@genai-fi/nanogpt 0.4.3 → 0.4.5

package/dist/{TiedEmbedding-CnJ1bx4q.js → TiedEmbedding-DznFwzcB.js}

@@ -1,26 +1,26 @@
-import { o as F, h as D, E as M, bb as fo, bc as mo, bd as mi, j as k, b8 as Mn, be as gi, x as L, bf as bi, bg as yi, bh as wi, bi as ki, bj as xi, bk as Ni, bl as vi, bm as go, bn as Si, bo, bp as Ai, bq as yo, br as Ci, p as Hn, al as wt, bs as wo, bt as Ii, bu as Di, bv as zi, c as On, s as V, b as w, bw as ko, bx as Ti, by as $i, bz as xo, bA as Ei, bB as Li, bC as Fi, bD as Mi, bE as Oi, bF as Ri, bG as _i, bH as Bi, k as No, aa as vo, bI as Wi, bJ as So, a5 as $, bK as Ls, bL as Ao, bM as Co, bN as Io, bO as Do, bP as zo, A as To, bQ as $o, bR as Eo, bS as Lo, bT as S, t as x, f as tt, n as Gi, bU as Be, bV as We, ab as Ft, a as Z, af as Fo, bW as Mo, bX as Oo, Z as ct, a0 as ee, ad as P, bY as Ro, bZ as _o, aN as lt, b_ as Bo, z as Q, b$ as Wo, c0 as Go, c1 as Po, c2 as Uo, c3 as Vo, c4 as jo, c5 as Ko, c6 as Ho, B as qo, c7 as Zo, c8 as Jo, c9 as Xo, aq as Yo, ca as Qo, C as tl, Y as he, cb as el, Q as nl, cc as sl, cd as il, ce as rl, at as al, cf as ol, a3 as ll, au as ul, cg as cl, a9 as hl, ch as pl, G as dl, aw as fl, ci as ml, cj as gl, ck as bl, cl as yl, ay as wl, a4 as kl, cm as xl, cn as Nl, co as vl, M as Sl, cp as Al, cq as Cl, cr as Il, X as Dl, _ as zl, aD as Tl, cs as $l, a6 as El, ct as Ll, aB as Fl, P as Ml, cu as Ol, O as qn, aE as Rl, cv as _l, cw as Bl, N as Wl, aH as Gl, aG as Pl, q as Ul, aV as Vl, cx as jl, aW as Kl, cy as Hl, aI as ql, ar as Zl, an as Jl, cz as Xl, T as Yl, ao as Ql, S as tu, u as eu, cA as nu, cB as su, cC as iu, aK as ru, cD as au, y as ou, cE as lu, a1 as uu, aM as cu, aL as hu, cF as Ie, cG as pu, g as du, cH as Fs, F as Bt, $ as Fe, D as fu, w as mu, ac as xe, cI as gu, cJ as bu, m as Ms, cK as yu, cL as Os, cM as wu } from "./index-C4JCoBvj.js";
-import { s as ku, a as xu, g as Nu, b as vu, V as d, N as G, r as bn, e as Su, l as Au, c as Zn, f as et, h as ye, i as Ge, j as Jn, k as Pi, m as Ui, t as Rt, R as Et, n as ht, A as Pt, o as K, p as le, q as Xn, u as pt, w as Ht, v as Pe, x as Ue, y as Yn, z as j, B as Ee, C as Gt, D as Cu, E as be, F as en, G as Qn, H as ue, I as Ct, J as nt, K as Ve, L as Iu, M as Du, O as je, P as Ut, Q as Lt, S as ts, T as oe, U as jt, W as Xe, X as ce, Y as zu, Z as Rn, _ as nn, $ as yn, a0 as Vi, a1 as It, a2 as Rs, a3 as Tu, a4 as ji, a5 as $u, a6 as Eu, a7 as Lu, a8 as Fu, a9 as Mu, aa as qt, ab as Ou, ac as es, ad as zt, ae as Ye, af as Ru, ag as _t, ah as ot, ai as Ki, aj as gt, ak as ne, al as _s, am as Ne, d as St, an as Bs, ao as Ke, ap as Hi, aq as ns, ar as _u, as as Bu, at as qi, au as Wu } from "./tfjs_backend-Cug-PH75.js";
-import { M as Gu, a as wn, f as Zi } from "./dropout-DfDdklfL.js";
-import { z as mt } from "./zeros-dnQxFgAD.js";
-import { o as pe } from "./ones-Bf3YR48P.js";
-import { v as Ji } from "./variable-LJT9Ld63.js";
-import { r as A } from "./reshape-Boe4DuIO.js";
-import { s as B } from "./sum-R28pucR5.js";
-import { m as Ot } from "./mat_mul-415y5Qn2.js";
-import { s as Kt } from "./split-DbcNm1-i.js";
-import { s as Pu, c as Xi } from "./sin-KmhiDuMa.js";
-import { g as Yi, d as ss, e as Ws, c as Uu } from "./axis_util-BgTGy5w8.js";
-import { a as Zt, e as Jt, l as Vu } from "./log_sum_exp-BswFnwOb.js";
-import { s as kn } from "./stack-D1YjmgKN.js";
-import { p as ju } from "./slice_util-19zDNNSn.js";
-import { c as is } from "./concat-CuRsVY-K.js";
-import { g as Qi } from "./gather-ZYRWhmXR.js";
-import { m as at, a as rs } from "./moments-CjeIaVdp.js";
-import { s as tr } from "./softmax-Cujsg4ay.js";
-import { m as ve } from "./max-CP_9O2Yd.js";
-import { t as Ku } from "./tensor-BVeHdl7V.js";
-import { r as Hu } from "./range-9AzeApCc.js";
-import { m as qu } from "./norm-CZM380I3.js";
+import { o as F, h as D, E as M, bb as fo, bc as mo, bd as mi, j as k, b9 as Mn, be as gi, x as L, bf as bi, bg as yi, bh as wi, bi as ki, bj as xi, bk as vi, bl as Ni, bm as go, bn as Si, bo, bp as Ai, bq as yo, br as Ci, p as Hn, O as wt, bs as wo, bt as Ii, bu as Di, bv as zi, c as On, s as V, b as w, bw as ko, bx as Ti, by as $i, bz as xo, bA as Ei, bB as Li, bC as Fi, bD as Mi, bE as Oi, bF as Ri, bG as _i, bH as Bi, k as vo, ag as No, bI as Wi, bJ as So, a8 as $, bK as Ls, bL as Ao, bM as Co, bN as Io, bO as Do, bP as zo, A as To, bQ as $o, bR as Eo, bS as Lo, bT as S, t as x, f as tt, n as Gi, bU as Be, bV as We, ad as Ft, a as Z, L as Fo, bW as Mo, bX as Oo, a0 as ct, a3 as ee, ai as P, bY as Ro, bZ as _o, aP as lt, b_ as Bo, z as Q, b$ as Wo, c0 as Go, c1 as Po, c2 as Uo, c3 as Vo, c4 as jo, c5 as Ko, c6 as Ho, B as qo, c7 as Zo, c8 as Jo, c9 as Xo, as as Yo, ca as Qo, C as tl, $ as he, cb as el, V as nl, cc as sl, cd as il, ce as rl, av as al, cf as ol, a6 as ll, aw as ul, cg as cl, af as hl, ch as pl, G as dl, ay as fl, ci as ml, cj as gl, ck as bl, cl as yl, aA as wl, a7 as kl, cm as xl, cn as vl, co as Nl, M as Sl, cp as Al, cq as Cl, cr as Il, _ as Dl, a1 as zl, aF as Tl, cs as $l, a9 as El, ct as Ll, aD as Fl, P as Ml, cu as Ol, U as qn, aG as Rl, cv as _l, cw as Bl, T as Wl, aJ as Gl, aI as Pl, q as Ul, aX as Vl, cx as jl, aY as Kl, cy as Hl, aK as ql, at as Zl, ap as Jl, cz as Xl, W as Yl, aq as Ql, S as tu, u as eu, cA as nu, cB as su, cC as iu, aM as ru, cD as au, y as ou, cE as lu, a4 as uu, aO as cu, aN as hu, cF as Ie, cG as pu, g as du, cH as Fs, F as Bt, a2 as Fe, D as fu, w as mu, ah as xe, cI as gu, cJ as bu, m as Ms, cK as yu, cL as Os, cM as wu } from "./index--6vO-cOz.js";
+import { s as ku, a as xu, g as vu, b as Nu, V as d, N as G, r as bn, e as Su, l as Au, c as Zn, f as et, h as ye, i as Ge, j as Jn, k as Pi, m as Ui, t as Rt, R as Et, n as ht, A as Pt, o as K, p as le, q as Xn, u as pt, w as Ht, v as Pe, x as Ue, y as Yn, z as j, B as Ee, C as Gt, D as Cu, E as be, F as en, G as Qn, H as ue, I as Ct, J as nt, K as Ve, L as Iu, M as Du, O as je, P as Ut, Q as Lt, S as ts, T as oe, U as jt, W as Xe, X as ce, Y as zu, Z as Rn, _ as nn, $ as yn, a0 as Vi, a1 as It, a2 as Rs, a3 as Tu, a4 as ji, a5 as $u, a6 as Eu, a7 as Lu, a8 as Fu, a9 as Mu, aa as qt, ab as Ou, ac as es, ad as zt, ae as Ye, af as Ru, ag as _t, ah as ot, ai as Ki, aj as gt, ak as ne, al as _s, am as ve, d as St, an as Bs, ao as Ke, ap as Hi, aq as ns, ar as _u, as as Bu, at as qi, au as Wu } from "./tfjs_backend-DuKis_xG.js";
+import { M as Gu, a as wn, f as Zi } from "./dropout-DFEXTPV0.js";
+import { z as mt } from "./zeros-8xl-W2DC.js";
+import { o as pe } from "./ones-D6kB8bdY.js";
+import { v as Ji } from "./variable-BJTZ3jOy.js";
+import { r as A } from "./reshape-z51Eu-re.js";
+import { s as B } from "./sum-DdkDf2MG.js";
+import { m as Ot } from "./mat_mul-BEHRPMh0.js";
