={id:this.nextDataId()};return this.tensorMap.set(r,{dtype:s,shape:e,values:t,refCount:1}),r}move(t,e,s,r,o){if(r==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.tensorMap.set(t,{dtype:r,shape:s,values:e,refCount:o})}submitQueue(){this.queue.submit([this.commandEncoder.finish()]),this.commandEncoder=null,this.dispatchCountInPass=0,this.commandQueueOwnedIds=new WeakSet,this.tensorDataPendingDisposal.forEach(t=>{this.releaseResource(t),this.tensorMap.delete(t)}),this.uniformPendingDisposal.forEach(t=>this.bufferManager.releaseBuffer(t)),this.stagingPendingDisposal.forEach(t=>this.bufferManager.releaseBuffer(t,!1)),this.tensorDataPendingDisposal=[],this.uniformPendingDisposal=[],this.stagingPendingDisposal=[]}ensureCommandEncoderReady(){this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder())}endComputePassEncoder(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}async checkCompileCompletionAsync(){let t;try{t=await Promise.all(Object.values(this.pipelineCache))}catch(e){throw new Error(e.message)}Object.keys(this.pipelineCache).map((e,s)=>{this.pipelineCache[e]=t[s]})}async getBufferData(t){if(W().getBool("WEBGPU_ENGINE_COMPILE_ONLY"))return console.warn("The data may be invalid since WEBGPU_ENGINE_COMPILE_ONLY is true, this can only be called when WEBGPU_ENGINE_COMPILE_ONLY is false"),null;const e=t.size,s=this.bufferManager.acquireBuffer(e,GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ);this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(t,0,s,0,e),this.submitQueue(),await s.mapAsync(GPUMapMode.READ);const r=s.getMappedRange().slice(0);return s.unmap(),s!=null&&this.bufferManager.releaseBuffer(s),W().getBool("WEBGPU_USE_PROFILE_TOOL")&&(w(this.dummyContext!==void 0,()=>"Fail to get context for profiling tool"),this.dummyContext.getCurrentTexture()),r}convertAndCacheOnCPU(t,e){const s=this.tensorMap.get(t);return s.values=e,s.values}readSync(t){const e=this.tensorMap.get(t),{values:s,complexTensorInfos:r}=e;if(s!=null||e.dtype==="string")return s;if(e.dtype==="complex64"){const g=this.readSync(r.real.dataId),m=this.readSync(r.imag.dataId),b=uo(Zl(g,m).buffer,"float32");return this.convertAndCacheOnCPU(t,b),b}this.hasReadSyncWarned||(this.hasReadSyncWarned=!0,console.warn("The performance of synchronously reading data from GPU to CPU is poor on the webgpu backend, please use asynchronous APIs instead."));const o=["opaque","premultiplied"],i=e.resource,a=i.size;w(a%4===0,()=>"Because there is 4 bytes for one pixel, buffer size must be multiple of 4.");const l=a/4,u=new ArrayBuffer(a),c=256,h=256,f=o.map(g=>new OffscreenCanvas(c,h)),d=new OffscreenCanvas(c,h);this.endComputePassEncoder(),f.map((g,m)=>{const b=g.getContext("webgpu");return b.configure({device:this.device,format:"bgra8unorm",usage:GPUTextureUsage.COPY_DST,alphaMode:o[m]}),b.getCurrentTexture()}).map((g,m)=>{const b=c*4,y=(_,T,P)=>{this.ensureCommandEncoderReady(),this.commandEncoder.copyBufferToTexture({buffer:i,bytesPerRow:b,offset:P},{texture:g},{width:_,height:T}),this.submitQueue();const B=d.getContext("2d",{willReadFrequently:!0});B.clearRect(0,0,_,T),B.drawImage(f[m],0,0);const Y=B.getImageData(0,0,_,T).data,j=o[m],F=new Uint8ClampedArray(u,P,_*T*4);for(let G=0;G<F.length;G+=4)if(j==="premultiplied")F[G+3]=Y[G+3];else{const q=Y[G];F[G]=Y[G+2],F[G+1]=Y[G+1],F[G+2]=q}},S=Math.floor(l/(c*h));let x=c,$=h,E=0;for(let _=0;_<S;_++)y(x,$,E),E+=c*h*4;const D=l%(c*h);$=Math.floor(D/c),$>0&&(y(x,$,E),E+=$*(c*4)),x=D%c,x>0&&y(x,1,E)});const p=uo(u,e.dtype);return this.convertAndCacheOnCPU(t,p),p}async read(t){if(!this.tensorMap.has(t))throw new Error(`Tensor ${t} was not registered!`);const e=this.tensorMap.get(t),{values:s}=e;if(s!=null)return s;let r;if(e.dtype==="complex64"){const o=await Promise.all([this.read(e.complexTensorInfos.real.dataId),this.read(e.complexTensorInfos.imag.dataId)]),i=o[0],a=o[1];r=Zl(i,a)}else{const o=await this.getBufferData(e.resource);r=uo(o,e.dtype)}return this.convertAndCacheOnCPU(t,r),r}copyBuffer(t){const e=t.size,s=t.usage,r=this.bufferManager.acquireBuffer(e,s);return this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(t,0,r,0,e),this.submitQueue(),r}createTensorFromGPUData(t,e,s){let r=t.buffer;if(s==="complex64")throw new Error("Cannot write to a complex64 dtype. ");const o={id:this.nextDataId()};this.tensorMap.set(o,{dtype:s,shape:e,values:null,refCount:1,external:t.zeroCopy});const i=this.tensorMap.get(o),a=sf(i.dtype)*z(i.shape);if(t.buffer.size<a)throw new Error(`GPUBuffer size(${t.buffer.size}) is smaller than tensor size(${a})!`);if((t.buffer.usage&(GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC))!==(GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC))throw new Error("GPUBuffer.usage should include GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC!");return t.zeroCopy!==!0&&(r=this.copyBuffer(r)),i.resource=r,To().makeTensorFromDataId(o,e,s,this)}readToGPU(t){const e=this.tensorMap.get(t),{values:s,dtype:r,shape:o,resource:i}=e;if(r==="complex64")throw new Error("Does not support reading buffer for complex64 dtype.");if(i==null)throw s!=null?new Error("Data is not on GPU but on CPU."):new Error("There is no data on GPU or CPU.");const a=i,l=a.size,u=a.usage,c=this.bufferManager.acquireBuffer(l,u);this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(i,0,c,0,l),this.submitQueue();const h=this.makeTensorInfo(o,r),f=To().makeTensorFromTensorInfo(h),d=this.tensorMap.get(h.dataId);return d.resource=c,{tensorRef:f,buffer:c}}bufferSync(t){const e=this.readSync(t.dataId);if(t.dtype==="string")try{const s=e.map(r=>Ks(r));return xt(t.shape,t.dtype,s)}catch{throw new Error("Failed to decode encoded string bytes into utf-8")}return xt(t.shape,t.dtype,e)}async time(t){!this.supportTimestampQuery&&!this.hasTimestampQueryWarned&&(console.warn("This device doesn't support timestamp-query extension. Start Chrome browser with flag --enable-dawn-features=allow_unsafe_apis to try it again. Otherwise, zero will be shown for the kernel time when profiling mode is enabled."),this.hasTimestampQueryWarned=!0);const e=this.activeTimers,s=[];let r=!1;this.programTimersStack==null?(this.programTimersStack=s,r=!0):this.activeTimers.push(s),this.activeTimers=s,t();const o=sn(this.activeTimers.map(u=>u.query)).filter(u=>u!=null),i=sn(this.activeTimers.map(u=>u.name)).filter(u=>u!=null);this.activeTimers=e,r&&(this.programTimersStack=null);const a={uploadWaitMs:this.uploadWaitMs,downloadWaitMs:this.downloadWaitMs,kernelMs:null,wallMs:null},l=await Promise.all(o);return a.kernelMs=Af(l),a.getExtraProfileInfo=()=>l.map((u,c)=>({name:i[c],ms:u})).map(u=>`${u.name}: ${u.ms}`).join(", "),this.uploadWaitMs=0,this.downloadWaitMs=0,a}makeTensorInfo(t,e,s){return e==="string"&&s!=null&&s.length>0&&Ws(s[0])&&(s=s.map(o=>nn(o))),{dataId:this.write(s,t,e),shape:t,dtype:e}}tensorToBinding(t){if(!t)return null;const s=this.tensorMap.get(t.dataId).resource;return s instanceof GPUBuffer?{buffer:s}:s instanceof GPUTexture?s.createView():s}uploadToGPU(t){const e=this.tensorMap.get(t);if(e.resource!=null)return;const s=sf(e.dtype)*z(e.shape);let r;const o=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST;if(e.values){if(r=this.bufferManager.acquireBuffer(s,o,!0),r.mapState==="unmapped"){const i=this.bufferManager.acquireBuffer(s,GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC,!0,!1),a=i.getMappedRange();e.dtype==="int32"||e.dtype==="bool"?new Int32Array(a).set(e.values):new Float32Array(a).set(e.values),i.unmap(),this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(i,0,r,0,s),this.stagingPendingDisposal.push(i)}else{const i=r.getMappedRange();e.dtype==="int32"||e.dtype==="bool"?new Int32Array(i).set(e.values):new Float32Array(i).set(e.values),r.unmap()}e.values=null}else r=this.bufferManager.acquireBuffer(s,o);e.resource=r}makeUniforms(t){let e=0,s=0;const r=[];let o=1;t.forEach(l=>{l.data.length===0&&(l.data=[1]);let u;switch(l.data.length){case 1:u=4;break;case 2:u=8;break;case 3:u=16;break;case 4:u=16;break;case 5:u=16;break;case 6:u=16;break;default:w(!1,()=>`Unsupported ${l.data.length}D shape`)}(s===5||s===6)&&(u=16),u>o&&(o=u),e=Math.ceil(e/u)*u,s=l.data.length,r.push(e),e+=l.data.length*4}),e=Math.ceil(e/o)*o;const i=new ArrayBuffer(e);t.forEach((l,u)=>{const c=r[u];l.type==="int32"?new Int32Array(i,c,l.data.length).set(l.data):l.type==="uint32"?new Uint32Array(i,c,l.data.length).set(l.data):new Float32Array(i,c,l.data.length).set(l.data)});const a=this.bufferManager.acquireBuffer(e,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);return this.queue.writeBuffer(a,0,i,0,e),this.uniformPendingDisposal.push(a),{offset:0,size:e,buffer:a}}runWebGPUProgram(t,e,s,r,o){if(o||(o=this.makeTensorInfo(t.outputShape,s)),z(o.shape)===0)return this.tensorMap.get(o.dataId).values=Cn(o.dtype,0),o;this.uploadToGPU(o.dataId),t.dispatch=Ox(this.device,t);const i=e.map((l,u)=>{if(l.dtype==="complex64")throw new Error("GPGPUProgram does not support complex64 input. For complex64 dtypes, please separate the program into real and imaginary parts.");return this.uploadToGPU(l.dataId),{dtype:this.tensorMap.get(l.dataId).dtype,shape:l.shape,name:t.variableNames[u]}});t.shaderKey=Ix(t,i,o);const a=W().getBool("WEBGPU_ENGINE_COMPILE_ONLY");return t.shaderKey in this.pipelineCache||(this.pipelineCache[t.shaderKey]=Sx(this.device,t,i,o,a)),t.pipeline=this.pipelineCache[t.shaderKey],a||this.recordAndSubmit(t,o,e,r),o}recordAndSubmit(t,e,s,r){if(t.pipeline instanceof Promise)throw new Error("Please call checkCompileCompletionAsync to ensure parallel compilation is done!");let o=[],i=[];const a="int32";if(t.pixelsOpType==null){o.push({type:"float32",data:[NaN]},{type:"float32",data:[1/0]}),i=s.concat(e).map(d=>d.shape);const f="int32";i.map(d=>{o.push({type:f,data:d});const p=jt(d);o.push({type:f,data:p})})}else{const f=jt(e.shape);o.push({type:a,data:f})}if(t.size){const f=z(t.outputShape);o.push({type:a,data:[t.outputComponent?f/t.outputComponent:f]})}r&&(o=[...o,...r]);const l=[this.tensorToBinding(e),...s.map(f=>this.tensorToBinding(f)),this.makeUniforms(o)];s.forEach(f=>{this.commandQueueOwnedIds.add(f.dataId)}),this.commandQueueOwnedIds.add(e.dataId);const u=this.device.createBindGroup({layout:t.pipeline.getBindGroupLayout(0),entries:l.map((f,d)=>({binding:d,resource:f}))}),c=this.activeTimers!=null;this.ensureCommandEncoderReady();const h={};c&&this.supportTimestampQuery?(this.endComputePassEncoder(),this.querySet==null&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.querySetCount})),h.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1},this.computePassEncoder=this.commandEncoder.beginComputePass(h)):this.computePassEncoder||(this.computePassEncoder=this.commandEncoder.beginComputePass(h)),this.computePassEncoder.setPipeline(t.pipeline),this.computePassEncoder.setBindGroup(0,u),this.computePassEncoder.dispatchWorkgroups(t.dispatch[0],t.dispatch[1],t.dispatch[2]),this.dispatchCountInPass++,(c||W().get("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE")<=this.dispatchCountInPass||t.pixelsOpType===to.DRAW)&&(this.endComputePassEncoder(),c?this.activeTimers.push({name:t.constructor.name,query:this.getQueryTime()}):this.submitQueue())}async getQueryTime(){if(!this.supportTimestampQuery)return 0;this.queryResolveBuffer==null&&(this.queryResolveBuffer=this.bufferManager.acquireBuffer(this.querySetCount*8,GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST|GPUBufferUsage.QUERY_RESOLVE)),this.commandEncoder.resolveQuerySet(this.querySet,0,this.querySetCount,this.queryResolveBuffer,0);const t=this.bufferManager.acquireBuffer(this.querySetCount*8,GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST);this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.querySetCount*8),this.submitQueue(),await t.mapAsync(GPUMapMode.READ);const e=new BigUint64Array(t.getMappedRange()),s=Number(e[1]-e[0])/1e6;return t.unmap(),this.bufferManager.releaseBuffer(t),s}shouldExecuteOnCPU(t,e=Mx){return W().getBool("WEBGPU_CPU_FORWARD")&&t.every(s=>this.tensorMap.get(s.dataId).resource==null&&z(s.shape)<e)}numDataIds(){return this.tensorMap.numDataIds()-this.tensorDataPendingDisposal.length}dispose(){this.disposed||(this.querySet!=null&&this.querySet.destroy(),this.bufferManager.dispose(),this.textureManager.dispose(),this.disposed=!0)}}Bs.nextDataId=0;/**