+import { s as Kt } from "./split-B_k_jwud.js";
+import { s as Pu, c as Xi } from "./sin-H567uayl.js";
+import { g as Yi, d as ss, e as Ws, c as Uu } from "./axis_util-QP0LdI1v.js";
+import { a as Zt, e as Jt, l as Vu } from "./log_sum_exp-CiEy1aUe.js";
+import { s as kn } from "./stack-CmqSdsfs.js";
+import { p as ju } from "./slice_util-BdhYwFY_.js";
+import { c as is } from "./concat-DvWM7HGZ.js";
+import { g as Qi } from "./gather-C5D8PxwA.js";
+import { m as at, a as rs } from "./moments-DYOHXoRV.js";
+import { s as tr } from "./softmax-Dsxflvdl.js";
+import { m as Ne } from "./max-BUShNgfh.js";
+import { t as Ku } from "./tensor-BGYi41cj.js";
+import { r as Hu } from "./range-C_vpUjBu.js";
+import { m as qu } from "./norm-DSva3hI3.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -108,7 +108,7 @@ function Qu(s, t, e, n, i, r, a = !1, o = "channelsLast") {
     [l, h, u, c] = s;
   else
     throw new Error(`Unknown dataFormat ${o}`);
-  const [p, f, , g] = t, [b, m] = rn(e), [v, y] = rn(n), C = _n(p, v), N = _n(f, y), { padInfo: I, outHeight: z, outWidth: _ } = nc(i, u, c, b, m, C, N, r, o), T = a ? g * h : g;
+  const [p, f, , g] = t, [b, m] = rn(e), [N, y] = rn(n), C = _n(p, N), v = _n(f, y), { padInfo: I, outHeight: z, outWidth: _ } = nc(i, u, c, b, m, C, v, r, o), T = a ? g * h : g;
   let E;
   return o === "channelsFirst" ? E = [l, T, z, _] : o === "channelsLast" && (E = [l, z, _, T]), {
     batchSize: l,
@@ -125,8 +125,8 @@ function Qu(s, t, e, n, i, r, a = !1, o = "channelsLast") {
     filterHeight: p,
     filterWidth: f,
     effectiveFilterHeight: C,
-    effectiveFilterWidth: N,
-    dilationHeight: v,
+    effectiveFilterWidth: v,
+    dilationHeight: N,
     dilationWidth: y,
     inShape: s,
     outShape: E,
@@ -156,8 +156,8 @@ function nc(s, t, e, n, i, r, a, o, l) {
     c = f[0], h = f[1];
   } else if (s === "same") {
     c = Math.ceil(t / n), h = Math.ceil(e / i);
-    const p = Math.max(0, (c - 1) * n + r - t), f = Math.max(0, (h - 1) * i + a - e), g = Math.floor(p / 2), b = p - g, m = Math.floor(f / 2), v = f - m;
-    u = { top: g, bottom: b, left: m, right: v, type: "SAME" };
+    const p = Math.max(0, (c - 1) * n + r - t), f = Math.max(0, (h - 1) * i + a - e), g = Math.floor(p / 2), b = p - g, m = Math.floor(f / 2), N = f - m;
+    u = { top: g, bottom: b, left: m, right: N, type: "SAME" };
   } else if (s === "valid")
     u = { top: 0, bottom: 0, left: 0, right: 0, type: "VALID" }, c = Math.ceil((t - r + 1) / n), h = Math.ceil((e - a + 1) / i);
   else if (typeof s == "object") {
@@ -419,15 +419,15 @@ function xc(s, t, e, n, i, r = "NHWC", a) {
   t.rank === 3 && (u = !0, l = A(t, [1, t.shape[0], t.shape[1], t.shape[2]]), o = [1, s[0], s[1], s[2]]), k(o.length === 4, () => `Error in conv2dDerInput: inShape must be length 4, but got length ${o.length}.`), k(l.rank === 4, () => `Error in conv2dDerInput: dy must be rank 4, but got rank ${l.rank}`), k(e.rank === 4, () => `Error in conv2dDerInput: filter must be rank 4, but got rank ${e.rank}`);
   const c = r === "NHWC" ? o[3] : o[1], h = r === "NHWC" ? l.shape[3] : l.shape[1];
   k(c === e.shape[2], () => `Error in conv2dDerInput: depth of input (${c}) must match input depth for filter ${e.shape[2]}.`), k(h === e.shape[3], () => `Error in conv2dDerInput: depth of output (${h}) must match output depth for filter ${e.shape[3]}.`), ft("conv2dDerInput", i, a);
-  const p = { dy: l, filter: e }, f = { strides: n, pad: i, dataFormat: r, dimRoundingMode: a, inputShape: o }, g = M.runKernel(Ni, p, f);
+  const p = { dy: l, filter: e }, f = { strides: n, pad: i, dataFormat: r, dimRoundingMode: a, inputShape: o }, g = M.runKernel(vi, p, f);
   return u ? A(g, [g.shape[1], g.shape[2], g.shape[3]]) : g;
 }
 const ls = /* @__PURE__ */ F({ conv2DBackpropInput_: xc });
-function Nc(s, t, e, n, i, r) {
+function vc(s, t, e, n, i, r) {
   const a = D(s, "x", "conv2dTranspose"), o = D(t, "filter", "conv2dTranspose");
   return ls(e, a, o, n, i, "NHWC", r);
 }
-const vc = /* @__PURE__ */ F({ conv2dTranspose_: Nc });
+const Nc = /* @__PURE__ */ F({ conv2dTranspose_: vc });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -448,7 +448,7 @@ function Sc(s, t, e, n, i = "NDHWC", r = [1, 1, 1]) {
   const a = D(s, "x", "conv3d"), o = D(t, "filter", "conv3d");
   let l = a, u = !1;
   a.rank === 4 && (u = !0, l = A(a, [1, a.shape[0], a.shape[1], a.shape[2], a.shape[3]])), k(l.rank === 5, () => `Error in conv3d: input must be rank 5, but got rank ${l.rank}.`), k(o.rank === 5, () => `Error in conv3d: filter must be rank 5, but got rank ${o.rank}.`), k(l.shape[4] === o.shape[3], () => `Error in conv3d: depth of input (${l.shape[4]}) must match input depth for filter ${o.shape[3]}.`), k(de(e, r), () => `Error in conv3D: Either strides or dilations must be 1. Got strides ${e} and dilations '${r}'`), k(i === "NDHWC", () => `Error in conv3d: got dataFormat of ${i} but only NDHWC is currently supported.`), k(Ae(r), () => "Error in conv3D: Dilated rates should be larger than 0."), k(Ae(e), () => "Error in conv3D: Strides should be larger than 0.");
-  const c = { x: l, filter: o }, h = { strides: e, pad: n, dataFormat: i, dilations: r }, p = M.runKernel(vi, c, h);
+  const c = { x: l, filter: o }, h = { strides: e, pad: n, dataFormat: i, dilations: r }, p = M.runKernel(Ni, c, h);
   return u ? A(p, [p.shape[1], p.shape[2], p.shape[3], p.shape[4]]) : p;
 }
 const Ac = /* @__PURE__ */ F({ conv3d_: Sc });
@@ -704,7 +704,7 @@ function Pc(s, t = -1) {
   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 On((i, r) => {
-    const o = ve(i, t, !0), l = V(i, o), u = V(L(l, "float32"), Zt(B(Jt(l), t, !0)));
+    const o = Ne(i, t, !0), l = V(i, o), u = V(L(l, "float32"), Zt(B(Jt(l), t, !0)));
     return r([u]), { value: u, gradFunc: (h, p) => {
       const [f] = p, g = !0, b = Jt(f);
       return V(h, w(B(h, t, g), b));
@@ -1008,9 +1008,9 @@ const hh = /* @__PURE__ */ F({ sinh_: ch });
  * =============================================================================
  */
 function ph(s, t = 0, e = 1, n, i) {
-  if (No(s), n != null && n === "bool")
+  if (vo(s), n != null && n === "bool")
     throw new Error("Unsupported data type $ { dtype }");
-  const r = new Gu(t, e, n, !0, i), a = vo(s, n);
+  const r = new Gu(t, e, n, !0, i), a = No(s, n);
   for (let o = 0; o < a.values.length; o++)
     a.values[o] = r.nextValue();
   return a.toTensor();
@@ -1094,8 +1094,8 @@ function gh({ x: s, filter: t, strides: e, pad: n, dataFormat: i = "NHWC", dilat
   const b = i === "NHWC" ? f.shape[3] : f.shape[1];
   k(p.shape[2] === b, () => `Error in conv2d: depth of input (${b}) must match input depth for filter ${p.shape[2]}.`), k(de(e, r), () => `Error in conv2D: Either strides or dilations must be 1. Got strides ${e} and dilations '${r}'`);
   const m = Qu(f.shape, p.shape, e, r, n, a);
-  let v;
-  o != null && (v = D(o, "bias", "fused conv2d"), [v] = Hn(v, h), i === "NHWC" ? wt(m.outShape, v.shape) : (k(v.shape.length <= 1, () => `Error in fused conv2d: only supports scalar or 1-D Tensor bias for NCHW format but got the bias of rank-${v.shape.length}.`), k(v.shape.length === 0 || v.shape[0] === m.outChannels || v.shape[0] === 1, () => `Error in fused conv2d: bias shape (${v.shape}) is not compatible with the number of output channels (${m.outChannels})`)));
+  let N;
+  o != null && (N = D(o, "bias", "fused conv2d"), [N] = Hn(N, h), i === "NHWC" ? wt(m.outShape, N.shape) : (k(N.shape.length <= 1, () => `Error in fused conv2d: only supports scalar or 1-D Tensor bias for NCHW format but got the bias of rank-${N.shape.length}.`), k(N.shape.length === 0 || N.shape[0] === m.outChannels || N.shape[0] === 1, () => `Error in fused conv2d: bias shape (${N.shape}) is not compatible with the number of output channels (${m.outChannels})`)));
   let y;
   if (u != null) {
     const z = u.shape;
@@ -1112,18 +1112,18 @@ function gh({ x: s, filter: t, strides: e, pad: n, dataFormat: i = "NHWC", dilat
   }
   const C = (z, _) => {
     k(i === "NHWC", () => `Error in gradient of fused conv2D: got dataFormat of ${i} but only NHWC is currently supported.`);
-    const [T, E, R, q] = _, bt = Nu(z, R, l);
+    const [T, E, R, q] = _, bt = vu(z, R, l);
     k(Se(r), () => `Error in gradient of fused conv2D: dilation rates greater than 1 are not yet supported in gradients. Got dilations '${r}'`);
     const ie = ls(E.shape, bt, T, e, n), re = cs(E, bt, T.shape, e, n), xt = [ie, re];
     if (q != null) {
-      const Tt = vu(q, bt);
+      const Tt = Nu(q, bt);
       xt.push(Tt);
     }
     return xt;
-  }, N = {
+  }, v = {
     x: f,
     filter: p,
-    bias: v,
+    bias: N,
     preluActivationWeights: y
   }, I = {
     strides: e,
@@ -1137,13 +1137,13 @@ function gh({ x: s, filter: t, strides: e, pad: n, dataFormat: i = "NHWC", dilat
   return o == null ? On((_, T, E) => {
     let R = (
       // tslint:disable-next-line: no-unnecessary-type-assertion
-      M.runKernel(Ls, N, I)
+      M.runKernel(Ls, v, I)
     );
     return E([T, _, R]), g && (R = A(R, [R.shape[1], R.shape[2], R.shape[3]])), { value: R, gradFunc: C };
   })(f, p) : On((_, T, E, R) => {
-    let q = M.runKernel(Ls, N, I);
+    let q = M.runKernel(Ls, v, I);
     return R([T, _, q, E]), g && (q = A(q, [q.shape[1], q.shape[2], q.shape[3]])), { value: q, gradFunc: C };
-  })(f, p, v);
+  })(f, p, N);
 }
 const bh = /* @__PURE__ */ F({ fusedConv2d_: gh });
 /**
@@ -1213,7 +1213,7 @@ const xh = F({ depthwiseConv2dNativeBackpropInput_: kh });
  * limitations under the License.
  * =============================================================================
  */
-class Nh {
+class vh {
   /**
    * Constructs a `tf.SGDOptimizer` that uses stochastic gradient descent.
    *
@@ -1377,7 +1377,7 @@ class Nh {
  * limitations under the License.
  * =============================================================================
  */
-const ge = Nh;
+const ge = vh;
 /**
  * @license
  * Copyright 2017 Google LLC. All Rights Reserved.
@@ -1394,9 +1394,9 @@ const ge = Nh;
  * limitations under the License.
  * =============================================================================
  */
-const vh = typeof requestAnimationFrame < "u" ? requestAnimationFrame : typeof setImmediate < "u" ? setImmediate : (s) => s();
+const Nh = typeof requestAnimationFrame < "u" ? requestAnimationFrame : typeof setImmediate < "u" ? setImmediate : (s) => s();
 function Sh() {
-  return new Promise((s) => vh(() => s()));
+  return new Promise((s) => Nh(() => s()));
 }
 /**
  * @license
@@ -1998,7 +1998,7 @@ class Mt {
   }
 }
 let Mh = 0;
-class Nn {
+class vn {
   constructor(t, e) {
     this.callArgs = e, this.id = Mh++, 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 n of t.inboundLayers)
@@ -2587,7 +2587,7 @@ class W extends Be {
     const u = [], c = [], h = [];
     for (const p of l)
       u.push(p.sourceLayer), c.push(p.nodeIndex), h.push(p.tensorIndex);
-    new Nn({
+    new vn({
       outboundLayer: this,
       inboundLayers: u,
       nodeIndices: c,
@@ -2751,7 +2751,7 @@ class He extends W {
     const n = t.dtype || "float32";
     this.batchInputShape = e, this.dtype = n, this.inputSpec = [{ shape: e }];
     const i = new Mt(this.dtype, this.batchInputShape, this, [], {}, this.name);
-    i.nodeIndex = 0, i.tensorIndex = 0, new Nn({
+    i.nodeIndex = 0, i.tensorIndex = 0, new vn({
       outboundLayer: this,
       inboundLayers: [],
       nodeIndices: [],
@@ -2932,16 +2932,16 @@ function Le(s, t, e, n) {
     const b = h[g], m = b.sourceLayer;
     if (m instanceof He)
       continue;
-    const v = [], y = [], C = [];
-    let N = !1;
+    const N = [], y = [], C = [];
+    let v = !1;
     for (const E of b.inputs) {
       const R = f.getValue(E), q = f.getMask(E);
-      v.push(R), y.push(q), q != null && (N = !0), i || (p[E.name]--, p[E.name] === 0 && !t.hasKey(E) && o.indexOf(E.name) === -1 && !R.isDisposed && E.sourceLayer.stateful !== !0 && C.push(R));
+      N.push(R), y.push(q), q != null && (v = !0), i || (p[E.name]--, p[E.name] === 0 && !t.hasKey(E) && o.indexOf(E.name) === -1 && !R.isDisposed && E.sourceLayer.stateful !== !0 && C.push(R));
     }
-    N && (e = e || {}, e.mask = y[0]);
-    const I = K(m.apply(v, e));
+    v && (e = e || {}, e.mask = y[0]);
+    const I = K(m.apply(N, e));
     let z = null;
-    m.supportsMasking && (z = m.computeMask(v, y));
+    m.supportsMasking && (z = m.computeMask(N, y));
     const _ = Kh(b), T = Array.isArray(_) ? _ : [_];
     for (let E = 0; E < T.length; ++E) {
       f.hasKey(T[E]) || f.add(T[E], I[E], Array.isArray(z) ? z[0] : z);
@@ -3713,7 +3713,7 @@ const wp = {
  * =============================================================================
  */
 const kp = {
-  kernelName: Ni,
+  kernelName: vi,
   inputsToSave: ["dy", "filter"],
   gradFunc: (s, t, e) => {
     const [n, i] = t, { strides: r, pad: a, dataFormat: o, dimRoundingMode: l } = e;
@@ -3747,7 +3747,7 @@ function xp(s, t, e, n, i) {
   const o = { x: r, dy: a }, l = { strides: n, pad: i, filterShape: e };
   return M.runKernel(el, o, l);
 }
-const Np = /* @__PURE__ */ F({ conv3DBackpropFilter_: xp });
+const vp = /* @__PURE__ */ F({ conv3DBackpropFilter_: xp });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -3764,8 +3764,8 @@ const Np = /* @__PURE__ */ F({ conv3DBackpropFilter_: xp });
  * limitations under the License.