+ */const Mx=W().getNumber("WEBGPU_CPU_HANDOFF_SIZE_THRESHOLD"),Ox=(n,t)=>{const e=n.limits.maxComputeWorkgroupsPerDimension,s=t.dispatchLayout,r=t.dispatch;if(r.every(i=>i<=e))return r;w(r[0]>e&&s.y===void 0&&s.z===void 0,()=>"Dispatch size exceeds WebGPU limits in Y or Z dimension.");let o=Math.ceil(Math.sqrt(r[0]));return o>e?(o=Math.ceil(Math.cbrt(r[0])),w(o<=e,()=>"Total dispatch size exceeds WebGPU maximum."),[o,o,o]):[o,o,1]};class Fs extends ba{nextDataId(){return Fs.nextDataId++}constructor(t,e){if(super(),this.commandQueueOwnedIds=new WeakSet,this.dispatchCountInPass=0,this.disposed=!1,this.downloadWaitMs=0,this.tensorDataPendingDisposal=[],this.queryResolveBuffer=null,this.querySet=null,this.querySetCount=2,this.stagingPendingDisposal=[],this.uniformPendingDisposal=[],this.uploadWaitMs=0,this.hasReadSyncWarned=!1,this.hasTimestampQueryWarned=!1,!rf())throw new Error("WebGPU is not supported on this device");this.pipelineCache={},this.device=t,this.queue=t.queue,this.commandEncoder=null,this.computePassEncoder=null,this.adapterInfo=new gx(e),this.supportTimestampQuery=this.device.features.has("timestamp-query"),this.thresholdToIncreaseWorkgroups=this.adapterInfo.intelGPUGeneration>=12?16:8,this.bufferManager=new bx(this.device),this.textureManager=new yx(this.device),this.tensorMap=new vf(this,To()),W().getBool("WEBGPU_USE_PROFILE_TOOL")&&(this.dummyCanvas=document.createElement("canvas"),this.dummyCanvas.width=1,this.dummyCanvas.height=1,this.dummyContext=this.dummyCanvas.getContext("webgpu"),this.dummyContext.configure({device:t,format:"bgra8unorm"}),document.body.appendChild(this.dummyCanvas))}floatPrecision(){return 32}disposeData(t,e=!1){if(!this.tensorMap.has(t))return!0;const s=this.tensorMap.get(t);return e?s.refCount=0:s.refCount--,s.refCount>0?!1:(s.complexTensorInfos!=null&&(this.disposeData(s.complexTensorInfos.real.dataId),this.disposeData(s.complexTensorInfos.imag.dataId)),this.commandQueueOwnedIds.has(t)?(this.tensorDataPendingDisposal.push(t),!0):(this.releaseResource(t),this.tensorMap.delete(t),!0))}memory(){return{numBytesInGPU:this.bufferManager.numBytesUsed,numBytesAllocatedInGPU:this.bufferManager.numBytesAllocated,unreliable:!1}}releaseResource(t){const e=this.tensorMap.get(t);if(!(!e||!e.resource)){if(e.external){e.resource=null;return}e.resource instanceof GPUBuffer?this.bufferManager.releaseBuffer(e.resource):e.resource instanceof GPUTexture&&this.textureManager.releaseTexture(e.resource),e.resource=null}}refCount(t){return this.tensorMap.has(t)?this.tensorMap.get(t).refCount:0}incRef(t){const e=this.tensorMap.get(t);e.refCount++}decRef(t){if(this.tensorMap.has(t)){const e=this.tensorMap.get(t);e.refCount--}}write(t,e,s){if(s==="complex64"&&t!=null)throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");const r={id:this.nextDataId()};return this.tensorMap.set(r,{dtype:s,shape:e,values:t,refCount:1}),r}move(t,e,s,r,o){if(r==="complex64")throw new Error("Cannot write to a complex64 dtype. Please use tf.complex(real, imag).");this.tensorMap.set(t,{dtype:r,shape:s,values:e,refCount:o})}submitQueue(){this.queue.submit([this.commandEncoder.finish()]),this.commandEncoder=null,this.dispatchCountInPass=0,this.commandQueueOwnedIds=new WeakSet,this.tensorDataPendingDisposal.forEach(t=>{this.releaseResource(t),this.tensorMap.delete(t)}),this.uniformPendingDisposal.forEach(t=>this.bufferManager.releaseBuffer(t)),this.stagingPendingDisposal.forEach(t=>this.bufferManager.releaseBuffer(t,!1)),this.tensorDataPendingDisposal=[],this.uniformPendingDisposal=[],this.stagingPendingDisposal=[]}ensureCommandEncoderReady(){this.commandEncoder||(this.commandEncoder=this.device.createCommandEncoder())}endComputePassEncoder(){this.computePassEncoder&&(this.computePassEncoder.end(),this.computePassEncoder=null)}async checkCompileCompletionAsync(){let t;try{t=await Promise.all(Object.values(this.pipelineCache))}catch(e){throw new Error(e.message)}Object.keys(this.pipelineCache).map((e,s)=>{this.pipelineCache[e]=t[s]})}async getBufferData(t){if(W().getBool("WEBGPU_ENGINE_COMPILE_ONLY"))return console.warn("The data may be invalid since WEBGPU_ENGINE_COMPILE_ONLY is true, this can only be called when WEBGPU_ENGINE_COMPILE_ONLY is false"),null;const e=t.size,s=this.bufferManager.acquireBuffer(e,GPUBufferUsage.COPY_DST|GPUBufferUsage.MAP_READ);this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(t,0,s,0,e),this.submitQueue(),await s.mapAsync(GPUMapMode.READ);const r=s.getMappedRange().slice(0);return s.unmap(),s!=null&&this.bufferManager.releaseBuffer(s),W().getBool("WEBGPU_USE_PROFILE_TOOL")&&(w(this.dummyContext!==void 0,()=>"Fail to get context for profiling tool"),this.dummyContext.getCurrentTexture()),r}convertAndCacheOnCPU(t,e){const s=this.tensorMap.get(t);return s.values=e,s.values}readSync(t){const e=this.tensorMap.get(t),{values:s,complexTensorInfos:r}=e;if(s!=null||e.dtype==="string")return s;if(e.dtype==="complex64"){const g=this.readSync(r.real.dataId),m=this.readSync(r.imag.dataId),b=uo(Zl(g,m).buffer,"float32");return this.convertAndCacheOnCPU(t,b),b}this.hasReadSyncWarned||(this.hasReadSyncWarned=!0,console.warn("The performance of synchronously reading data from GPU to CPU is poor on the webgpu backend, please use asynchronous APIs instead."));const o=["opaque","premultiplied"],i=e.resource,a=i.size;w(a%4===0,()=>"Because there is 4 bytes for one pixel, buffer size must be multiple of 4.");const l=a/4,u=new ArrayBuffer(a),c=256,h=256,f=o.map(g=>new OffscreenCanvas(c,h)),d=new OffscreenCanvas(c,h);this.endComputePassEncoder(),f.map((g,m)=>{const b=g.getContext("webgpu");return b.configure({device:this.device,format:"bgra8unorm",usage:GPUTextureUsage.COPY_DST,alphaMode:o[m]}),b.getCurrentTexture()}).map((g,m)=>{const b=c*4,y=(_,T,P)=>{this.ensureCommandEncoderReady(),this.commandEncoder.copyBufferToTexture({buffer:i,bytesPerRow:b,offset:P},{texture:g},{width:_,height:T}),this.submitQueue();const B=d.getContext("2d",{willReadFrequently:!0});B.clearRect(0,0,_,T),B.drawImage(f[m],0,0);const Y=B.getImageData(0,0,_,T).data,j=o[m],F=new Uint8ClampedArray(u,P,_*T*4);for(let G=0;G<F.length;G+=4)if(j==="premultiplied")F[G+3]=Y[G+3];else{const q=Y[G];F[G]=Y[G+2],F[G+1]=Y[G+1],F[G+2]=q}},S=Math.floor(l/(c*h));let x=c,$=h,E=0;for(let _=0;_<S;_++)y(x,$,E),E+=c*h*4;const D=l%(c*h);$=Math.floor(D/c),$>0&&(y(x,$,E),E+=$*(c*4)),x=D%c,x>0&&y(x,1,E)});const p=uo(u,e.dtype);return this.convertAndCacheOnCPU(t,p),p}async read(t){if(!this.tensorMap.has(t))throw new Error(`Tensor ${t} was not registered!`);const e=this.tensorMap.get(t),{values:s}=e;if(s!=null)return s;let r;if(e.dtype==="complex64"){const o=await Promise.all([this.read(e.complexTensorInfos.real.dataId),this.read(e.complexTensorInfos.imag.dataId)]),i=o[0],a=o[1];r=Zl(i,a)}else{const o=await this.getBufferData(e.resource);r=uo(o,e.dtype)}return this.convertAndCacheOnCPU(t,r),r}copyBuffer(t){const e=t.size,s=t.usage,r=this.bufferManager.acquireBuffer(e,s);return this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(t,0,r,0,e),this.submitQueue(),r}createTensorFromGPUData(t,e,s){let r=t.buffer;if(s==="complex64")throw new Error("Cannot write to a complex64 dtype. ");const o={id:this.nextDataId()};this.tensorMap.set(o,{dtype:s,shape:e,values:null,refCount:1,external:t.zeroCopy});const i=this.tensorMap.get(o),a=sf(i.dtype)*z(i.shape);if(t.buffer.size<a)throw new Error(`GPUBuffer size(${t.buffer.size}) is smaller than tensor size(${a})!`);if((t.buffer.usage&(GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC))!==(GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC))throw new Error("GPUBuffer.usage should include GPUBufferUsage.STORAGE | GPUBufferUsage.COPY_SRC!");return t.zeroCopy!==!0&&(r=this.copyBuffer(r)),i.resource=r,To().makeTensorFromDataId(o,e,s,this)}readToGPU(t){const e=this.tensorMap.get(t),{values:s,dtype:r,shape:o,resource:i}=e;if(r==="complex64")throw new Error("Does not support reading buffer for complex64 dtype.");if(i==null)throw s!=null?new Error("Data is not on GPU but on CPU."):new Error("There is no data on GPU or CPU.");const a=i,l=a.size,u=a.usage,c=this.bufferManager.acquireBuffer(l,u);this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(i,0,c,0,l),this.submitQueue();const h=this.makeTensorInfo(o,r),f=To().makeTensorFromTensorInfo(h),d=this.tensorMap.get(h.dataId);return d.resource=c,{tensorRef:f,buffer:c}}bufferSync(t){const e=this.readSync(t.dataId);if(t.dtype==="string")try{const s=e.map(r=>Ys(r));return xt(t.shape,t.dtype,s)}catch{throw new Error("Failed to decode encoded string bytes into utf-8")}return xt(t.shape,t.dtype,e)}async time(t){!this.supportTimestampQuery&&!this.hasTimestampQueryWarned&&(console.warn("This device doesn't support timestamp-query extension. Start Chrome browser with flag --enable-dawn-features=allow_unsafe_apis to try it again. Otherwise, zero will be shown for the kernel time when profiling mode is enabled."),this.hasTimestampQueryWarned=!0);const e=this.activeTimers,s=[];let r=!1;this.programTimersStack==null?(this.programTimersStack=s,r=!0):this.activeTimers.push(s),this.activeTimers=s,t();const o=rn(this.activeTimers.map(u=>u.query)).filter(u=>u!=null),i=rn(this.activeTimers.map(u=>u.name)).filter(u=>u!=null);this.activeTimers=e,r&&(this.programTimersStack=null);const a={uploadWaitMs:this.uploadWaitMs,downloadWaitMs:this.downloadWaitMs,kernelMs:null,wallMs:null},l=await Promise.all(o);return a.kernelMs=Af(l),a.getExtraProfileInfo=()=>l.map((u,c)=>({name:i[c],ms:u})).map(u=>`${u.name}: ${u.ms}`).join(", "),this.uploadWaitMs=0,this.downloadWaitMs=0,a}makeTensorInfo(t,e,s){return e==="string"&&s!=null&&s.length>0&&Gs(s[0])&&(s=s.map(o=>sn(o))),{dataId:this.write(s,t,e),shape:t,dtype:e}}tensorToBinding(t){if(!t)return null;const s=this.tensorMap.get(t.dataId).resource;return s instanceof GPUBuffer?{buffer:s}:s instanceof GPUTexture?s.createView():s}uploadToGPU(t){const e=this.tensorMap.get(t);if(e.resource!=null)return;const s=sf(e.dtype)*z(e.shape);let r;const o=GPUBufferUsage.STORAGE|GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST;if(e.values){if(r=this.bufferManager.acquireBuffer(s,o,!0),r.mapState==="unmapped"){const i=this.bufferManager.acquireBuffer(s,GPUBufferUsage.MAP_WRITE|GPUBufferUsage.COPY_SRC,!0,!1),a=i.getMappedRange();e.dtype==="int32"||e.dtype==="bool"?new Int32Array(a).set(e.values):new Float32Array(a).set(e.values),i.unmap(),this.ensureCommandEncoderReady(),this.endComputePassEncoder(),this.commandEncoder.copyBufferToBuffer(i,0,r,0,s),this.stagingPendingDisposal.push(i)}else{const i=r.getMappedRange();e.dtype==="int32"||e.dtype==="bool"?new Int32Array(i).set(e.values):new Float32Array(i).set(e.values),r.unmap()}e.values=null}else r=this.bufferManager.acquireBuffer(s,o);e.resource=r}makeUniforms(t){let e=0,s=0;const r=[];let o=1;t.forEach(l=>{l.data.length===0&&(l.data=[1]);let u;switch(l.data.length){case 1:u=4;break;case 2:u=8;break;case 3:u=16;break;case 4:u=16;break;case 5:u=16;break;case 6:u=16;break;default:w(!1,()=>`Unsupported ${l.data.length}D shape`)}(s===5||s===6)&&(u=16),u>o&&(o=u),e=Math.ceil(e/u)*u,s=l.data.length,r.push(e),e+=l.data.length*4}),e=Math.ceil(e/o)*o;const i=new ArrayBuffer(e);t.forEach((l,u)=>{const c=r[u];l.type==="int32"?new Int32Array(i,c,l.data.length).set(l.data):l.type==="uint32"?new Uint32Array(i,c,l.data.length).set(l.data):new Float32Array(i,c,l.data.length).set(l.data)});const a=this.bufferManager.acquireBuffer(e,GPUBufferUsage.COPY_DST|GPUBufferUsage.UNIFORM);return this.queue.writeBuffer(a,0,i,0,e),this.uniformPendingDisposal.push(a),{offset:0,size:e,buffer:a}}runWebGPUProgram(t,e,s,r,o){if(o||(o=this.makeTensorInfo(t.outputShape,s)),z(o.shape)===0)return this.tensorMap.get(o.dataId).values=Cn(o.dtype,0),o;this.uploadToGPU(o.dataId),t.dispatch=Ox(this.device,t);const i=e.map((l,u)=>{if(l.dtype==="complex64")throw new Error("GPGPUProgram does not support complex64 input. For complex64 dtypes, please separate the program into real and imaginary parts.");return this.uploadToGPU(l.dataId),{dtype:this.tensorMap.get(l.dataId).dtype,shape:l.shape,name:t.variableNames[u]}});t.shaderKey=Ix(t,i,o);const a=W().getBool("WEBGPU_ENGINE_COMPILE_ONLY");return t.shaderKey in this.pipelineCache||(this.pipelineCache[t.shaderKey]=Sx(this.device,t,i,o,a)),t.pipeline=this.pipelineCache[t.shaderKey],a||this.recordAndSubmit(t,o,e,r),o}recordAndSubmit(t,e,s,r){if(t.pipeline instanceof Promise)throw new Error("Please call checkCompileCompletionAsync to ensure parallel compilation is done!");let o=[],i=[];const a="int32";if(t.pixelsOpType==null){o.push({type:"float32",data:[NaN]},{type:"float32",data:[1/0]}),i=s.concat(e).map(d=>d.shape);const f="int32";i.map(d=>{o.push({type:f,data:d});const p=Kt(d);o.push({type:f,data:p})})}else{const f=Kt(e.shape);o.push({type:a,data:f})}if(t.size){const f=z(t.outputShape);o.push({type:a,data:[t.outputComponent?f/t.outputComponent:f]})}r&&(o=[...o,...r]);const l=[this.tensorToBinding(e),...s.map(f=>this.tensorToBinding(f)),this.makeUniforms(o)];s.forEach(f=>{this.commandQueueOwnedIds.add(f.dataId)}),this.commandQueueOwnedIds.add(e.dataId);const u=this.device.createBindGroup({layout:t.pipeline.getBindGroupLayout(0),entries:l.map((f,d)=>({binding:d,resource:f}))}),c=this.activeTimers!=null;this.ensureCommandEncoderReady();const h={};c&&this.supportTimestampQuery?(this.endComputePassEncoder(),this.querySet==null&&(this.querySet=this.device.createQuerySet({type:"timestamp",count:this.querySetCount})),h.timestampWrites={querySet:this.querySet,beginningOfPassWriteIndex:0,endOfPassWriteIndex:1},this.computePassEncoder=this.commandEncoder.beginComputePass(h)):this.computePassEncoder||(this.computePassEncoder=this.commandEncoder.beginComputePass(h)),this.computePassEncoder.setPipeline(t.pipeline),this.computePassEncoder.setBindGroup(0,u),this.computePassEncoder.dispatchWorkgroups(t.dispatch[0],t.dispatch[1],t.dispatch[2]),this.dispatchCountInPass++,(c||W().get("WEBGPU_DEFERRED_SUBMIT_BATCH_SIZE")<=this.dispatchCountInPass||t.pixelsOpType===eo.DRAW)&&(this.endComputePassEncoder(),c?this.activeTimers.push({name:t.constructor.name,query:this.getQueryTime()}):this.submitQueue())}async getQueryTime(){if(!this.supportTimestampQuery)return 0;this.queryResolveBuffer==null&&(this.queryResolveBuffer=this.bufferManager.acquireBuffer(this.querySetCount*8,GPUBufferUsage.COPY_SRC|GPUBufferUsage.COPY_DST|GPUBufferUsage.QUERY_RESOLVE)),this.commandEncoder.resolveQuerySet(this.querySet,0,this.querySetCount,this.queryResolveBuffer,0);const t=this.bufferManager.acquireBuffer(this.querySetCount*8,GPUBufferUsage.MAP_READ|GPUBufferUsage.COPY_DST);this.commandEncoder.copyBufferToBuffer(this.queryResolveBuffer,0,t,0,this.querySetCount*8),this.submitQueue(),await t.mapAsync(GPUMapMode.READ);const e=new BigUint64Array(t.getMappedRange()),s=Number(e[1]-e[0])/1e6;return t.unmap(),this.bufferManager.releaseBuffer(t),s}shouldExecuteOnCPU(t,e=Mx){return W().getBool("WEBGPU_CPU_FORWARD")&&t.every(s=>this.tensorMap.get(s.dataId).resource==null&&z(s.shape)<e)}numDataIds(){return this.tensorMap.numDataIds()-this.tensorDataPendingDisposal.length}dispose(){this.disposed||(this.querySet!=null&&this.querySet.destroy(),this.bufferManager.dispose(),this.textureManager.dispose(),this.disposed=!0)}}Fs.nextDataId=0;/**
  * @license
  * Copyright 2022 Google Inc. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -3483,7 +3483,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */rf()&&Xp("webgpu",async()=>{const n={powerPreference:W().get("WEBGPU_USE_LOW_POWER_GPU")?"low-power":"high-performance"},t=await navigator.gpu.requestAdapter(n),e={},s=[];t.features.has("timestamp-query")&&s.push("timestamp-query"),t.features.has("bgra8unorm-storage")&&s.push(["bgra8unorm-storage"]),e.requiredFeatures=s;const r=t.limits;e.requiredLimits={maxComputeWorkgroupStorageSize:r.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.maxStorageBufferBindingSize,maxBufferSize:r.maxBufferSize,maxComputeWorkgroupSizeX:r.maxComputeWorkgroupSizeX,maxComputeInvocationsPerWorkgroup:r.maxComputeInvocationsPerWorkgroup};const o=await t.requestDevice(e),i=await t.requestAdapterInfo();return new Bs(o,i)},3);/**
+ */rf()&&Xp("webgpu",async()=>{const n={powerPreference:W().get("WEBGPU_USE_LOW_POWER_GPU")?"low-power":"high-performance"},t=await navigator.gpu.requestAdapter(n),e={},s=[];t.features.has("timestamp-query")&&s.push("timestamp-query"),t.features.has("bgra8unorm-storage")&&s.push(["bgra8unorm-storage"]),e.requiredFeatures=s;const r=t.limits;e.requiredLimits={maxComputeWorkgroupStorageSize:r.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.maxStorageBufferBindingSize,maxBufferSize:r.maxBufferSize,maxComputeWorkgroupSizeX:r.maxComputeWorkgroupSizeX,maxComputeInvocationsPerWorkgroup:r.maxComputeInvocationsPerWorkgroup};const o=await t.requestDevice(e),i=await t.requestAdapterInfo();return new Fs(o,i)},3);/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -3612,7 +3612,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function of(n){const{backend:t,attrs:e}=n,{shape:s,value:r}=e;let{dtype:o}=e;if(o=o||is(r),o==="string"){const i=bt(o,z(s));return i.fill(r),t.makeTensorInfo(s,o,i)}else{const i=new Gx(s),a=[{type:"float32",data:[r]}];return t.runWebGPUProgram(i,[],o,a)}}/**
+ */function of(n){const{backend:t,attrs:e}=n,{shape:s,value:r}=e;let{dtype:o}=e;if(o=o||as(r),o==="string"){const i=bt(o,z(s));return i.fill(r),t.makeTensorInfo(s,o,i)}else{const i=new Gx(s),a=[{type:"float32",data:[r]}];return t.runWebGPUProgram(i,[],o,a)}}/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -3657,7 +3657,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function Mt(n){return(t,e,s,r,o)=>{const i=zt(t,e),a=i.length,l=jt(i),u=z(i),c=Cn(o,u),h=t.length,f=e.length,d=jt(t),p=jt(e),g=or(t,i),m=or(e,i);if(g.length+m.length===0)for(let b=0;b<c.length;++b)c[b]=n(s[b%s.length],r[b%r.length]);else for(let b=0;b<c.length;++b){const y=ho(b,a,l),S=y.slice(-h);g.forEach(D=>S[D]=0);const x=co(S,h,d),$=y.slice(-f);m.forEach(D=>$[D]=0);const E=co($,f,p);c[b]=n(s[x],r[E])}return[c,i]}}/**
+ */function Mt(n){return(t,e,s,r,o)=>{const i=zt(t,e),a=i.length,l=Kt(i),u=z(i),c=Cn(o,u),h=t.length,f=e.length,d=Kt(t),p=Kt(e),g=ir(t,i),m=ir(e,i);if(g.length+m.length===0)for(let b=0;b<c.length;++b)c[b]=n(s[b%s.length],r[b%r.length]);else for(let b=0;b<c.length;++b){const y=ho(b,a,l),S=y.slice(-h);g.forEach(D=>S[D]=0);const x=co(S,h,d),$=y.slice(-f);m.forEach(D=>$[D]=0);const E=co($,f,p);c[b]=n(s[x],r[E])}return[c,i]}}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -3672,7 +3672,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function jx(n,t,e,s){if(s==="int32"){const r=Int32Array.from(n);return[t,"int32",r]}if(s==="bool"){const r=Hs([0],e),[o,i]=Mt((a,l)=>a!==l?1:0)(t,[],n,r,"bool");return[i,"bool",o]}throw new Error(`Error in Cast: failed to cast ${e} to ${s}`)}/**
+ */function jx(n,t,e,s){if(s==="int32"){const r=Int32Array.from(n);return[t,"int32",r]}if(s==="bool"){const r=Ks([0],e),[o,i]=Mt((a,l)=>a!==l?1:0)(t,[],n,r,"bool");return[i,"bool",o]}throw new Error(`Error in Cast: failed to cast ${e} to ${s}`)}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -4062,7 +4062,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function bS(n,t,e,s,r){const o=t.length,i=z(t),a=jt(t),l=jt(r),u=Cn(e,z(r));for(let c=0;c<i;++c){const h=ho(c,o,a),f=new Array(h.length);for(let p=0;p<f.length;p++)f[p]=h[s[p]];const d=co(f,o,l);u[d]=n[c]}return u}/**