  * =============================================================================
  */
-const vp = {
-  kernelName: vi,
+const Np = {
+  kernelName: Ni,
   inputsToSave: ["x", "filter"],
   gradFunc: (s, t, e) => {
     const { dilations: n, strides: i, pad: r } = e;
@@ -3773,7 +3773,7 @@ const vp = {
     const [a, o] = t;
     return {
       x: () => er(a.shape, s, o, i, r),
-      filter: () => Np(a, s, o.shape, i, r)
+      filter: () => vp(a, s, o.shape, i, r)
     };
   }
 };
@@ -4106,29 +4106,29 @@ const Op = {
   gradFunc: (s, t, e) => {
     const { varianceEpsilon: n } = e, [i, r, a, o] = t, l = o ?? tt(1), u = lt(r.shape, i.shape), c = [];
     if (r.rank === 1) {
-      for (let N = 0; N < i.shape.length - 1; ++N)
-        c.push(i.shape[N]);
+      for (let v = 0; v < i.shape.length - 1; ++v)
+        c.push(i.shape[v]);
       c.push(1);
     }
     const h = V(i, r), p = w(s, l), f = rh($(a, tt(n))), g = w(w(w(f, f), f), tt(-0.5));
     return {
       x: () => r.rank === 1 ? A(w(w(s, Ee(A(f, [1, 1, 1, r.shape[0]]), c)), l), i.shape) : A(w(w(s, f), l), i.shape),
       mean: () => {
-        let N = w(w(f, tt(-1)), p);
-        return r.rank === 1 && (N = B(N, u)), A(N, r.shape);
+        let v = w(w(f, tt(-1)), p);
+        return r.rank === 1 && (v = B(v, u)), A(v, r.shape);
       },
       variance: () => {
-        let N = w(w(g, h), p);
-        return r.rank === 1 && (N = B(N, u)), A(N, r.shape);
+        let v = w(w(g, h), p);
+        return r.rank === 1 && (v = B(v, u)), A(v, r.shape);
       },
       scale: () => {
-        const N = w(h, f);
-        let I = w(s, N);
+        const v = w(h, f);
+        let I = w(s, v);
         return r.rank === 1 && (I = B(I, u)), A(I, r.shape);
       },
       offset: () => {
-        let N = s;
-        return r.rank === 1 && (N = B(N, u)), A(N, r.shape);
+        let v = s;
+        return r.rank === 1 && (v = B(v, u)), A(v, r.shape);
       }
     };
   }
@@ -4154,11 +4154,11 @@ const Rp = {
   inputsToSave: ["x", "indices"],
   gradFunc: (s, t, e) => {
     const [n, i] = t, { axis: r, batchDims: a } = e, o = he(r, n.shape)[0], l = (u, c, h) => () => {
-      const p = u.shape, f = c.size, g = p.slice(0, o), b = g.length, m = p.slice(r, p.length).slice(1), v = m.length, y = qs(0, b), C = qs(b + 1, b + 1 + v), N = Zs([
+      const p = u.shape, f = c.size, g = p.slice(0, o), b = g.length, m = p.slice(r, p.length).slice(1), N = m.length, y = qs(0, b), C = qs(b + 1, b + 1 + N), v = Zs([
         g,
         [f],
         m
-      ]), I = A(h, N), z = A(c, [f]), _ = Zs([[b], y, C]), T = j(I, _);
+      ]), I = A(h, v), z = A(c, [f]), _ = Zs([[b], y, C]), T = j(I, _);
       let E = fh(T, z, u.shape[o]);
       const R = ss(_);
       return E = j(E, R), E;
@@ -4407,7 +4407,7 @@ const Kp = {
  */
 function Hp(s, t, e, n = 5, i = 1, r = 1, a = 0.5) {
   const o = { x: s, y: t, dy: e }, l = { depthRadius: n, bias: i, alpha: r, beta: a };
-  return M.runKernel(Nl, o, l);
+  return M.runKernel(vl, o, l);
 }
 const qp = F({ localResponseNormalizationBackprop_: Hp });
 /**
@@ -4427,7 +4427,7 @@ const qp = F({ localResponseNormalizationBackprop_: Hp });
  * =============================================================================
  */
 const Zp = {
-  kernelName: vl,
+  kernelName: Nl,
   inputsToSave: ["x"],
   outputsToSave: [!0],
   gradFunc: (s, t, e) => {
@@ -5132,7 +5132,7 @@ const xd = {
  * limitations under the License.
  * =============================================================================
  */
-const Nd = {
+const vd = {
   kernelName: Ul,
   inputsToSave: ["x"],
   gradFunc: (s, t) => {
@@ -5156,7 +5156,7 @@ const Nd = {
  * limitations under the License.
  * =============================================================================
  */
-const vd = {
+const Nd = {
   kernelName: Vl,
   inputsToSave: ["images"],
   gradFunc: (s, t, e) => {
@@ -5930,7 +5930,7 @@ const Xd = [
   yp,
   kp,
   wp,
-  vp,
+  Np,
   Sp,
   Ap,
   Cp,
@@ -5979,8 +5979,8 @@ const Xd = [
   wd,
   kd,
   xd,
-  Nd,
   vd,
+  Nd,
   Sd,
   Ad,
   Cd,
@@ -6427,7 +6427,7 @@ class yt {
   }
 }
 yt.constructors = {};
-function Nr(s, t, e, n, i, r, a, o, l) {
+function vr(s, t, e, n, i, r, a, o, l) {
   const u = new nf(), c = [
     new ef(),
     ...yt.createCallbacks(t)
@@ -6473,13 +6473,13 @@ function pn(s, t) {
     return P(s, i);
   });
 }
-function vn(s, t) {
+function Nn(s, t) {
   return x(() => at(je(V(t, s)), -1));
 }
 function xs(s, t) {
   return x(() => at(Fe(V(t, s)), -1));
 }
-function Ns(s, t) {
+function vs(s, t) {
   return x(() => {
     const e = V(s, t), n = Ct(Fe(s), nt(), Number.MAX_VALUE), i = Fe(P(e, n));
     return w(100, at(i, -1));
@@ -6505,7 +6505,7 @@ function of(s, t) {
 }
 function lf(s, t) {
   return x(() => {
-    const e = B(w(s, t), -1), n = ve(w(V(1, s), t), -1);
+    const e = B(w(s, t), -1), n = Ne(w(V(1, s), t), -1);
     return Ie(0, $(1, V(n, e)));
   });
 }
@@ -6560,16 +6560,16 @@ function pf(s, t) {
     return at(V(t, w(s, e)), -1);
   });
 }
-function vr(s, t) {
+function Nr(s, t) {
   return x(() => {
     const e = pn(s, -1), n = pn(t, -1), i = w(e, n);
     return pt(B(i, -1));
   });
 }
 const fn = {
-  meanSquaredError: vn,
+  meanSquaredError: Nn,
   meanAbsoluteError: xs,
-  meanAbsolutePercentageError: Ns,
+  meanAbsolutePercentageError: vs,
   meanSquaredLogarithmicError: rf,
   squaredHinge: af,
   hinge: of,
@@ -6580,7 +6580,7 @@ const fn = {
   binaryCrossentropy: Sn,
   kullbackLeiblerDivergence: hf,
   poisson: pf,
-  cosineProximity: vr
+  cosineProximity: Nr
 };
 function $n(s) {
   if (typeof s == "string") {
@@ -6627,7 +6627,7 @@ function gf(s, t) {
 function bf(s, t) {
   return s.rank === t.rank && (s = ts(s, [s.rank - 1])), t = sn(t, -1), t.dtype !== s.dtype && (t = L(t, s.dtype)), L(Xt(s, t), "float32");
 }
-const yf = vn, wf = vn, kf = xs, xf = xs, Nf = Ns, vf = Ns, Cr = Oe, Sf = vr, Ir = dn, mn = {
+const yf = Nn, wf = Nn, kf = xs, xf = xs, vf = vs, Nf = vs, Cr = Oe, Sf = Nr, Ir = dn, mn = {
   binaryAccuracy: Sr,
   categoricalAccuracy: Ar,
   precision: mf,
@@ -6637,8 +6637,8 @@ const yf = vn, wf = vn, kf = xs, xf = xs, Nf = Ns, vf = Ns, Cr = Oe, Sf = vr, Ir
   MSE: wf,
   mae: kf,
   MAE: xf,
-  mape: Nf,
-  MAPE: vf,
+  mape: vf,
+  MAPE: Nf,
   cosine: Sf
 };
 function Af(s) {
@@ -6924,7 +6924,7 @@ const Ef = (s) => {
   const e = t[0].split("/");
   return !isNaN(parseInt(e[e.length - 1], 10));
 };
-class vt extends W {
+class Nt extends W {
   constructor(t) {
     if (super({}), this.containerNodes = /* @__PURE__ */ new Set(), this.name = t.name, this.name == null) {
       const y = this.getClassName().toLowerCase();
@@ -6934,12 +6934,12 @@ class vt extends W {
       throw new d(`The list of inputs passed to the model is redundant. All inputs should only appear once. Found: ${this.inputs.map((y) => y.name)}`);
     jt(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 C = y.sourceLayer, N = y.nodeIndex, I = y.tensorIndex;
-      this.outputLayers.push(C), this.outputLayersNodeIndices.push(N), this.outputLayersTensorIndices.push(I);
+      const C = y.sourceLayer, v = y.nodeIndex, I = y.tensorIndex;
+      this.outputLayers.push(C), this.outputLayersNodeIndices.push(v), this.outputLayersTensorIndices.push(I);
     }
     for (const y of this.inputs) {
-      const C = y.sourceLayer, N = y.nodeIndex, I = y.tensorIndex;
-      Ut(N === 0, "input layer has >1 nodes"), Ut(I === 0, "input layer has >1 tensors"), this.inputLayers.push(C), this.inputLayersNodeIndices.push(N), this.inputLayersTensorIndices.push(I);
+      const C = y.sourceLayer, v = y.nodeIndex, I = y.tensorIndex;
+      Ut(v === 0, "input layer has >1 nodes"), Ut(I === 0, "input layer has >1 tensors"), this.inputLayers.push(C), this.inputLayersNodeIndices.push(v), this.inputLayersTensorIndices.push(I);
     }
     this.inputNames = [], this.outputNames = [], this.feedInputShapes = [], this.feedInputNames = [], this.feedOutputNames = [];
     for (let y = 0; y < this.inputLayers.length; y++) {
@@ -6951,21 +6951,21 @@ class vt extends W {
     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 = {}, n = {}, i = {}, r = {}, a = {}, o = [], l = (y, C, N, I, z, _) => {
+    const e = {}, n = {}, i = {}, r = {}, a = {}, o = [], l = (y, C, v, I, z, _) => {
       (I == null || z == null || _ == null) && (I = y.sourceLayer, z = y.nodeIndex, _ = y.tensorIndex);
       const T = I.inboundNodes[z];
-      if (N.indexOf(T) !== -1)
+      if (v.indexOf(T) !== -1)
         throw new Et(`The tensor ${y.name} at layer "${I.name}" is part of a cycle.`);
       if (C.indexOf(T) !== -1)
         return;
-      this.containerNodes.add(vt.nodeKey(I, z)), I.id in a || (a[I.id] = Object.keys(a).length), N.indexOf(T) === -1 && N.push(T);
+      this.containerNodes.add(Nt.nodeKey(I, z)), I.id in a || (a[I.id] = Object.keys(a).length), v.indexOf(T) === -1 && v.push(T);
       const E = T.inboundLayers.length;
       for (let R = 0; R < E; R++) {
         const q = T.inputTensors[R], bt = T.inboundLayers[R], ie = T.nodeIndices[R], re = T.tensorIndices[R];
-        l(q, C, N, bt, ie, re);
+        l(q, C, v, bt, ie, re);
       }
-      for (C.push(T); N.indexOf(T) >= 0; )
-        N.splice(N.indexOf(T), 1);
+      for (C.push(T); v.indexOf(T) >= 0; )
+        v.splice(v.indexOf(T), 1);
       o.push(T);
     }, u = [], c = [];
     for (const y of this.outputs)
@@ -6974,8 +6974,8 @@ class vt extends W {
     for (const y of h) {
       n[y.id] = y, y.id in e || (e[y.id] = 0);
       let C = e[y.id];
-      const N = i[y.outboundLayer.id] == null ? 0 : i[y.outboundLayer.id];
-      C = Math.max(C, N), i[y.outboundLayer.id] = C, r[y.outboundLayer.id] = y.outboundLayer, e[y.id] = C;
+      const v = i[y.outboundLayer.id] == null ? 0 : i[y.outboundLayer.id];
+      C = Math.max(C, v), i[y.outboundLayer.id] = C, r[y.outboundLayer.id] = y.outboundLayer, e[y.id] = C;
       for (let I = 0; I < y.inboundLayers.length; I++) {
         const z = y.inboundLayers[I], _ = y.nodeIndices[I], T = z.inboundNodes[_], E = e[T.id] == null ? 0 : e[T.id];
         e[T.id] = Math.max(C + 1, E), n[T.id] = T;
@@ -6995,35 +6995,35 @@ class vt extends W {
     this.layers = [];
     for (const y of g) {
       const C = f[y];
-      C.sort((N, I) => {
-        const z = a[N.id], _ = a[I.id];
+      C.sort((v, I) => {
+        const z = a[v.id], _ = a[I.id];
         return z < _ ? -1 : z > _ ? 1 : 0;
       });
-      for (const N of C)
-        N instanceof vt && this.internalContainerRefs.push(N), this.layers.push(N);
+      for (const v of C)
+        v instanceof Nt && this.internalContainerRefs.push(v), this.layers.push(v);
     }
     this.layersByDepth = f, g = Object.keys(p).map((y) => parseInt(y, 10)).sort(Xe);
     const b = this.inputs.slice(), m = [];
     for (const y of g)
       for (const C of p[y]) {
-        const N = C.outboundLayer;
-        if (N != null) {
+        const v = C.outboundLayer;
+        if (v != null) {
           for (const I of C.inputTensors)
             if (b.indexOf(I) === -1)
-              throw new Et(`Graph disconnected: cannot obtain value for tensor ${I} at layer "${N.name}". The following previous layers were accessed without issue: ${m}`);
+              throw new Et(`Graph disconnected: cannot obtain value for tensor ${I} at layer "${v.name}". The following previous layers were accessed without issue: ${m}`);
           for (const I of C.outputTensors)
             b.push(I);
-          m.push(N.name);
+          m.push(v.name);
         }
       }
     this.nodesByDepth = p;
-    const v = this.layers.map((y) => y.name);
-    for (const y of v) {
-      const C = v.filter((N) => N === y).length;
+    const N = this.layers.map((y) => y.name);
+    for (const y of N) {
+      const C = N.filter((v) => v === y).length;
       if (C !== 1)
-        throw new Et(`The name "${y}" is used ${C} times in the model. All layer names should be unique. Layer names: ` + JSON.stringify(v));
+        throw new Et(`The name "${y}" is used ${C} times in the model. All layer names should be unique. Layer names: ` + JSON.stringify(N));
     }
-    this.outboundNodes = [], this.inboundNodes = [], new Nn({
+    this.outboundNodes = [], this.inboundNodes = [], new vn({
       outboundLayer: this,
       inboundLayers: [],
       nodeIndices: [],
@@ -7255,8 +7255,8 @@ class vt extends W {
             continue;
           const h = [];
           for (let b = 0; b < u.inboundLayers.length; b++) {
-            const m = u.inboundLayers[b], v = u.nodeIndices[b], y = u.tensorIndices[b], C = `${m.name}_${v}_${y}`, N = n[C];
-            h.push(N);
+            const m = u.inboundLayers[b], N = u.nodeIndices[b], y = u.tensorIndices[b], C = `${m.name}_${N}_${y}`, v = n[C];
+            h.push(v);
           }
           const p = c.computeOutputShape(ht(h)), f = ln(p), g = c.inboundNodes.indexOf(u);
           for (let b = 0; b < f.length; b++) {
@@ -7301,16 +7301,16 @@ class vt extends W {
         for (const b of p)
           b.id in n && g.push(n[b.id]);
         if (g.length === p.length) {
-          let b = {}, m, v, y, C;
+          let b = {}, m, N, y, C;
           if (c.callArgs != null && (b = c.callArgs), g.length === 1) {
-            const [N, I] = g[0];
-            b.mask == null && (b.mask = I), y = K(h.call(N, b)), C = K(h.computeMask(N, I)), m = [N], v = [I];
+            const [v, I] = g[0];
+            b.mask == null && (b.mask = I), y = K(h.call(v, b)), C = K(h.computeMask(v, I)), m = [v], N = [I];
           } else
-            m = g.map((N) => N[0]), v = g.map((N) => N[1]), b.mask == null && (b.mask = v), y = K(h.call(m, b)), C = K(h.computeMask(m, v));
+            m = g.map((v) => v[0]), N = g.map((v) => v[1]), b.mask == null && (b.mask = N), y = K(h.call(m, b)), C = K(h.computeMask(m, N));
           if (h.activityRegularizer)
             throw new G("LayersModel invocation with concrete Tensor value(s) in the presence of activity regularizer(s) is not supported yet.");
-          for (let N = 0; N < f.length; ++N) {
-            const I = f[N], z = y[N], _ = C[N];
+          for (let v = 0; v < f.length; ++v) {
+            const I = f[v], z = y[v], _ = C[v];
             n[I.id] = [z, _];
           }
         }
@@ -7336,9 +7336,9 @@ class vt extends W {
     const e = {};
     let n;
     for (const i of this.layers) {
-      n = i instanceof vt ? 1 : 0;
+      n = i instanceof Nt ? 1 : 0;
       for (let r = 0; r < i.inboundNodes.length; r++) {
-        const a = vt.nodeKey(i, r);
+        const a = Nt.nodeKey(i, r);
         this.containerNodes.has(a) && (e[a] = n, n += 1);
       }
     }
@@ -7371,7 +7371,7 @@ class vt extends W {
       const t = [];
       for (const e of this.layers)
         for (let n = 0; n < e.inboundNodes.length; ++n) {
-          const i = vt.nodeKey(e, n);
+          const i = Nt.nodeKey(e, n);
           this.containerNodes.has(i) && t.push(...e.calculateLosses());
         }
       return t;
@@ -7382,7 +7382,7 @@ class vt extends W {
     for (const a of this.layers) {
       const o = a.getClassName(), l = a.getConfig(), u = [];
       for (let h = 0; h < a.inboundNodes.length; h++) {