+ */function bS(n,t,e,s,r){const o=t.length,i=z(t),a=Kt(t),l=Kt(r),u=Cn(e,z(r));for(let c=0;c<i;++c){const h=ho(c,o,a),f=new Array(h.length);for(let p=0;p<f.length;p++)f[p]=h[s[p]];const d=co(f,o,l);u[d]=n[c]}return u}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -4092,7 +4092,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function wS(n,t,e){n.forEach((s,r)=>{if(s<0||s>=e){const o=ho(r,t.length,jt(t)).join(",");throw new Error(`indices[${o}] = ${s} is not in [0, ${e})`)}})}function xS(n,t){for(let e=0;e<n.length;++e){const s=n[e],r=e===n.length-1?t:n[e+1].length;if(s.length===0)throw new Error("Ragged splits may not be empty");if(s[0]<0)throw new Error("Ragged splits must be non-negative");if(s[s.length-1]>r)throw new Error("Ragged splits must not point past values");for(let o=1;o<s.length;++o)if(s[o-1]>s[o])throw new Error("Ragged splits must be sorted in ascending order")}}function SS(n,t,e,s){const r=[];let o=0;const i=t.length-1+e.length,a=new Array(i).fill(null).map(()=>[0]);xS(e,s);let l=1;for(let u=0;u<t.length-1;++u){l*=t[u];const c=t[u+1];for(let h=1;h<l+1;++h)a[u].push(h*c)}for(let u=0;u<n.length;++u){let c=n[u],h=n[u]+1;for(let f=0;f<e.length;++f){const d=e[f],p=f+t.length-1;if(p>=0){const g=a[p],m=g[g.length-1]-d[c];for(let b=c;b<h;++b)a[p].push(d[b+1]+m)}c=d[c],h=d[h]}h!==c&&(r.push([c,h]),o+=h-c)}return{outSplits:a,valueSlices:r,numValues:o}}function $S(n){const t=[];for(let e=0;e<n.length;++e){const s=n[e].length,r=bt("int32",s);t.push(r),n[e].forEach((o,i)=>r[i]=o)}return t}function lf(n,t){const e=n.slice(0,t);for(;e.length<t;)e.push(1);for(let s=t;s<n.length;s++)e[t-1]*=n[s];return e}function vS(n,t,e,s,r,o){const i=lf(t,2)[1],a=lf(o,2)[1];let l=0;for(const u of e)for(let c=u[0];c<u[1];++c){for(let h=0;h<s;++h)r[l*a+h]=n[c*i+h];++l}}function IS(n,t,e,s,r){const o=t.slice();o[0]=r;const i=bt(e,z(o)),a=n.length,l=a===0?0:a/t[0];return vS(n,t,s,l,i,o),[i,o]}function AS(n,t,e,s,r,o,i,a){if(n.length===0)throw new Error("paramsNestedSplits must be non empty");if(t[0].length===0)throw new Error("Split tensors must not be scalars");const l=t[0][0]-1;if(wS(o,i,l),s.length===0)throw new Error("params.rank must be nonzero");const u=s[0],{outSplits:c,valueSlices:h,numValues:f}=SS(o,i,n,u),d=$S(c),p=IS(e,s,r,h,f);return[d,p[0],p[1]]}/**
+ */function wS(n,t,e){n.forEach((s,r)=>{if(s<0||s>=e){const o=ho(r,t.length,Kt(t)).join(",");throw new Error(`indices[${o}] = ${s} is not in [0, ${e})`)}})}function xS(n,t){for(let e=0;e<n.length;++e){const s=n[e],r=e===n.length-1?t:n[e+1].length;if(s.length===0)throw new Error("Ragged splits may not be empty");if(s[0]<0)throw new Error("Ragged splits must be non-negative");if(s[s.length-1]>r)throw new Error("Ragged splits must not point past values");for(let o=1;o<s.length;++o)if(s[o-1]>s[o])throw new Error("Ragged splits must be sorted in ascending order")}}function SS(n,t,e,s){const r=[];let o=0;const i=t.length-1+e.length,a=new Array(i).fill(null).map(()=>[0]);xS(e,s);let l=1;for(let u=0;u<t.length-1;++u){l*=t[u];const c=t[u+1];for(let h=1;h<l+1;++h)a[u].push(h*c)}for(let u=0;u<n.length;++u){let c=n[u],h=n[u]+1;for(let f=0;f<e.length;++f){const d=e[f],p=f+t.length-1;if(p>=0){const g=a[p],m=g[g.length-1]-d[c];for(let b=c;b<h;++b)a[p].push(d[b+1]+m)}c=d[c],h=d[h]}h!==c&&(r.push([c,h]),o+=h-c)}return{outSplits:a,valueSlices:r,numValues:o}}function $S(n){const t=[];for(let e=0;e<n.length;++e){const s=n[e].length,r=bt("int32",s);t.push(r),n[e].forEach((o,i)=>r[i]=o)}return t}function lf(n,t){const e=n.slice(0,t);for(;e.length<t;)e.push(1);for(let s=t;s<n.length;s++)e[t-1]*=n[s];return e}function vS(n,t,e,s,r,o){const i=lf(t,2)[1],a=lf(o,2)[1];let l=0;for(const u of e)for(let c=u[0];c<u[1];++c){for(let h=0;h<s;++h)r[l*a+h]=n[c*i+h];++l}}function IS(n,t,e,s,r){const o=t.slice();o[0]=r;const i=bt(e,z(o)),a=n.length,l=a===0?0:a/t[0];return vS(n,t,s,l,i,o),[i,o]}function AS(n,t,e,s,r,o,i,a){if(n.length===0)throw new Error("paramsNestedSplits must be non empty");if(t[0].length===0)throw new Error("Split tensors must not be scalars");const l=t[0][0]-1;if(wS(o,i,l),s.length===0)throw new Error("params.rank must be nonzero");const u=s[0],{outSplits:c,valueSlices:h,numValues:f}=SS(o,i,n,u),d=$S(c),p=IS(e,s,r,h,f);return[d,p[0],p[1]]}/**
  * @license
  * Copyright 2022 Google LLC.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -4122,7 +4122,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */var le=Ie;class eo{constructor(t,e,s,r,o,i,a,l,u,c){this.shape=t,this.shapeShape=e,this.values=s,this.valuesShape=r,this.valuesDType=o,this.defaultValue=i,this.defaultValueShape=a,this.rowPartitionValues=l,this.rowPartitionValuesShapes=u,this.rowPartitionTypes=hy(c),this.raggedRank=fy(this.rowPartitionTypes)}getRowPartitionTypeByDimension(t){return this.rowPartitionTypes[0]===le.FIRST_DIM_SIZE?this.rowPartitionTypes[t+1]:this.rowPartitionTypes[t]}getRowPartitionTensor(t){return this.rowPartitionTypes[0]===le.FIRST_DIM_SIZE?this.rowPartitionValues[t+1]:this.rowPartitionValues[t]}getMaxWidth(t){const e=this.getRowPartitionTensor(t-1);switch(this.getRowPartitionTypeByDimension(t-1)){case le.VALUE_ROWIDS:return eo.getMaxWidthValueRowID(e);case le.ROW_SPLITS:return eo.getMaxWidthRowSplit(e);default:throw new Error(`Cannot handle partition type ${le[this.getRowPartitionTypeByDimension(t-1)]}`)}}static getMaxWidthRowSplit(t){const e=t.length;if(e===0||e===1)return 0;let s=0;for(let r=0;r<e-1;++r){const o=t[r+1]-t[r];o>s&&(s=o)}return s}static getMaxWidthValueRowID(t){const e=t.length;if(e===0)return 0;let s=0,r=t[0],o=0;for(let i=1;i<e;++i){const a=t[i];a!==r&&(r=a,o=Math.max(i-s,o),s=i)}return Math.max(e-s,o)}tensorShapeFromTensor(t,e,s=!0){if(e.length===0){if(t[0]===-1)return[];throw new Error("The only valid scalar shape tensor is the fully unknown shape specified as -1.")}return hf(t,s)}calculateOutputSize(t){const e=this.valuesShape,s=this.defaultValueShape;dy(s,e);const r=this.tensorShapeFromTensor(this.shape,this.shapeShape),i=cy(this.raggedRank,r,e);i[0]<0&&(i[0]=t);for(let a=1;a<=this.raggedRank;++a)i[a]<0&&(i[a]=this.getMaxWidth(a));return i}calculateFirstParentOutputIndex(t,e,s){const r=Math.min(t,s),o=[];let i=0;for(let a=0;a<r;++a,i+=e)o.push(i);for(let a=r;a<t;++a)o.push(-1);return w(o.length===t,()=>"Final length of result must be equal to firstDimension."),o}calculateOutputIndexRowSplit(t,e,s,r){const o=t.length,i=[];for(let a=0;a<o-1;++a){const l=t[a+1]-t[a];let u=Math.min(r,l),c=e[a];c===-1&&(u=0);for(let h=0;h<u;++h)i.push(c),c+=s;for(let h=0;h<l-u;++h)i.push(-1)}if(o>0&&i.length!==t[o-1])throw new Error("Invalid row split size.");return i}calculateOutputIndexValueRowID(t,e,s,r){const o=t.length,i=[];if(o===0)return[];let a=0,l=t[0];if(l>=e.length)throw new Error(`Got currentValueRowId=${l}, which is not less than ${e.length}`);let u=e[l];i.push(u);for(let c=1;c<o;++c){const h=t[c];if(h===l)u>=0&&(++a,a<r?u+=s:u=-1);else{if(a=0,l=h,h>=e.length)throw new Error(`Got nextValueRowId=${h} which is not less than ${e.length}`);u=e[h]}i.push(u)}if(i.length!==t.length)throw new Error("Invalid row ids.");return i}calculateOutputIndex(t,e,s,r){const o=this.getRowPartitionTensor(t),i=this.getRowPartitionTypeByDimension(t);switch(i){case le.VALUE_ROWIDS:return this.calculateOutputIndexValueRowID(o,e,s,r);case le.ROW_SPLITS:if(o.length-1>e.length)throw new Error(`Row partition size is greater than output size: ${o.length-1} > ${e.length}`);return this.calculateOutputIndexRowSplit(o,e,s,r);default:throw new Error(`Unsupported partition type: ${le[i]}`)}}getFirstDimensionSize(){const t=this.rowPartitionValues[0];if(this.rowPartitionTypes.length===0)throw new Error("No row_partition_types given.");const e=this.rowPartitionTypes[0];switch(e){case le.FIRST_DIM_SIZE:return t[0];case le.VALUE_ROWIDS:throw new Error("Cannot handle VALUE_ROWIDS in first dimension.");case le.ROW_SPLITS:return this.rowPartitionValuesShapes[0][0]-1;default:throw new Error(`Cannot handle type ${le[e]}`)}}compute(){if(this.rowPartitionValues[0].length<=0)throw new Error("Invalid first partition input. Tensor requires at least one element.");const e=this.getFirstDimensionSize(),s=this.calculateOutputSize(e),r=new Array(this.raggedRank+1);r[r.length-1]=1;for(let l=r.length-2;l>=0;--l)r[l]=r[l+1]*s[l+1];const o=hf(s,!1),i=bt(this.valuesDType,z(o));if(r[0]*s[0]>0){let l=this.calculateFirstParentOutputIndex(e,r[0],s[0]);for(let u=1;u<=this.raggedRank;++u)l=this.calculateOutputIndex(u-1,l,r[u],s[u]);this.setOutput(this.raggedRank,l,i,o)}return[o,i]}setOutput(t,e,s,r){if(s.length===0)return;const o=this.values,i=s;let a=r.slice();a=a.slice(t+1);const l=z(a),u=e.length;let c=this.defaultValue;if(c.length!==l&&c.length!==1){const p=this.defaultValueShape;k(()=>{const g=L(c,p);c=sr(g,a).dataSync()})}let h=0,f=0,d=0;for(let p=0;p<=u;++p){let g=p<u?e[p]:-1;if(g===d){++d;continue}if(f<d){const m=o.subarray(h*l),b=i.subarray(f*l),y=(d-f)*l;cf(b,m,y)}if(p>=u){const m=s.length;g=Math.floor(m/l)}if(g>d)if(this.defaultValue.length===1)i.subarray(d*l,g*l).fill(this.defaultValue[0]),d=g;else for(;g>d;){const m=i.slice(d*l);cf(m,c,l),++d}g<0?(h=p+1,f=d):(h=p,f=d,d=f+1)}}}function cf(n,t,e){for(let s=0;s<e;s++)n[s]=t[s]}function hf(n,t){const e=[];for(let s of n){if(s<0){if(!t)throw new Error(`Dimension ${s} must be >= 0`);if(s<-1)throw new Error(`Dimension ${s} must be >= -1`);s=-1}e.push(s)}return e}function CS(n,t,e,s,r,o,i,a,l,u){return new eo(n,t,e,s,r,o,i,a,l,u).compute()}/**
+ */var le=Ie;class no{constructor(t,e,s,r,o,i,a,l,u,c){this.shape=t,this.shapeShape=e,this.values=s,this.valuesShape=r,this.valuesDType=o,this.defaultValue=i,this.defaultValueShape=a,this.rowPartitionValues=l,this.rowPartitionValuesShapes=u,this.rowPartitionTypes=hy(c),this.raggedRank=fy(this.rowPartitionTypes)}getRowPartitionTypeByDimension(t){return this.rowPartitionTypes[0]===le.FIRST_DIM_SIZE?this.rowPartitionTypes[t+1]:this.rowPartitionTypes[t]}getRowPartitionTensor(t){return this.rowPartitionTypes[0]===le.FIRST_DIM_SIZE?this.rowPartitionValues[t+1]:this.rowPartitionValues[t]}getMaxWidth(t){const e=this.getRowPartitionTensor(t-1);switch(this.getRowPartitionTypeByDimension(t-1)){case le.VALUE_ROWIDS:return no.getMaxWidthValueRowID(e);case le.ROW_SPLITS:return no.getMaxWidthRowSplit(e);default:throw new Error(`Cannot handle partition type ${le[this.getRowPartitionTypeByDimension(t-1)]}`)}}static getMaxWidthRowSplit(t){const e=t.length;if(e===0||e===1)return 0;let s=0;for(let r=0;r<e-1;++r){const o=t[r+1]-t[r];o>s&&(s=o)}return s}static getMaxWidthValueRowID(t){const e=t.length;if(e===0)return 0;let s=0,r=t[0],o=0;for(let i=1;i<e;++i){const a=t[i];a!==r&&(r=a,o=Math.max(i-s,o),s=i)}return Math.max(e-s,o)}tensorShapeFromTensor(t,e,s=!0){if(e.length===0){if(t[0]===-1)return[];throw new Error("The only valid scalar shape tensor is the fully unknown shape specified as -1.")