-        const p = a.inboundNodes[h], f = vt.nodeKey(a, h);
+        const p = a.inboundNodes[h], f = Nt.nodeKey(a, h);
         let g = {};
         if (this.containerNodes.has(f)) {
           if (p.callArgs)
@@ -7394,9 +7394,9 @@ class vt extends W {
           if (p.inboundLayers.length > 0) {
             const b = [];
             for (let m = 0; m < p.inboundLayers.length; m++) {
-              const v = p.inboundLayers[m], y = p.nodeIndices[m], C = p.tensorIndices[m], N = vt.nodeKey(v, y);
-              let I = e[N];
-              I == null && (I = 0), b.push([v.name, I, C, g]);
+              const N = p.inboundLayers[m], y = p.nodeIndices[m], C = p.tensorIndices[m], v = Nt.nodeKey(N, y);
+              let I = e[v];
+              I == null && (I = 0), b.push([N.name, I, C, g]);
             }
             u.push(b);
           }
@@ -7408,7 +7408,7 @@ class vt extends W {
     t.layers = n;
     const i = [];
     for (let a = 0; a < this.inputLayers.length; a++) {
-      const o = this.inputLayers[a], l = this.inputLayersNodeIndices[a], u = vt.nodeKey(o, l);
+      const o = this.inputLayers[a], l = this.inputLayersNodeIndices[a], u = Nt.nodeKey(o, l);
       if (!this.containerNodes.has(u))
         continue;
       let c = e[u];
@@ -7419,7 +7419,7 @@ class vt extends W {
     t.inputLayers = i;
     const r = [];
     for (let a = 0; a < this.outputLayers.length; a++) {
-      const o = this.outputLayers[a], l = this.outputLayersNodeIndices[a], u = vt.nodeKey(o, l);
+      const o = this.outputLayers[a], l = this.outputLayersNodeIndices[a], u = Nt.nodeKey(o, l);
       if (!this.containerNodes.has(u))
         continue;
       let c = e[u];
@@ -7444,21 +7444,21 @@ class vt extends W {
   /** @nocollapse */
   static fromConfig(t, e, n = {}, i = !1) {
     const r = {}, a = {};
-    function o(m, v) {
-      m.name in a ? a[m.name].push(v) : a[m.name] = [v];
+    function o(m, N) {
+      m.name in a ? a[m.name].push(N) : a[m.name] = [N];
     }
-    function l(m, v) {
+    function l(m, N) {
       const y = [];
       let C;
-      for (const N of v) {
-        const I = N[0], z = N[1], _ = N[2];
-        if (C = N[3] == null ? {} : N[3], !(I in r)) {
-          o(m, v);
+      for (const v of N) {
+        const I = v[0], z = v[1], _ = v[2];
+        if (C = v[3] == null ? {} : v[3], !(I in r)) {
+          o(m, N);
           return;
         }
         const T = r[I];
         if (T.inboundNodes.length <= z) {
-          o(m, v);
+          o(m, N);
           return;
         }
         const E = T.inboundNodes[z];
@@ -7467,11 +7467,11 @@ class vt extends W {
       y.length > 0 && m.apply(ht(y), C);
     }
     function u(m) {
-      const v = m.name, y = Wt(m, e.customObjects != null ? e.customObjects : {});
-      y.setFastWeightInitDuringBuild(i), r[v] = y, m.inboundNodes.forEach((N) => {
-        if (!(N instanceof Array))
-          throw new d(`Corrupted configuration, expected array for nodeData: ${N}`);
-        o(y, N);
+      const N = m.name, y = Wt(m, e.customObjects != null ? e.customObjects : {});
+      y.setFastWeightInitDuringBuild(i), r[N] = y, m.inboundNodes.forEach((v) => {
+        if (!(v instanceof Array))
+          throw new d(`Corrupted configuration, expected array for nodeData: ${v}`);
+        o(y, v);
       });
     }
     const c = e.name, h = e.layers;
@@ -7479,26 +7479,26 @@ class vt extends W {
       u(m);
     for (; !zu(a); )
       for (const m of h) {
-        const v = r[m.name];
-        if (v.name in a) {
-          const y = a[v.name];
-          delete a[v.name];
+        const N = r[m.name];
+        if (N.name in a) {
+          const y = a[N.name];
+          delete a[N.name];
           for (const C of y)
-            l(v, C);
+            l(N, C);
         }
       }
     const p = [], f = [], g = e.inputLayers;
     for (const m of g) {
-      const v = m[0], y = m[1], C = m[2];
-      Ut(v in r);
-      const I = r[v].inboundNodes[y].outputTensors;
+      const N = m[0], y = m[1], C = m[2];
+      Ut(N in r);
+      const I = r[N].inboundNodes[y].outputTensors;
       p.push(I[C]);
     }
     const b = e.outputLayers;
     for (const m of b) {
-      const v = m[0], y = m[1], C = m[2];
-      Ut(v in r);
-      const I = r[v].inboundNodes[y].outputTensors;
+      const N = m[0], y = m[1], C = m[2];
+      Ut(N in r);
+      const I = r[N].inboundNodes[y].outputTensors;
       f.push(I[C]);
     }
     return new t({ inputs: p, outputs: f, name: c });
@@ -7652,7 +7652,7 @@ async function Rf(s, t, e) {
     const o = s.makeTrainFunction(), l = s.getDedupedMetricsNames();
     let u;
     i ? u = l.slice().concat(l.map((m) => "val_" + m)) : u = l.slice();
-    const c = xr(e.callbacks, e.yieldEvery), h = e.verbose == null ? 1 : e.verbose, { callbackList: p, history: f } = Nr(
+    const c = xr(e.callbacks, e.yieldEvery), h = e.verbose == null ? 1 : e.verbose, { callbackList: p, history: f } = vr(
       c,
       h,
       e.epochs,
@@ -7669,39 +7669,39 @@ async function Rf(s, t, e) {
     for (; g < e.epochs; ) {
       const m = {};
       await p.onEpochBegin(g);
-      let v = 0, y = 0;
-      for (n || (b = await t.iterator()); !n || v < e.batchesPerEpoch; ) {
+      let N = 0, y = 0;
+      for (n || (b = await t.iterator()); !n || N < e.batchesPerEpoch; ) {
         const C = await b.next();
         if (n && C.done) {
-          console.warn(`You provided \`batchesPerEpoch\` as ${e.batchesPerEpoch}, but your dataset iterator ran out of data after ${v} 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.`);
+          console.warn(`You provided \`batchesPerEpoch\` as ${e.batchesPerEpoch}, but your dataset iterator ran out of data after ${N} 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 (C.value != null) {
-          const { xs: N, ys: I } = Er(s, C.value), z = {};
-          z.batch = y, z.size = N[0].shape[0], await p.onBatchBegin(y, z);
+          const { xs: v, ys: I } = Er(s, C.value), z = {};
+          z.batch = y, z.size = v[0].shape[0], await p.onBatchBegin(y, z);
           const _ = [];
           if (e.classWeight != null) {
             const R = Tr(e.classWeight, s.outputNames);
             for (let q = 0; q < R.length; ++q)
               _.push(await $r(I[q], null, R[q]));
           }
-          const T = N.concat(I).concat(_), E = o(T);
+          const T = v.concat(I).concat(_), E = o(T);
           Z(T);
           for (let R = 0; R < l.length; ++R) {
             const q = l[R], bt = E[R];
             z[q] = bt, Bt(bt);
           }
-          await p.onBatchEnd(y, z), kr(z), y++, v++;
+          await p.onBatchEnd(y, z), kr(z), y++, N++;
         }
-        if (n ? v >= e.batchesPerEpoch : C.done) {
+        if (n ? N >= e.batchesPerEpoch : C.done) {
           if (i) {
-            let N;
-            ai(e.validationData) ? N = K(await s.evaluateDataset(e.validationData, { batches: e.validationBatches })) : N = K(s.evaluate(r, a, {
+            let v;
+            ai(e.validationData) ? v = K(await s.evaluateDataset(e.validationData, { batches: e.validationBatches })) : v = K(s.evaluate(r, a, {
               batchSize: e.validationBatchSize == null ? Mf : e.validationBatchSize,
               verbose: 0
             }));
             for (let I = 0; I < s.metricsNames.length; ++I)
-              m[`val_${s.metricsNames[I]}`] = N[I];
+              m[`val_${s.metricsNames[I]}`] = v[I];
           }
           break;
         }
@@ -7745,8 +7745,8 @@ async function Wf(s, t, e) {
             r.push(tt(0));
         const g = p[0].shape[0];
         for (let b = 0; b < f.length; ++b) {
-          const m = f[b], v = r[b];
-          r[b] = x(() => $(r[b], w(g, m))), l > 0 && Z(v);
+          const m = f[b], N = r[b];
+          r[b] = x(() => $(r[b], w(g, m))), l > 0 && Z(N);
         }
         Z(f), o += g, ++l;
       }
@@ -7802,7 +7802,7 @@ function Lr(s) {
   }
   return t;
 }
-function Nt(s, t) {
+function vt(s, t) {
   if (s == null)
     return;
   const e = [];
@@ -7917,7 +7917,7 @@ function Pf(s, t, e) {
 }
 function Uf(s, t, e) {
   const n = [
-    vn,
+    Nn,
     Sn,
     Oe
   ];
@@ -7986,7 +7986,7 @@ function Vf(s, t) {
   }
 }
 const jf = "layers-model";
-class we extends vt {
+class we extends Nt {
   constructor(t) {
     super(t), this.isTraining = !1;
   }
@@ -8094,8 +8094,8 @@ class we extends vt {
             if (typeof g == "string" && ["accuracy", "acc", "crossentropy", "ce"].indexOf(g) !== -1) {
               const m = this.internalOutputShapes[a];
               m[m.length - 1] === 1 || this.lossFunctions[a] === Sn ? ["accuracy", "acc"].indexOf(g) !== -1 ? p = Sr : ["crossentropy", "ce"].indexOf(g) !== -1 && (p = gf) : this.lossFunctions[a] === dn ? ["accuracy", "acc"].indexOf(g) !== -1 ? p = bf : ["crossentropy", "ce"].indexOf(g) !== -1 && (p = Ir) : ["accuracy", "acc"].indexOf(g) !== -1 ? p = Ar : ["crossentropy", "ce"].indexOf(g) !== -1 && (p = Cr);
-              let v;
-              ["accuracy", "acc"].indexOf(g) !== -1 ? v = "acc" : ["crossentropy", "ce"].indexOf(g) !== -1 && (v = "ce"), f = p, h = "" + v;
+              let N;
+              ["accuracy", "acc"].indexOf(g) !== -1 ? N = "acc" : ["crossentropy", "ce"].indexOf(g) !== -1 && (N = "ce"), f = p, h = "" + N;
             } else
               f = Af(g), h = "" + tn(g);
             let b;
@@ -8160,7 +8160,7 @@ class we extends vt {
       const l = this.testFunction, u = this.testLoop(l, o, i, n.verbose, n.steps);
       return ht(u);
     } finally {
-      Nt(a[0], t), Nt(a[1], e);
+      vt(a[0], t), vt(a[1], e);
     }
   }
   // TODO(cais): Add code snippet below once real dataset objects are
@@ -8326,7 +8326,7 @@ class we extends vt {
       const i = e.batchSize == null ? 32 : e.batchSize;
       return En(i), this.predictLoop(n, i);
     } finally {
-      Nt(n, t);
+      vt(n, t);
     }
   }
   /**
@@ -8400,8 +8400,8 @@ class we extends vt {
             for (let m = 0; m < b.length; ++m)
               o.push(tt(0));
           for (let m = 0; m < b.length; ++m) {
-            const v = b[m];
-            o[m] = $(o[m], w(p - h, v));
+            const N = b[m];
+            o[m] = $(o[m], w(p - h, N));
           }
         }
         for (let c = 0; c < o.length; ++c)
@@ -8443,18 +8443,18 @@ class we extends vt {
         let g;
         for (let b = 0; b < this.lossFunctions.length; ++b) {
           const m = this.lossFunctions[b];
-          let v = m(i[b], f[b]);
-          r[b] != null && (v = Ff(v, r[b]));
-          const y = at(v);
-          e.push(y), b === 0 ? g = v : g = $(g, v);
+          let N = m(i[b], f[b]);
+          r[b] != null && (N = Ff(N, r[b]));
+          const y = at(N);
+          e.push(y), b === 0 ? g = N : g = $(g, N);
         }
         for (let b = 0; b < this.metricsTensors.length; ++b) {
           let m;
           if (this.outputs.length > 1 && b < this.outputs.length)
             m = e[b];
           else {
-            const v = this.metricsTensors[b][0], y = this.metricsTensors[b][1];
-            m = at(v(i[y], f[y]));
+            const N = this.metricsTensors[b][0], y = this.metricsTensors[b][1];
+            m = at(N(i[y], f[y]));
           }
           Bt(m), a.push(m);
         }
@@ -8533,7 +8533,7 @@ class we extends vt {
       En(f);
       const b = await this.standardizeUserData(t, e, n.sampleWeight, n.classWeight, !1, f);
       i = b[0], r = b[1], p = b[2];
-      let m = !1, v;
+      let m = !1, N;
       if (n.validationData != null && n.validationData.length > 0) {
         if (m = !0, n.validationData.length === 2)
           l = n.validationData[0], u = n.validationData[1];
@@ -8547,21 +8547,21 @@ class we extends vt {
           !0,
           f
         );
-        c = R[0], h = R[1], v = c.concat(h);
+        c = R[0], h = R[1], N = c.concat(h);
       } else if (n.validationSplit != null && n.validationSplit > 0 && n.validationSplit < 1) {
         m = !0;
         const E = Math.floor(i[0].shape[0] * (1 - n.validationSplit)), R = i[0].shape[0];
-        c = Te(i, E, R), a = i, i = Te(i, 0, E), h = Te(r, E, R), o = r, r = Te(r, 0, E), v = c.concat(h);
+        c = Te(i, E, R), a = i, i = Te(i, 0, E), h = Te(r, E, R), o = r, r = Te(r, 0, E), N = c.concat(h);
       } else n.validationSteps != null && (m = !0);
       const y = i.concat(r).concat(p);
       this.checkTrainableWeightsConsistency();
-      const C = this.makeTrainFunction(), N = this.getDedupedMetricsNames();
+      const C = this.makeTrainFunction(), v = this.getDedupedMetricsNames();
       let I, z;
-      m ? (this.makeTestFunction(), I = this.testFunction, z = N.slice().concat(N.map((E) => "val_" + E))) : (I = null, v = [], z = N.slice());
+      m ? (this.makeTestFunction(), I = this.testFunction, z = v.slice().concat(v.map((E) => "val_" + E))) : (I = null, N = [], z = v.slice());
       const _ = xr(n.callbacks, n.yieldEvery);
-      return await this.fitLoop(C, y, N, f, n.epochs, n.verbose, _, I, v, n.shuffle, z, n.initialEpoch, null, null);
+      return await this.fitLoop(C, y, v, f, n.epochs, n.verbose, _, I, N, n.shuffle, z, n.initialEpoch, null, null);
     } finally {
-      this.isTraining = !1, Nt(i, t), Nt(r, e), Nt(a, t), Nt(o, e), Nt(c, l), Nt(h, u), p != null && Z(p);
+      this.isTraining = !1, vt(i, t), vt(r, e), vt(a, t), vt(o, e), vt(c, l), vt(h, u), p != null && Z(p);
     }
   }
   /**
@@ -8597,20 +8597,20 @@ class we extends vt {
     if (l != null && u != null && (b = !0), g != null && (b = !0, f == null))
       throw new d("Can only use `validationSteps` when doing step-wise training, i.e., `stepsPerEpoch` must be set.");
     const m = this.checkNumSamples(e, i, f, "steps_per_epoch");
-    let v;
-    m != null && (v = It(0, m)), a == null && (a = 1);
-    const { callbackList: y, history: C } = Nr(o, a, r, p, m, f, i, b, h);
+    let N;
+    m != null && (N = It(0, m)), a == null && (a = 1);
+    const { callbackList: y, history: C } = vr(o, a, r, p, m, f, i, b, h);
     y.setModel(this), this.history = C, await y.onTrainBegin(), this.stopTraining_ = !1;
-    for (let N = p; N < r; ++N) {
-      await y.onEpochBegin(N);
+    for (let v = p; v < r; ++v) {
+      await y.onEpochBegin(v);
       const I = {};
       if (f != null)
         throw new G("stepsPerEpoch mode is not implemented yet.");
       {
         if (c === "batch")
           throw new G("batch shuffling is not implemneted yet");
-        c && bu(v);
-        const z = Rn(v), _ = Ln(m, i);
+        c && bu(N);
+        const z = Rn(N), _ = Ln(m, i);
         for (let T = 0; T < _.length; ++T) {
           const E = {};
           if (await y.onBatchBegin(T, E), x(() => {
@@ -8633,7 +8633,7 @@ class we extends vt {
         }
         z.dispose();
       }
-      if (await y.onEpochEnd(N, I), this.stopTraining_)
+      if (await y.onEpochEnd(v, I), this.stopTraining_)
         break;
     }
     return await y.onTrainEnd(), await this.history.syncData(), this.history;
@@ -8693,7 +8693,7 @@ class we extends vt {
       const c = await u.data();
       l.push(c[0]);
     }
-    return Z(o), Nt(n[0], t), Nt(n[1], e), ht(l);
+    return Z(o), vt(n[0], t), vt(n[1], e), ht(l);
   }
   /**
    * Extract weight values of the model.