}return hf(t,s)}calculateOutputSize(t){const e=this.valuesShape,s=this.defaultValueShape;dy(s,e);const r=this.tensorShapeFromTensor(this.shape,this.shapeShape),i=cy(this.raggedRank,r,e);i[0]<0&&(i[0]=t);for(let a=1;a<=this.raggedRank;++a)i[a]<0&&(i[a]=this.getMaxWidth(a));return i}calculateFirstParentOutputIndex(t,e,s){const r=Math.min(t,s),o=[];let i=0;for(let a=0;a<r;++a,i+=e)o.push(i);for(let a=r;a<t;++a)o.push(-1);return w(o.length===t,()=>"Final length of result must be equal to firstDimension."),o}calculateOutputIndexRowSplit(t,e,s,r){const o=t.length,i=[];for(let a=0;a<o-1;++a){const l=t[a+1]-t[a];let u=Math.min(r,l),c=e[a];c===-1&&(u=0);for(let h=0;h<u;++h)i.push(c),c+=s;for(let h=0;h<l-u;++h)i.push(-1)}if(o>0&&i.length!==t[o-1])throw new Error("Invalid row split size.");return i}calculateOutputIndexValueRowID(t,e,s,r){const o=t.length,i=[];if(o===0)return[];let a=0,l=t[0];if(l>=e.length)throw new Error(`Got currentValueRowId=${l}, which is not less than ${e.length}`);let u=e[l];i.push(u);for(let c=1;c<o;++c){const h=t[c];if(h===l)u>=0&&(++a,a<r?u+=s:u=-1);else{if(a=0,l=h,h>=e.length)throw new Error(`Got nextValueRowId=${h} which is not less than ${e.length}`);u=e[h]}i.push(u)}if(i.length!==t.length)throw new Error("Invalid row ids.");return i}calculateOutputIndex(t,e,s,r){const o=this.getRowPartitionTensor(t),i=this.getRowPartitionTypeByDimension(t);switch(i){case le.VALUE_ROWIDS:return this.calculateOutputIndexValueRowID(o,e,s,r);case le.ROW_SPLITS:if(o.length-1>e.length)throw new Error(`Row partition size is greater than output size: ${o.length-1} > ${e.length}`);return this.calculateOutputIndexRowSplit(o,e,s,r);default:throw new Error(`Unsupported partition type: ${le[i]}`)}}getFirstDimensionSize(){const t=this.rowPartitionValues[0];if(this.rowPartitionTypes.length===0)throw new Error("No row_partition_types given.");const e=this.rowPartitionTypes[0];switch(e){case le.FIRST_DIM_SIZE:return t[0];case le.VALUE_ROWIDS:throw new Error("Cannot handle VALUE_ROWIDS in first dimension.");case le.ROW_SPLITS:return this.rowPartitionValuesShapes[0][0]-1;default:throw new Error(`Cannot handle type ${le[e]}`)}}compute(){if(this.rowPartitionValues[0].length<=0)throw new Error("Invalid first partition input. Tensor requires at least one element.");const e=this.getFirstDimensionSize(),s=this.calculateOutputSize(e),r=new Array(this.raggedRank+1);r[r.length-1]=1;for(let l=r.length-2;l>=0;--l)r[l]=r[l+1]*s[l+1];const o=hf(s,!1),i=bt(this.valuesDType,z(o));if(r[0]*s[0]>0){let l=this.calculateFirstParentOutputIndex(e,r[0],s[0]);for(let u=1;u<=this.raggedRank;++u)l=this.calculateOutputIndex(u-1,l,r[u],s[u]);this.setOutput(this.raggedRank,l,i,o)}return[o,i]}setOutput(t,e,s,r){if(s.length===0)return;const o=this.values,i=s;let a=r.slice();a=a.slice(t+1);const l=z(a),u=e.length;let c=this.defaultValue;if(c.length!==l&&c.length!==1){const p=this.defaultValueShape;k(()=>{const g=L(c,p);c=rr(g,a).dataSync()})}let h=0,f=0,d=0;for(let p=0;p<=u;++p){let g=p<u?e[p]:-1;if(g===d){++d;continue}if(f<d){const m=o.subarray(h*l),b=i.subarray(f*l),y=(d-f)*l;cf(b,m,y)}if(p>=u){const m=s.length;g=Math.floor(m/l)}if(g>d)if(this.defaultValue.length===1)i.subarray(d*l,g*l).fill(this.defaultValue[0]),d=g;else for(;g>d;){const m=i.slice(d*l);cf(m,c,l),++d}g<0?(h=p+1,f=d):(h=p,f=d,d=f+1)}}}function cf(n,t,e){for(let s=0;s<e;s++)n[s]=t[s]}function hf(n,t){const e=[];for(let s of n){if(s<0){if(!t)throw new Error(`Dimension ${s} must be >= 0`);if(s<-1)throw new Error(`Dimension ${s} must be >= -1`);s=-1}e.push(s)}return e}function CS(n,t,e,s,r,o,i,a,l,u){return new no(n,t,e,s,r,o,i,a,l,u).compute()}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -4167,7 +4167,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function TS(n,t,e,s,r,o,i,a,l,u){const c=[s/r,r],h=n.values,f=t.values;if(s===0)return xt(e,t.dtype);const d=l instanceof Xs?l:xt(c,t.dtype);typeof l=="string"||typeof l=="number"?d.values.fill(l):typeof l=="boolean"&&d.values.fill(+l);for(let p=0;p<o;p++){const g=[];let m=0;for(let b=0;b<i;b++){const y=h[p*i+b];g.push(y),m+=y*a[b]}if(m<0||m>=s/r)throw new Error(`Invalid indices: ${g} does not index into ${e}`);for(let b=0;b<r;b++)u?d.values[m*r+b]+=f[p*r+b]:d.values[m*r+b]=t.rank===0?f[0]:f[p*r+b]}return d}/**
+ */function TS(n,t,e,s,r,o,i,a,l,u){const c=[s/r,r],h=n.values,f=t.values;if(s===0)return xt(e,t.dtype);const d=l instanceof Js?l:xt(c,t.dtype);typeof l=="string"||typeof l=="number"?d.values.fill(l):typeof l=="boolean"&&d.values.fill(+l);for(let p=0;p<o;p++){const g=[];let m=0;for(let b=0;b<i;b++){const y=h[p*i+b];g.push(y),m+=y*a[b]}if(m<0||m>=s/r)throw new Error(`Invalid indices: ${g} does not index into ${e}`);for(let b=0;b<r;b++)u?d.values[m*r+b]+=f[p*r+b]:d.values[m*r+b]=t.rank===0?f[0]:f[p*r+b]}return d}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the License);
@@ -4197,7 +4197,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function DS(n,t,e,s,r){const o=sy(s,t,e),i=z(e),a=jt(s);if(o){const h=ry(t,a);return r==="string"?n.slice(h,h+i):n.subarray(h,h+i)}const l=r==="string"?tu(n):n,u=xt(s,r,l),c=xt(e,r);for(let h=0;h<c.size;++h){const f=c.indexToLoc(h),d=f.map((p,g)=>p+t[g]);c.set(u.get(...d),...f)}return r==="string"?Ry(c.values):c.values}/**
+ */function DS(n,t,e,s,r){const o=sy(s,t,e),i=z(e),a=Kt(s);if(o){const h=ry(t,a);return r==="string"?n.slice(h,h+i):n.subarray(h,h+i)}const l=r==="string"?tu(n):n,u=xt(s,r,l),c=xt(e,r);for(let h=0;h<c.size;++h){const f=c.indexToLoc(h),d=f.map((p,g)=>p+t[g]);c.set(u.get(...d),...f)}return r==="string"?Ry(c.values):c.values}/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -4317,7 +4317,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class zS{constructor(t,e,s,r,o,i){this.separator=nn(t),this.nGramWidths=e,this.leftPad=nn(s),this.rightPad=nn(r),this.padWidth=o,this.preserveShort=i}getPadWidth(t){return Math.min(this.padWidth<0?t-1:this.padWidth,t-1)}getNumNGrams(t,e){const s=this.getPadWidth(e);return Math.max(0,t+2*s-e+1)}createNGrams(t,e,s,r,o,i){for(let a=0;a<o;++a){const l=this.getPadWidth(i),u=Math.max(0,l-a),c=Math.max(0,l-(o-(a+1))),h=i-(u+c),f=e+(u>0?0:a-l);let d=0;d+=u*this.leftPad.length;for(let y=0;y<h;++y)d+=t[f+y].length;d+=c*this.rightPad.length;const p=u+c+h-1;d+=p*this.separator.length,s[r+a]=new Uint8Array(d);const g=s[r+a];let m=0;const b=y=>y.forEach(S=>g[m++]=S);for(let y=0;y<u;++y)b(this.leftPad),b(this.separator);for(let y=0;y<h-1;++y)b(t[f+y]),b(this.separator);if(h>0){b(t[f+h-1]);for(let y=0;y<c;++y)b(this.separator),b(this.rightPad)}else{for(let y=0;y<c-1;++y)b(this.rightPad),b(this.separator);b(this.rightPad)}}}compute(t,e){const s=t.length,r=e.length;if(r>0){let l=e[0];if(l!==0)throw new Error(`First split value must be 0, got ${l}`);for(let u=1;u<r;++u){let c=e[u]>=l;if(c=c&&e[u]<=s,!c)throw new Error(`Invalid split value ${e[u]}, must be in [${l}, ${s}]`);l=e[u]}if(l!==s)throw new Error(`Last split value must be data size. Expected ${s}, got ${l}`)}const o=r-1,i=bt("int32",r);if(s===0||r===0){const l=new Array(s);for(let u=0;u<=o;++u)i[u]=0;return[l,i]}i[0]=0;for(let l=1;l<=o;++l){const u=e[l]-e[l-1];let c=0;this.nGramWidths.forEach(h=>{c+=this.getNumNGrams(u,h)}),this.preserveShort&&u>0&&c===0&&(c=1),i[l]=i[l-1]+c}const a=new Array(i[o]);for(let l=0;l<o;++l){const u=e[l];let c=i[l];if(this.nGramWidths.forEach(h=>{const f=e[l+1]-e[l],d=this.getNumNGrams(f,h);this.createNGrams(t,u,a,c,d,h),c+=d}),this.preserveShort&&c===i[l]){const h=e[l+1]-e[l];if(h===0)continue;const f=h+2*this.padWidth;this.createNGrams(t,u,a,c,1,f)}}return[a,i]}}function US(n,t,e,s,r,o,i,a){return new zS(e,s,r,o,i,a).compute(n,t)}/**
+ */class zS{constructor(t,e,s,r,o,i){this.separator=sn(t),this.nGramWidths=e,this.leftPad=sn(s),this.rightPad=sn(r),this.padWidth=o,this.preserveShort=i}getPadWidth(t){return Math.min(this.padWidth<0?t-1:this.padWidth,t-1)}getNumNGrams(t,e){const s=this.getPadWidth(e);return Math.max(0,t+2*s-e+1)}createNGrams(t,e,s,r,o,i){for(let a=0;a<o;++a){const l=this.getPadWidth(i),u=Math.max(0,l-a),c=Math.max(0,l-(o-(a+1))),h=i-(u+c),f=e+(u>0?0:a-l);let d=0;d+=u*this.leftPad.length;for(let y=0;y<h;++y)d+=t[f+y].length;d+=c*this.rightPad.length;const p=u+c+h-1;d+=p*this.separator.length,s[r+a]=new Uint8Array(d);const g=s[r+a];let m=0;const b=y=>y.forEach(S=>g[m++]=S);for(let y=0;y<u;++y)b(this.leftPad),b(this.separator);for(let y=0;y<h-1;++y)b(t[f+y]),b(this.separator);if(h>0){b(t[f+h-1]);for(let y=0;y<c;++y)b(this.separator),b(this.rightPad)}else{for(let y=0;y<c-1;++y)b(this.rightPad),b(this.separator);b(this.rightPad)}}}compute(t,e){const s=t.length,r=e.length;if(r>0){let l=e[0];if(l!==0)throw new Error(`First split value must be 0, got ${l}`);for(let u=1;u<r;++u){let c=e[u]>=l;if(c=c&&e[u]<=s,!c)throw new Error(`Invalid split value ${e[u]}, must be in [${l}, ${s}]`);l=e[u]}if(l!==s)throw new Error(`Last split value must be data size. Expected ${s}, got ${l}`)}const o=r-1,i=bt("int32",r);if(s===0||r===0){const l=new Array(s);for(let u=0;u<=o;++u)i[u]=0;return[l,i]}i[0]=0;for(let l=1;l<=o;++l){const u=e[l]-e[l-1];let c=0;this.nGramWidths.forEach(h=>{c+=this.getNumNGrams(u,h)}),this.preserveShort&&u>0&&c===0&&(c=1),i[l]=i[l-1]+c}const a=new Array(i[o]);for(let l=0;l<o;++l){const u=e[l];let c=i[l];if(this.nGramWidths.forEach(h=>{const f=e[l+1]-e[l],d=this.getNumNGrams(f,h);this.createNGrams(t,u,a,c,d,h),c+=d}),this.preserveShort&&c===i[l]){const h=e[l+1]-e[l];if(h===0)continue;const f=h+2*this.padWidth;this.createNGrams(t,u,a,c,1,f)}}return[a,i]}}function US(n,t,e,s,r,o,i,a){return new zS(e,s,r,o,i,a).compute(n,t)}/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -4392,7 +4392,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const Fs=(n,t)=>{const e=t.value-n.value;return e===0?n.index-t.index:e};function ff(n,t,e=0,s=n.length-1){for(;s>e;){if(s-e>600){const a=s-e+1,l=t-e+1,u=Math.log(a),c=.5*Math.exp(2*u/3),h=.5*Math.sqrt(u*c*(a-c)/a)*Math.sign(l-a/2),f=Math.max(e,Math.floor(t-l*c/a+h)),d=Math.min(s,Math.floor(t+(a-l)*c/a+h));ff(n,t,f,d)}const r=n[t];let o=e,i=s;for(En(n,e,t),Fs(n[s],r)>0&&En(n,e,s);o<i;){for(En(n,o,i),o++,i--;Fs(n[o],r)<0;)o=o+1;for(;Fs(n[i],r)>0;)i=i-1}Fs(n[e],r)===0?En(n,e,i):(i=i+1,En(n,i,s)),i<=t&&(e=i+1),t<=i&&(s=i-1)}}function HS(n,t,e,s,r){const o=t[t.length-1],[i,a]=[n.length/o,o],l=Cn(e,i*s),u=Cn("int32",i*s);for(let h=0;h<i;h++){const f=h*a,d=n.subarray(f,f+a);let p=new Array(d.length);d.forEach((y,S)=>p[S]={value:y,index:S}),s<p.length&&(ff(p,s),p=p.slice(0,s)),r&&p.sort(Fs);const g=h*s,m=l.subarray(g,g+s),b=u.subarray(g,g+s);for(let y=0;y<s;y++)m[y]=p[y].value,b[y]=p[y].index}const c=t.slice();return c[c.length-1]=s,[xt(c,e,l),xt(c,"int32",u)]}/**
+ */const zs=(n,t)=>{const e=t.value-n.value;return e===0?n.index-t.index:e};function ff(n,t,e=0,s=n.length-1){for(;s>e;){if(s-e>600){const a=s-e+1,l=t-e+1,u=Math.log(a),c=.5*Math.exp(2*u/3),h=.5*Math.sqrt(u*c*(a-c)/a)*Math.sign(l-a/2),f=Math.max(e,Math.floor(t-l*c/a+h)),d=Math.min(s,Math.floor(t+(a-l)*c/a+h));ff(n,t,f,d)}const r=n[t];let o=e,i=s;for(En(n,e,t),zs(n[s],r)>0&&En(n,e,s);o<i;){for(En(n,o,i),o++,i--;zs(n[o],r)<0;)o=o+1;for(;zs(n[i],r)>0;)i=i-1}zs(n[e],r)===0?En(n,e,i):(i=i+1,En(n,i,s)),i<=t&&(e=i+1),t<=i&&(s=i-1)}}function HS(n,t,e,s,r){const o=t[t.length-1],[i,a]=[n.length/o,o],l=Cn(e,i*s),u=Cn("int32",i*s);for(let h=0;h<i;h++){const f=h*a,d=n.subarray(f,f+a);let p=new Array(d.length);d.forEach((y,S)=>p[S]={value:y,index:S}),s<p.length&&(ff(p,s),p=p.slice(0,s)),r&&p.sort(zs);const g=h*s,m=l.subarray(g,g+s),b=u.subarray(g,g+s);for(let y=0;y<s;y++)m[y]=p[y].value,b[y]=p[y].index}const c=t.slice();return c[c.length-1]=s,[xt(c,e,l),xt(c,"int32",u)]}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -4407,7 +4407,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function KS(n,t,e,s){const r=os(t,e)[0],o=[1,e[0],1];for(let p=0;p<r;p++)o[0]*=e[p];o[1]=e[r];for(let p=r+1;p<e.length;p++)o[2]*=e[p];const i=new Map,a=new Int32Array(e[r]),l=new Xs(o,s,n),u=[],c=o[0]===1&&o[2]===1;for(let p=0;p<e[r];p++){let g;if(c)g=n[p].toString();else{const b=[];for(let y=0;y<o[0];y++)for(let S=0;S<o[2];S++)b.push(l.get(y,p,S));g=b.join(",")}const m=i.get(g);if(m!=null)a[p]=m;else{const b=i.size;i.set(g,b),a[p]=b,u.push(p)}}const h=o.slice();h[1]=i.size;const f=new Xs(h,s);u.forEach((p,g)=>{for(let m=0;m<o[0];m++)for(let b=0;b<o[2];b++)f.set(l.get(m,p,b),m,g,b)});const d=e.slice();return d[r]=h[1],{outputValues:f.values,outputShape:d,indices:a}}/**
+ */function KS(n,t,e,s){const r=is(t,e)[0],o=[1,e[0],1];for(let p=0;p<r;p++)o[0]*=e[p];o[1]=e[r];for(let p=r+1;p<e.length;p++)o[2]*=e[p];const i=new Map,a=new Int32Array(e[r]),l=new Js(o,s,n),u=[],c=o[0]===1&&o[2]===1;for(let p=0;p<e[r];p++){let g;if(c)g=n[p].toString();else{const b=[];for(let y=0;y<o[0];y++)for(let S=0;S<o[2];S++)b.push(l.get(y,p,S));g=b.join(",")}const m=i.get(g);if(m!=null)a[p]=m;else{const b=i.size;i.set(g,b),a[p]=b,u.push(p)}}const h=o.slice();h[1]=i.size;const f=new Js(h,s);u.forEach((p,g)=>{for(let m=0;m<o[0];m++)for(let b=0;b<o[2];b++)f.set(l.get(m,p,b),m,g,b)});const d=e.slice();return d[r]=h[1],{outputValues:f.values,outputShape:d,indices:a}}/**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -4452,7 +4452,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class e2{constructor(t,e){this.variableNames=["source"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.rank=e.length,this.dispatchLayout=ae(this.outputShape),this.dispatch=It(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.start=t,this.uniforms=`start : ${Tt(t.length)}, `,this.shaderKey="slice"}getUserCode(){const t=Tt(this.rank),e=n2(this.rank);let s;return this.start.length===1?s=this.outputShape.map((o,i)=>"sourceLoc = uniforms.start + coords;"):s=this.outputShape.map((o,i)=>`sourceLoc.${fa[i]} = uniforms.start.${vn(i)} + coords.${fa[i]};`),`
+ */class e2{constructor(t,e){this.variableNames=["source"],this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=e,this.rank=e.length,this.dispatchLayout=ae(this.outputShape),this.dispatch=It(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.start=t,this.uniforms=`start : ${Tt(t.length)}, `,this.shaderKey="slice"}getUserCode(){const t=Tt(this.rank),e=n2(this.rank);let s;return this.start.length===1?s=this.outputShape.map((o,i)=>"sourceLoc = uniforms.start + coords;"):s=this.outputShape.map((o,i)=>`sourceLoc.${fa[i]} = uniforms.start.${In(i)} + coords.${fa[i]};`),`
       ${wt("index")} {
         if (index < uniforms.size) {
           var sourceLoc : ${t};
@@ -4738,7 +4738,7 @@ out_color[outIdx] = vec4f(
       }`:a=`
       fn activation(a : ${i}, coords : vec${s}<i32>) -> ${i} {
         ${r}
-      }`,a}function no(n,t){return`
+      }`,a}function so(n,t){return`
       ${n?"value = value + getBiasByOutputCoords(coords);":""}
       ${t?"value = activation(value, coords);":""}
       `}/**
@@ -4783,7 +4783,7 @@ out_color[outIdx] = vec4f(
     {
       var value = valueIn;
       let coords = vec3<i32>(batch, row, col);
-      ${no(n,t)}
+      ${so(n,t)}
       setOutputAtCoords(coords[0], coords[1], coords[2], value);
     }
   }
@@ -5125,7 +5125,7 @@ out_color[outIdx] = vec4f(
         var value = valueIn;
         let outWidth = ${n?"uniforms.outShape[2]":"uniforms.outShape[3]"};
         ${d}
-        ${no(r,o)}
+        ${so(r,o)}
         setOutputAtCoords(coords[0], coords[1], coords[2], coords[3], value);
         }
       }`}class R${constructor(t,e,s,r,o=!1,i=null,a=!1,l=!1){this.variableNames=["x","W"],this.uniforms="filterDims : vec2<i32>, pads : vec2<i32>, strides : vec2<i32>, dilations : vec2<i32>, dimAOuter : i32, dimBOuter : i32, dimInner : i32,",this.outputShape=t.outShape,this.isChannelsLast=t.dataFormat==="channelsLast",this.isVec4=((t.inChannels%4===0||t.inChannels%3===0)&&this.isChannelsLast||t.outWidth%4===0&&!this.isChannelsLast)&&t.outChannels%4===0,this.dispatchLayout=this.isChannelsLast?{x:[3],y:[1,2],z:[0]}:{x:[2,3],y:[1],z:[0]},this.workgroupSize=Px(this.dispatchLayout,this.outputShape,this.isVec4),this.elementsPerThread=Lx(this.dispatchLayout,this.outputShape,this.isVec4),this.dispatch=It(this.dispatchLayout,this.outputShape,this.workgroupSize,this.elementsPerThread),this.isVec4?(this.outputComponent=4,this.isChannelsLast&&t.inChannels%4!==0?(this.innerElementSize=3,this.variableComponents=[1,4]):(this.innerElementSize=4,this.variableComponents=[4,4]),o&&(this.variableNames.push("bias"),this.variableComponents.push(4)),a&&(this.variableNames.push("preluActivationWeights"),this.variableComponents.push(4))):(this.innerElementSize=this.elementsPerThread[0],o&&this.variableNames.push("bias"),a&&this.variableNames.push("preluActivationWeights")),this.sequentialAccessByThreads=l,this.addBias=o,this.activation=i,this.hasPreluActivationWeights=a,this.tileAOuter=this.workgroupSize[1]*this.elementsPerThread[1],this.tileBOuter=this.workgroupSize[0]*this.elementsPerThread[0],this.tileInner=Math.max(this.workgroupSize[0]*this.innerElementSize,this.workgroupSize[1]),this.fitAOuter=e%this.tileAOuter===0,this.fitBOuter=s%this.tileBOuter===0,this.fitInner=r%this.tileInner===0,this.shaderKey=`conv2DMM_${this.elementsPerThread}_${this.activation}}_${this.fitAOuter}_${this.fitBOuter}_${this.fitInner}_${this.isVec4}_${this.innerElementSize}_${this.isChannelsLast}_${this.sequentialAccessByThreads}`}getUserCode(){const t=this.isVec4?pa(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner):ma(this.elementsPerThread,this.workgroupSize,!this.isChannelsLast,this.tileInner,!1,null,this.sequentialAccessByThreads),e=this.isVec4?[this.innerElementSize,4,4]:[1,1,1];return`
@@ -5168,7 +5168,7 @@ out_color[outIdx] = vec4f(
          let coords = ${this.isChannelsLast?"vec4<i32>(batch, row, col, chan);":"vec4<i32>(batch, chan, row, col);"}
          if (coordsInBounds4D(coords, uniforms.outShape)) {
            var value = valueIn;
-           ${no(this.addBias,this.activation)}
+           ${so(this.addBias,this.activation)}
            setOutputAtCoords(coords.x, coords.y, coords.z, coords.w, value);
          }
        }
@@ -5389,7 +5389,7 @@ out_color[outIdx] = vec4f(
       if (index < uniforms.size) {
         let coords = getCoordsFromIndex(index);
         var value = getXByOutputIndex(index);
-        ${no(this.addBias,this.activation)}
+        ${so(this.addBias,this.activation)}
         setOutputAtIndex(index, value);
       }
     }
@@ -5423,7 +5423,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function mf({a:n,b:t,transposeA:e,transposeB:s,backend:r,bias:o=null,preluActivationWeights:i=null,leakyreluAlpha:a=0,activation:l=null}){const u=n.shape.length,c=t.shape.length,h=e?n.shape[u-2]:n.shape[u-1],f=s?t.shape[c-1]:t.shape[c-2],d=e?n.shape[u-1]:n.shape[u-2],p=s?t.shape[c-2]:t.shape[c-1],g=n.shape.slice(0,-2),m=t.shape.slice(0,-2),b=z(g),y=z(m),x=zt(n.shape.slice(0,-2),t.shape.slice(0,-2)).concat([d,p]);w(h===f,()=>`Error in matMul: inner shapes (${h}) and (${f}) of Tensors with shapes ${n.shape} and ${t.shape} and transposeA=${e} and transposeB=${s} must match.`);const $=e?[b,h,d]:[b,d,h],E=s?[y,p,f]:[y,f,p],D=dt({inputs:{x:n},backend:r,attrs:{shape:$}}),_=dt({inputs:{x:t},backend:r,attrs:{shape:E}}),T=[D,_],P=Math.max(b,y),B=[D,_],Y=[{type:"int32",data:[d]},{type:"int32",data:[p]},{type:"int32",data:[h]}];let j,F;const G=[P,d,p];let q=W().get("WEBGPU_MATMUL_PROGRAM_TYPE");if(q<0){const Xt=W().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),Jt=Xt>0?Xt:r.thresholdToIncreaseWorkgroups,An=P*Math.ceil(d/32)*Math.ceil(p/32);An<=Jt||d<=8&&An<=Jt*2?P*d*p<=128?q=Ne.MatMulReduceProgram:P===1&&f>=2e3?q=Ne.MatMulSplitKProgram:q=Ne.MatMulSmallOutputSizeProgram:q=Ne.MatMulPackedProgram}switch(q){case Ne.MatMulReduceProgram:j=new O$(G,e,s,o,l,i);break;case Ne.MatMulSplitKProgram:{if(F=of({backend:r,attrs:{shape:G,value:0,dtype:n.dtype}}),j=new z$(G,f,e,s),o||l){F=r.runWebGPUProgram(j,B,n.dtype,Y,F);const Jt=new U$(F.shape,o,l,i);let An=null;const ro=[F];o&&ro.push(o),i&&ro.push(i),l==="leakyrelu"&&(An=[{type:"float32",data:[a]}],Jt.uniforms+=" alpha : f32,");const yf=r.runWebGPUProgram(Jt,ro,F.dtype,An);T.push(F);const yv=dt({inputs:{x:yf},backend:r,attrs:{shape:x}});T.push(yf);for(const wv of T)r.disposeData(wv.dataId);return yv}break}case Ne.MatMulSmallOutputSizeProgram:j=new F$($,E,G,e,s,o,l,i);break;case Ne.MatMulPackedProgram:const Xt=r.adapterInfo.isIntel();j=new N$($,G,e,s,o,l,i,Xt);break;default:throw new Error(`Unsupported MatMulProgramType ${q}.`)}o&&B.push(o),i&&B.push(i),l==="leakyrelu"&&(Y.push({type:"float32",data:[a]}),j.uniforms+=" alpha : f32,"),F=r.runWebGPUProgram(j,B,n.dtype,Y,F);const De=dt({inputs:{x:F},backend:r,attrs:{shape:x}});T.push(F);for(const Xt of T)r.disposeData(Xt.dataId);return De}/**
+ */function mf({a:n,b:t,transposeA:e,transposeB:s,backend:r,bias:o=null,preluActivationWeights:i=null,leakyreluAlpha:a=0,activation:l=null}){const u=n.shape.length,c=t.shape.length,h=e?n.shape[u-2]:n.shape[u-1],f=s?t.shape[c-1]:t.shape[c-2],d=e?n.shape[u-1]:n.shape[u-2],p=s?t.shape[c-2]:t.shape[c-1],g=n.shape.slice(0,-2),m=t.shape.slice(0,-2),b=z(g),y=z(m),x=zt(n.shape.slice(0,-2),t.shape.slice(0,-2)).concat([d,p]);w(h===f,()=>`Error in matMul: inner shapes (${h}) and (${f}) of Tensors with shapes ${n.shape} and ${t.shape} and transposeA=${e} and transposeB=${s} must match.`);const $=e?[b,h,d]:[b,d,h],E=s?[y,p,f]:[y,f,p],D=dt({inputs:{x:n},backend:r,attrs:{shape:$}}),_=dt({inputs:{x:t},backend:r,attrs:{shape:E}}),T=[D,_],P=Math.max(b,y),B=[D,_],Y=[{type:"int32",data:[d]},{type:"int32",data:[p]},{type:"int32",data:[h]}];let j,F;const G=[P,d,p];let q=W().get("WEBGPU_MATMUL_PROGRAM_TYPE");if(q<0){const qt=W().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),jt=qt>0?qt:r.thresholdToIncreaseWorkgroups,Xe=P*Math.ceil(d/32)*Math.ceil(p/32);Xe<=jt||d<=8&&Xe<=jt*2?P*d*p<=128?q=Ne.MatMulReduceProgram:P===1&&f>=2e3?q=Ne.MatMulSplitKProgram:q=Ne.MatMulSmallOutputSizeProgram:q=Ne.MatMulPackedProgram}switch(q){case Ne.MatMulReduceProgram:j=new O$(G,e,s,o,l,i);break;case Ne.MatMulSplitKProgram:{if(F=of({backend:r,attrs:{shape:G,value:0,dtype:n.dtype}}),j=new z$(G,f,e,s),o||l){F=r.runWebGPUProgram(j,B,n.dtype,Y,F);const jt=new U$(F.shape,o,l,i);let Xe=null;const os=[F];o&&os.push(o),i&&os.push(i),l==="leakyrelu"&&(Xe=[{type:"float32",data:[a]}],jt.uniforms+=" alpha : f32,");const yf=r.runWebGPUProgram(jt,os,F.dtype,Xe);T.push(F);const yv=dt({inputs:{x:yf},backend:r,attrs:{shape:x}});T.push(yf);for(const wv of T)r.disposeData(wv.dataId);return yv}break}case Ne.MatMulSmallOutputSizeProgram:j=new F$($,E,G,e,s,o,l,i);break;case Ne.MatMulPackedProgram:const qt=r.adapterInfo.isIntel();j=new N$($,G,e,s,o,l,i,qt);break;default:throw new Error(`Unsupported MatMulProgramType ${q}.`)}o&&B.push(o),i&&B.push(i),l==="leakyrelu"&&(Y.push({type:"float32",data:[a]}),j.uniforms+=" alpha : f32,"),F=r.runWebGPUProgram(j,B,n.dtype,Y,F);const De=dt({inputs:{x:F},backend:r,attrs:{shape:x}});T.push(F);for(const qt of T)r.disposeData(qt.dataId);return De}/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -5438,7 +5438,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function so(n,t){const e=n.length;return e>=3?t?[...n.slice(0,-3),n[e-3]*n[e-2],n[e-1]]:[...n.slice(0,-3),n[e-3],n[e-2]*n[e-1]]:!t&&e===1&&n[0]>1?[n[0],1]:null}function W$({x:n,filter:t,convInfo:e,backend:s,bias:r=null,preluActivationWeights:o=null,leakyreluAlpha:i=0,activation:a=null}){const l=e.dataFormat==="channelsLast",u=!l,c=!1,h=l&&e.filterHeight===e.inHeight&&e.filterWidth===e.inWidth&&e.padInfo.type==="VALID",f=[];let d,p;if(h){const b=e.inHeight*e.inWidth*e.inChannels;d=dt({inputs:{x:n},backend:s,attrs:{shape:[1,e.batchSize,b]}}),p=dt({inputs:{x:t},backend:s,attrs:{shape:[1,b,e.outChannels]}})}else d=dt({inputs:{x:n},backend:s,attrs:{shape:l?[e.batchSize,e.inHeight*e.inWidth,e.inChannels]:[e.batchSize,e.inChannels,e.inHeight*e.inWidth]}}),p=dt({inputs:{x:t},backend:s,attrs:{shape:[1,e.inChannels,e.outChannels]}});if(f.push(d),f.push(p),o!=null){const b=so(o.shape,l);b!=null&&(o=dt({inputs:{x:o},backend:s,attrs:{shape:b}}),f.push(o))}if(r!=null){const b=so(r.shape,l);b!=null&&(r=dt({inputs:{x:r},backend:s,attrs:{shape:b}}),f.push(r))}const g=mf({a:l?d:p,b:l?p:d,transposeA:u,transposeB:c,backend:s,bias:r,activation:a,preluActivationWeights:o,leakyreluAlpha:i}),m=dt({inputs:{x:g},backend:s,attrs:{shape:e.outShape}});f.push(g);for(const b of f)s.disposeData(b.dataId);return m}function G$({x:n,filter:t,convInfo:e,backend:s,bias:r=null,preluActivationWeights:o=null,leakyreluAlpha:i=0,activation:a=null}){const{filterWidth:l,filterHeight:u,inChannels:c,strideWidth:h,strideHeight:f,padInfo:d,outWidth:p,outHeight:g,dilationWidth:m,dilationHeight:b,dataFormat:y}=e,S=y==="channelsLast",x=l*u*c,$=g*p,E=S?[e.batchSize,$,x]:[e.batchSize,x,$],D=new L$(E,S),_=[{type:"int32",data:[d.top,d.left]},{type:"int32",data:[f,h]},{type:"int32",data:[b,m]},{type:"int32",data:[p]},{type:"int32",data:[c*l]},{type:"int32",data:[c]}],T=s.runWebGPUProgram(D,[n],n.dtype,_),P=[];P.push(T);const B=dt({inputs:{x:t},backend:s,attrs:{shape:[1,x,-1]}});if(P.push(B),o!=null){const q=so(o.shape,S);q!=null&&(o=dt({inputs:{x:o},backend:s,attrs:{shape:q}}),P.push(o))}if(r!=null){const q=so(r.shape,S);q!=null&&(r=dt({inputs:{x:r},backend:s,attrs:{shape:q}}),P.push(r))}const F=mf({a:S?T:B,b:S?B:T,transposeA:!S,transposeB:!1,backend:s,bias:r,activation:a,preluActivationWeights:o,leakyreluAlpha:i}),G=dt({inputs:{x:F},backend:s,attrs:{shape:e.outShape}});P.push(F);for(const q of P)s.disposeData(q.dataId);return G}function V$({x:n,filter:t,convInfo:e,backend:s,bias:r=null,preluActivationWeights:o=null,leakyreluAlpha:i=0,activation:a=null}){const l=r!=null,u=o!=null,c=e.dataFormat==="channelsLast",h=c&&e.filterHeight===e.inHeight&&e.filterWidth===e.inWidth&&e.padInfo.type==="VALID",f=W().getBool("WEBGPU_USE_NAIVE_CONV2D_DEBUG");if(!f&&(h||e.filterHeight===1&&e.filterWidth===1&&e.dilationHeight===1&&e.dilationWidth===1&&e.strideHeight===1&&e.strideWidth===1&&(e.padInfo.type==="SAME"||e.padInfo.type==="VALID")))return W$({x:n,filter:t,convInfo:e,backend:s,bias:r,activation:a,preluActivationWeights:o,leakyreluAlpha:i});const d=W().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),p=d>-1?d:s.thresholdToIncreaseWorkgroups,g=e.batchSize*Math.ceil(e.outHeight*e.outWidth/32)*Math.ceil(e.outChannels/32);if(W().getBool("WEBGPU_CONV_SEPARATE_IM2COL_SHADER")||g<=p)return G$({x:n,filter:t,convInfo:e,backend:s,bias:r,preluActivationWeights:o,leakyreluAlpha:i,activation:a});let m;const b=[e.padInfo.top,e.padInfo.left],y=[{type:"int32",data:[e.filterHeight,e.filterWidth]},{type:"int32",data:[...b]},{type:"int32",data:[e.strideHeight,e.strideWidth]},{type:"int32",data:[e.dilationHeight,e.dilationWidth]}];if(f)m=new P$(e,l,a,u);else{const E=c?e.outHeight*e.outWidth:e.outChannels,D=c?e.outChannels:e.outHeight*e.outWidth,_=e.filterHeight*e.filterWidth*e.inChannels;y.push({type:"int32",data:[E]},{type:"int32",data:[D]},{type:"int32",data:[_]});const T=s.adapterInfo.isIntel();m=new R$(e,E,D,_,l,a,u,T)}const S=[],x=[n,t];l&&(!c&&r.shape.length===1&&(r=dt({inputs:{x:r},backend:s,attrs:{shape:[r.shape[0],1,1]}}),S.push(r)),x.push(r)),u&&(!c&&o.shape.length===1&&(o=dt({inputs:{x:o},backend:s,attrs:{shape:[o.shape[0],1,1]}}),S.push(o)),x.push(o)),a==="leakyrelu"&&(y.push({type:"float32",data:[i]}),m.uniforms+=" alpha : f32,");const $=s.runWebGPUProgram(m,x,n.dtype,y);for(const E of S)s.disposeData(E.dataId);return $}/**
+ */function ro(n,t){const e=n.length;return e>=3?t?[...n.slice(0,-3),n[e-3]*n[e-2],n[e-1]]:[...n.slice(0,-3),n[e-3],n[e-2]*n[e-1]]:!t&&e===1&&n[0]>1?[n[0],1]:null}function W$({x:n,filter:t,convInfo:e,backend:s,bias:r=null,preluActivationWeights:o=null,leakyreluAlpha:i=0,activation:a=null}){const l=e.dataFormat==="channelsLast",u=!l,c=!1,h=l&&e.filterHeight===e.inHeight&&e.filterWidth===e.inWidth&&e.padInfo.type==="VALID",f=[];let d,p;if(h){const b=e.inHeight*e.inWidth*e.inChannels;d=dt({inputs:{x:n},backend:s,attrs:{shape:[1,e.batchSize,b]}}),p=dt({inputs:{x:t},backend:s,attrs:{shape:[1,b,e.outChannels]}})}else d=dt({inputs:{x:n},backend:s,attrs:{shape:l?[e.batchSize,e.inHeight*e.inWidth,e.inChannels]:[e.batchSize,e.inChannels,e.inHeight*e.inWidth]}}),p=dt({inputs:{x:t},backend:s,attrs:{shape:[1,e.inChannels,e.outChannels]}});if(f.push(d),f.push(p),o!=null){const b=ro(o.shape,l);b!=null&&(o=dt({inputs:{x:o},backend:s,attrs:{shape:b}}),f.push(o))}if(r!=null){const b=ro(r.shape,l);b!=null&&(r=dt({inputs:{x:r},backend:s,attrs:{shape:b}}),f.push(r))}const g=mf({a:l?d:p,b:l?p:d,transposeA:u,transposeB:c,backend:s,bias:r,activation:a,preluActivationWeights:o,leakyreluAlpha:i}),m=dt({inputs:{x:g},backend:s,attrs:{shape:e.outShape}});f.push(g);for(const b of f)s.disposeData(b.dataId);return m}function G$({x:n,filter:t,convInfo:e,backend:s,bias:r=null,preluActivationWeights:o=null,leakyreluAlpha:i=0,activation:a=null}){const{filterWidth:l,filterHeight:u,inChannels:c,strideWidth:h,strideHeight:f,padInfo:d,outWidth:p,outHeight:g,dilationWidth:m,dilationHeight:b,dataFormat:y}=e,S=y==="channelsLast",x=l*u*c,$=g*p,E=S?[e.batchSize,$,x]:[e.batchSize,x,$],D=new L$(E,S),_=[{type:"int32",data:[d.top,d.left]},{type:"int32",data:[f,h]},{type:"int32",data:[b,m]},{type:"int32",data:[p]},{type:"int32",data:[c*l]},{type:"int32",data:[c]}],T=s.runWebGPUProgram(D,[n],n.dtype,_),P=[];P.push(T);const B=dt({inputs:{x:t},backend:s,attrs:{shape:[1,x,-1]}});if(P.push(B),o!=null){const q=ro(o.shape,S);q!=null&&(o=dt({inputs:{x:o},backend:s,attrs:{shape:q}}),P.push(o))}if(r!=null){const q=ro(r.shape,S);q!=null&&(r=dt({inputs:{x:r},backend:s,attrs:{shape:q}}),P.push(r))}const F=mf({a:S?T:B,b:S?B:T,transposeA:!S,transposeB:!1,backend:s,bias:r,activation:a,preluActivationWeights:o,leakyreluAlpha:i}),G=dt({inputs:{x:F},backend:s,attrs:{shape:e.outShape}});P.push(F);for(const q of P)s.disposeData(q.dataId);return G}function V$({x:n,filter:t,convInfo:e,backend:s,bias:r=null,preluActivationWeights:o=null,leakyreluAlpha:i=0,activation:a=null}){const l=r!=null,u=o!=null,c=e.dataFormat==="channelsLast",h=c&&e.filterHeight===e.inHeight&&e.filterWidth===e.inWidth&&e.padInfo.type==="VALID",f=W().getBool("WEBGPU_USE_NAIVE_CONV2D_DEBUG");if(!f&&(h||e.filterHeight===1&&e.filterWidth===1&&e.dilationHeight===1&&e.dilationWidth===1&&e.strideHeight===1&&e.strideWidth===1&&(e.padInfo.type==="SAME"||e.padInfo.type==="VALID")))return W$({x:n,filter:t,convInfo:e,backend:s,bias:r,activation:a,preluActivationWeights:o,leakyreluAlpha:i});const d=W().getNumber("WEBGPU_THRESHOLD_TO_INCREASE_WORKGROUPS_FOR_MATMUL"),p=d>-1?d:s.thresholdToIncreaseWorkgroups,g=e.batchSize*Math.ceil(e.outHeight*e.outWidth/32)*Math.ceil(e.outChannels/32);if(W().getBool("WEBGPU_CONV_SEPARATE_IM2COL_SHADER")||g<=p)return G$({x:n,filter:t,convInfo:e,backend:s,bias:r,preluActivationWeights:o,leakyreluAlpha:i,activation:a});let m;const b=[e.padInfo.top,e.padInfo.left],y=[{type:"int32",data:[e.filterHeight,e.filterWidth]},{type:"int32",data:[...b]},{type:"int32",data:[e.strideHeight,e.strideWidth]},{type:"int32",data:[e.dilationHeight,e.dilationWidth]}];if(f)m=new P$(e,l,a,u);else{const E=c?e.outHeight*e.outWidth:e.outChannels,D=c?e.outChannels:e.outHeight*e.outWidth,_=e.filterHeight*e.filterWidth*e.inChannels;y.push({type:"int32",data:[E]},{type:"int32",data:[D]},{type:"int32",data:[_]});const T=s.adapterInfo.isIntel();m=new R$(e,E,D,_,l,a,u,T)}const S=[],x=[n,t];l&&(!c&&r.shape.length===1&&(r=dt({inputs:{x:r},backend:s,attrs:{shape:[r.shape[0],1,1]}}),S.push(r)),x.push(r)),u&&(!c&&o.shape.length===1&&(o=dt({inputs:{x:o},backend:s,attrs:{shape:[o.shape[0],1,1]}}),S.push(o)),x.push(o)),a==="leakyrelu"&&(y.push({type:"float32",data:[i]}),m.uniforms+=" alpha : f32,");const $=s.runWebGPUProgram(m,x,n.dtype,y);for(const E of S)s.disposeData(E.dataId);return $}/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -5599,7 +5599,7 @@ out_color[outIdx] = vec4f(
           }
         }
       }
-    `}}function J$(n){const t=n.length;if(t>6)throw Error(`Transpose for rank ${t} is not yet supported`);const e=new Array(t);for(let s=0;s<n.length;s++)e[n[s]]=`coords.${vn(s)}`;return e.join()}/**
+    `}}function J$(n){const t=n.length;if(t>6)throw Error(`Transpose for rank ${t} is not yet supported`);const e=new Array(t);for(let s=0;s<n.length;s++)e[n[s]]=`coords.${In(s)}`;return e.join()}/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -5692,7 +5692,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */const tv={mean:"float32",all:"bool",any:"bool"};function ev(n,t,e,s,r){const o=n.shape.length,i=[],a=os(t,n.shape);let l=a;const u=mg(l,o);let c=n;u!=null&&(c=Z$({inputs:{x:n},attrs:{perm:u},backend:r}),l=gg(l.length,o),i.push(c)),pg(s,l,o);const[h,f]=Wo(c.shape,l);let d=h;e&&(d=Cl(h,a));let p;if(r.shouldExecuteOnCPU([c])){const g=r.tensorMap.get(c.dataId).values;switch(s){case"max":const m=JS(g,z(f),d,n.dtype);p=r.makeTensorInfo(d,n.dtype,m);break;case"prod":const{outVals:b,outShape:y,outDtype:S}=ZS(c.shape,c.dtype,g,l);p=r.makeTensorInfo(y,S,b);break;default:throw new Error(`${s} CPU implementation is not yet supported.`)}}else{const g=z(f),b=z(c.shape)/g,y={windowSize:g,inSize:g,batchSize:b,outSize:1},S=tv[s]||Tp(n.dtype),x=[{type:"int32",data:[g]}],$=new Q$(y,s,r.device.limits.maxComputeWorkgroupSizeX),E=r.runWebGPUProgram($,[c],S,x);i.push(E),p=dt({inputs:{x:E},attrs:{shape:d},backend:r})}return i.forEach(g=>r.disposeData(g.dataId)),p}/**
+ */const tv={mean:"float32",all:"bool",any:"bool"};function ev(n,t,e,s,r){const o=n.shape.length,i=[],a=is(t,n.shape);let l=a;const u=mg(l,o);let c=n;u!=null&&(c=Z$({inputs:{x:n},attrs:{perm:u},backend:r}),l=gg(l.length,o),i.push(c)),pg(s,l,o);const[h,f]=Wo(c.shape,l);let d=h;e&&(d=Cl(h,a));let p;if(r.shouldExecuteOnCPU([c])){const g=r.tensorMap.get(c.dataId).values;switch(s){case"max":const m=JS(g,z(f),d,n.dtype);p=r.makeTensorInfo(d,n.dtype,m);break;case"prod":const{outVals:b,outShape:y,outDtype:S}=ZS(c.shape,c.dtype,g,l);p=r.makeTensorInfo(y,S,b);break;default:throw new Error(`${s} CPU implementation is not yet supported.`)}}else{const g=z(f),b=z(c.shape)/g,y={windowSize:g,inSize:g,batchSize:b,outSize:1},S=tv[s]||Tp(n.dtype),x=[{type:"int32",data:[g]}],$=new Q$(y,s,r.device.limits.maxComputeWorkgroupSizeX),E=r.runWebGPUProgram($,[c],S,x);i.push(E),p=dt({inputs:{x:E},attrs:{shape:d},backend:r})}return i.forEach(g=>r.disposeData(g.dataId)),p}/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -5813,7 +5813,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */class uv{constructor(t){this.uniforms="",this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=ms(t,1),this.variableNames=t.map((e,s)=>`T${s}`),this.dispatchLayout=ae(this.outputShape),this.dispatch=It(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.offsetLength=t.length-1;for(let e=0;e<this.offsetLength;e++)this.uniforms+=`offset${e} : i32,`;this.shaderKey="concat"}getUserCode(){const t=[];if(this.offsetLength>0){t.push("if (yC < uniforms.offset0){ setOutputAtCoords(coords.x, coords.y, getT0(yR, yC)); }");for(let o=1;o<this.offsetLength;o++)t.push(`else if (yC < uniforms.offset${[o]}){ setOutputAtCoords(coords.x, coords.y, getT${o}(yR, yC - uniforms.offset${o-1})); }`);const s=this.offsetLength,r=this.offsetLength-1;t.push(`else { setOutputAtCoords(coords.x, coords.y, getT${s}(yR, yC - uniforms.offset${r})); }`)}else t.push("setOutputAtCoords(coords.x, coords.y, getT0(yR, yC));");return`
+ */class uv{constructor(t){this.uniforms="",this.workPerThread=1,this.workgroupSize=[64,1,1],this.size=!0,this.outputShape=gs(t,1),this.variableNames=t.map((e,s)=>`T${s}`),this.dispatchLayout=ae(this.outputShape),this.dispatch=It(this.dispatchLayout,this.outputShape,this.workgroupSize,[this.workPerThread,1,1]),this.offsetLength=t.length-1;for(let e=0;e<this.offsetLength;e++)this.uniforms+=`offset${e} : i32,`;this.shaderKey="concat"}getUserCode(){const t=[];if(this.offsetLength>0){t.push("if (yC < uniforms.offset0){ setOutputAtCoords(coords.x, coords.y, getT0(yR, yC)); }");for(let o=1;o<this.offsetLength;o++)t.push(`else if (yC < uniforms.offset${[o]}){ setOutputAtCoords(coords.x, coords.y, getT${o}(yR, yC - uniforms.offset${o-1})); }`);const s=this.offsetLength,r=this.offsetLength-1;t.push(`else { setOutputAtCoords(coords.x, coords.y, getT${s}(yR, yC - uniforms.offset${r})); }`)}else t.push("setOutputAtCoords(coords.x, coords.y, getT0(yR, yC));");return`
       ${wt("index")} {
         for(var i = 0; i < ${this.workPerThread}; i = i + 1) {
           let flatIndex = index * ${this.workPerThread} + i;
@@ -5887,7 +5887,7 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function zs(n,t,e){const s=n[0].dtype;if(s==="complex64"){const p=n.map(S=>fv({inputs:{input:S},backend:e})),g=n.map(S=>hv({inputs:{input:S},backend:e})),m=zs(p,t,e),b=zs(g,t,e),y=cv({inputs:{real:m,imag:b},backend:e});return p.forEach(S=>e.disposeData(S.dataId)),g.forEach(S=>e.disposeData(S.dataId)),e.disposeData(m.dataId),e.disposeData(b.dataId),y}let r=e.shouldExecuteOnCPU(n);if(s==="string"&&(r=!0),r){const p=n.map($=>{const D=[-1,z($.shape.slice(t))];return dt({inputs:{x:$},backend:e,attrs:{shape:D}})}),g=p.map($=>({vals:e.readSync($.dataId),shape:$.shape})),m=ms(p.map($=>$.shape),1),b=p[0].shape[0]===1,y=XS(g,m,s,b),S=ms(n.map($=>$.shape),t),x=e.makeTensorInfo(S,s,y);return p.forEach($=>e.disposeData($.dataId)),x}const o=e.device.limits.maxStorageBuffersPerShaderStage-1;if(n.length>o){const p=[];for(let m=0;m<n.length;m+=o){const b=n.slice(m,m+o);p.push(zs(b,t,e))}const g=zs(p,t,e);for(const m of p)e.disposeData(m.dataId);return g}const{tensors2D:i,outShape:a}=dv(n,t,e),l=i.map(p=>p.shape),u=new uv(l),c=[],h=new Array(l.length-1);if(h.length>0){h[0]=l[0][1],c.push({type:"int32",data:[h[0]]});for(let p=1;p<h.length;p++)h[p]=h[p-1]+l[p][1],c.push({type:"int32",data:[h[p]]})}const f=e.runWebGPUProgram(u,i,i[0].dtype,c);i.forEach(p=>e.disposeData(p.dataId));const d=dt({inputs:{x:f},backend:e,attrs:{shape:a}});return e.disposeData(f.dataId),d}function dv(n,t,e){const s=ms(n.map(o=>o.shape),t);return{tensors2D:n.map(o=>dt({inputs:{x:o},backend:e,attrs:{shape:[z(o.shape.slice(0,t)),z(o.shape.slice(t))]}})),outShape:s}}/**
+ */function Us(n,t,e){const s=n[0].dtype;if(s==="complex64"){const p=n.map(S=>fv({inputs:{input:S},backend:e})),g=n.map(S=>hv({inputs:{input:S},backend:e})),m=Us(p,t,e),b=Us(g,t,e),y=cv({inputs:{real:m,imag:b},backend:e});return p.forEach(S=>e.disposeData(S.dataId)),g.forEach(S=>e.disposeData(S.dataId)),e.disposeData(m.dataId),e.disposeData(b.dataId),y}let r=e.shouldExecuteOnCPU(n);if(s==="string"&&(r=!0),r){const p=n.map($=>{const D=[-1,z($.shape.slice(t))];return dt({inputs:{x:$},backend:e,attrs:{shape:D}})}),g=p.map($=>({vals:e.readSync($.dataId),shape:$.shape})),m=gs(p.map($=>$.shape),1),b=p[0].shape[0]===1,y=XS(g,m,s,b),S=gs(n.map($=>$.shape),t),x=e.makeTensorInfo(S,s,y);return p.forEach($=>e.disposeData($.dataId)),x}const o=e.device.limits.maxStorageBuffersPerShaderStage-1;if(n.length>o){const p=[];for(let m=0;m<n.length;m+=o){const b=n.slice(m,m+o);p.push(Us(b,t,e))}const g=Us(p,t,e);for(const m of p)e.disposeData(m.dataId);return g}const{tensors2D:i,outShape:a}=dv(n,t,e),l=i.map(p=>p.shape),u=new uv(l),c=[],h=new Array(l.length-1);if(h.length>0){h[0]=l[0][1],c.push({type:"int32",data:[h[0]]});for(let p=1;p<h.length;p++)h[p]=h[p-1]+l[p][1],c.push({type:"int32",data:[h[p]]})}const f=e.runWebGPUProgram(u,i,i[0].dtype,c);i.forEach(p=>e.disposeData(p.dataId));const d=dt({inputs:{x:f},backend:e,attrs:{shape:a}});return e.disposeData(f.dataId),d}function dv(n,t,e){const s=gs(n.map(o=>o.shape),t);return{tensors2D:n.map(o=>dt({inputs:{x:o},backend:e,attrs:{shape:[z(o.shape.slice(0,t)),z(o.shape.slice(t))]}})),outShape:s}}/**
  * @license
  * Copyright 2021 Google LLC. All Rights Reserved.
  * Licensed under the Apache License, Version 2.0 (the "License");
@@ -5902,4 +5902,4 @@ out_color[outIdx] = vec4f(
  * See the License for the specific language governing permissions and
  * limitations under the License.
  * =============================================================================
- */function pv(n){const{inputs:t,backend:e,attrs:s}=n,{axis:r}=s,o=os(r,t[0].shape)[0],i=t.map(u=>u.shape);uy(i,o);const a=ms(t.map(u=>u.shape),o);if(z(a)===0)return e.makeTensorInfo(a,t[0].dtype,[]);const l=t.filter(u=>z(u.shape)>0);return l.length===1?Ye({inputs:{x:l[0]},backend:e}):zs(l,o,e)}const mv=[Fx,Vx,r2,j$,ov,lv,{kernelName:Ca,backendName:"webgpu",kernelFunc:pv},zx];for(const n of mv)lp({...n,backendName:"webgpu-oidn"});async function gv(){try{const n={powerPreference:"high-performance"},t=await navigator.gpu.requestAdapter(n),e={},s=[];t.features.has("timestamp-query")&&s.push("timestamp-query"),t.features.has("bgra8unorm-storage")&&s.push(["bgra8unorm-storage"]),e.requiredFeatures=s;const r=t.limits;e.requiredLimits={maxComputeWorkgroupStorageSize:r.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.maxStorageBufferBindingSize,maxBufferSize:r.maxBufferSize,maxComputeWorkgroupSizeX:r.maxComputeWorkgroupSizeX,maxComputeInvocationsPerWorkgroup:r.maxComputeInvocationsPerWorkgroup};const o=await t.requestDevice(e),i=await t.requestAdapterInfo();return gf(o,i)}catch{}}async function gf(n,t){const e=new Bs(n,t);return A.registerBackend("webgpu-oidn",()=>e),await A.setBackend("webgpu-oidn"),e}async function bf(n,t,e){const s=await(t?gf(t.device,t.adapterInfo):gv()),r=ga(n);return new Yh(r,s,e)}async function bv(n,t,e){return fetch(n).then(s=>s.arrayBuffer()).then(s=>bf(s,t,e))}Nt.UNet=Yh,Nt.initUNetFromBuffer=bf,Nt.initUNetFromURL=bv,Nt.parseTZA=ga,Object.defineProperty(Nt,Symbol.toStringTag,{value:"Module"})});
+ */function pv(n){const{inputs:t,backend:e,attrs:s}=n,{axis:r}=s,o=is(r,t[0].shape)[0],i=t.map(u=>u.shape);uy(i,o);const a=gs(t.map(u=>u.shape),o);if(z(a)===0)return e.makeTensorInfo(a,t[0].dtype,[]);const l=t.filter(u=>z(u.shape)>0);return l.length===1?Ye({inputs:{x:l[0]},backend:e}):Us(l,o,e)}const mv=[Fx,Vx,r2,j$,ov,lv,{kernelName:Ca,backendName:"webgpu",kernelFunc:pv},zx];for(const n of mv)lp({...n,backendName:"webgpu-oidn"});async function gv(){try{const n={powerPreference:"high-performance"},t=await navigator.gpu.requestAdapter(n),e={},s=[];t.features.has("timestamp-query")&&s.push("timestamp-query"),t.features.has("bgra8unorm-storage")&&s.push(["bgra8unorm-storage"]),e.requiredFeatures=s;const r=t.limits;e.requiredLimits={maxComputeWorkgroupStorageSize:r.maxComputeWorkgroupStorageSize,maxComputeWorkgroupsPerDimension:r.maxComputeWorkgroupsPerDimension,maxStorageBufferBindingSize:r.maxStorageBufferBindingSize,maxBufferSize:r.maxBufferSize,maxComputeWorkgroupSizeX:r.maxComputeWorkgroupSizeX,maxComputeInvocationsPerWorkgroup:r.maxComputeInvocationsPerWorkgroup};const o=await t.requestDevice(e),i=await t.requestAdapterInfo();return gf(o,i)}catch{}}async function gf(n,t){const e=new Fs(n,t);return A.registerBackend("webgpu-oidn",()=>e),await A.setBackend("webgpu-oidn"),e}async function bf(n,t,e){const s=await(t?gf(t.device,t.adapterInfo):gv()),r=ga(n);return new Yh(r,s,e)}async function bv(n,t,e){return fetch(n).then(s=>s.arrayBuffer()).then(s=>bf(s,t,e))}Nt.UNet=Yh,Nt.initUNetFromBuffer=bf,Nt.initUNetFromURL=bv,Nt.parseTZA=ga,Object.defineProperty(Nt,Symbol.toStringTag,{value:"Module"})});