@@ -9039,7 +9039,7 @@ class Re extends we {
           throw new d("All layers in a Sequential model should have a single output tensor. For multi-output layers, use the functional API.");
         this.checkShape(t), this.outputs = [t.inboundNodes[0].outputTensors[0]], this.inputs = dr(this.outputs[0]);
       }
-      this.inboundNodes = [], new Nn({
+      this.inboundNodes = [], new vn({
         outboundLayer: this,
         inboundLayers: [],
         nodeIndices: [],
@@ -9563,7 +9563,7 @@ class Vr extends ut {
 }
 Vr.className = "tanh";
 S(Vr);
-let vs = class extends ut {
+let Ns = class extends ut {
   /**
    * Calculate the activation function.
    *
@@ -9580,8 +9580,8 @@ let vs = class extends ut {
     return tr(t, e);
   }
 };
-vs.className = "softmax";
-S(vs);
+Ns.className = "softmax";
+S(Ns);
 class jr extends ut {
   /**
    * Calculate the activation function of log softmax:
@@ -9858,7 +9858,7 @@ na.className = "ThresholdedReLU";
 S(na);
 class sa extends W {
   constructor(t) {
-    super(t ?? {}), this.DEFAULT_AXIS = 1, t == null && (t = {}), this.softmax = new vs().apply, this.axis = t.axis == null ? this.DEFAULT_AXIS : t.axis;
+    super(t ?? {}), this.DEFAULT_AXIS = 1, t == null && (t = {}), this.softmax = new Ns().apply, this.axis = t.axis == null ? this.DEFAULT_AXIS : t.axis;
   }
   call(t, e) {
     return x(() => {
@@ -10132,8 +10132,8 @@ class ra extends Ze {
       this.dataFormat === "channelsFirst" ? (a = 2, o = 3) : (a = 1, o = 2);
       const l = i[a], u = i[o], c = this.kernelSize[0], h = this.kernelSize[1], p = this.strides[0], f = this.strides[1], g = $t(l, p, c, this.padding), b = $t(u, f, h, this.padding), m = [r, g, b, this.filters];
       this.dataFormat !== "channelsLast" && (n = j(n, [0, 2, 3, 1]));
-      let v = vc(n, this.kernel.read(), m, this.strides, this.padding);
-      return this.dataFormat !== "channelsLast" && (v = j(v, [0, 3, 1, 2])), this.bias != null && (v = zt(v, this.bias.read(), this.dataFormat)), this.activation != null && (v = this.activation.apply(v)), v;
+      let N = Nc(n, this.kernel.read(), m, this.strides, this.padding);
+      return this.dataFormat !== "channelsLast" && (N = j(N, [0, 3, 1, 2])), this.bias != null && (N = zt(N, this.bias.read(), this.dataFormat)), this.activation != null && (N = this.activation.apply(N)), N;
     });
   }
   computeOutputShape(t) {
@@ -10173,7 +10173,7 @@ class aa extends Je {
       const i = n.shape, r = i[0];
       let a, o, l;
       this.dataFormat === "channelsFirst" ? (l = 2, a = 3, o = 4) : (l = 1, a = 2, o = 3);
-      const u = i[l], c = i[a], h = i[o], p = this.kernelSize[0], f = this.kernelSize[1], g = this.kernelSize[2], b = this.strides[0], m = this.strides[1], v = this.strides[2], y = $t(u, b, p, this.padding), C = $t(c, m, f, this.padding), N = $t(h, v, g, this.padding), I = [r, y, C, N, this.filters];
+      const u = i[l], c = i[a], h = i[o], p = this.kernelSize[0], f = this.kernelSize[1], g = this.kernelSize[2], b = this.strides[0], m = this.strides[1], N = this.strides[2], y = $t(u, b, p, this.padding), C = $t(c, m, f, this.padding), v = $t(h, N, g, this.padding), I = [r, y, C, v, this.filters];
       this.dataFormat !== "channelsLast" && (n = j(n, [0, 2, 3, 4, 1]));
       let z = Dc(n, this.kernel.read(), I, this.strides, this.padding);
       return this.dataFormat !== "channelsLast" && (z = j(z, [0, 4, 1, 2, 3])), this.bias !== null && (z = zt(z, this.bias.read(), this.dataFormat)), this.activation !== null && (z = this.activation.apply(z)), z;
@@ -10419,16 +10419,16 @@ function da(s, t, e, n = !1, i, r, a = !1, o = !1) {
     const f = t.shape[0], g = en(t);
     let b;
     i != null && (b = en(i));
-    for (let v = 0; v < f; ++v) {
-      const y = g[v], C = x(() => s(y, p));
+    for (let N = 0; N < f; ++N) {
+      const y = g[N], C = x(() => s(y, p));
       if (i == null)
         h = C[0], p = C[1];
       else {
-        const N = x(() => {
-          const I = b[v], z = V(Dt(I), I), _ = $(w(C[0], I), w(p[0], z)), T = p.map((E, R) => $(w(C[1][R], I), w(E, z)));
+        const v = x(() => {
+          const I = b[N], z = V(Dt(I), I), _ = $(w(C[0], I), w(p[0], z)), T = p.map((E, R) => $(w(C[1][R], I), w(E, z)));
           return { output: _, newStates: T };
         });
-        h = N.output, p = N.newStates;
+        h = v.output, p = v.newStates;
       }
       o && c.push(h);
     }
@@ -10642,7 +10642,7 @@ class In extends W {
 }
 class As extends In {
   constructor(t) {
-    super(t), this.DEFAULT_ACTIVATION = "tanh", this.DEFAULT_KERNEL_INITIALIZER = "glorotNormal", this.DEFAULT_RECURRENT_INITIALIZER = "orthogonal", this.DEFAULT_BIAS_INITIALIZER = "zeros", this.units = t.units, ot(this.units, "units"), this.activation = Qt(t.activation == null ? this.DEFAULT_ACTIVATION : t.activation), this.useBias = t.useBias == null ? !0 : t.useBias, this.kernelInitializer = J(t.kernelInitializer || this.DEFAULT_KERNEL_INITIALIZER), this.recurrentInitializer = J(t.recurrentInitializer || this.DEFAULT_RECURRENT_INITIALIZER), this.biasInitializer = J(t.biasInitializer || this.DEFAULT_BIAS_INITIALIZER), this.kernelRegularizer = X(t.kernelRegularizer), this.recurrentRegularizer = X(t.recurrentRegularizer), this.biasRegularizer = X(t.biasRegularizer), this.kernelConstraint = rt(t.kernelConstraint), this.recurrentConstraint = rt(t.recurrentConstraint), this.biasConstraint = rt(t.biasConstraint), this.dropout = Ne([1, qt([0, t.dropout == null ? 0 : t.dropout])]), this.recurrentDropout = Ne([
+    super(t), this.DEFAULT_ACTIVATION = "tanh", this.DEFAULT_KERNEL_INITIALIZER = "glorotNormal", this.DEFAULT_RECURRENT_INITIALIZER = "orthogonal", this.DEFAULT_BIAS_INITIALIZER = "zeros", this.units = t.units, ot(this.units, "units"), this.activation = Qt(t.activation == null ? this.DEFAULT_ACTIVATION : t.activation), this.useBias = t.useBias == null ? !0 : t.useBias, this.kernelInitializer = J(t.kernelInitializer || this.DEFAULT_KERNEL_INITIALIZER), this.recurrentInitializer = J(t.recurrentInitializer || this.DEFAULT_RECURRENT_INITIALIZER), this.biasInitializer = J(t.biasInitializer || this.DEFAULT_BIAS_INITIALIZER), this.kernelRegularizer = X(t.kernelRegularizer), this.recurrentRegularizer = X(t.recurrentRegularizer), this.biasRegularizer = X(t.biasRegularizer), this.kernelConstraint = rt(t.kernelConstraint), this.recurrentConstraint = rt(t.recurrentConstraint), this.biasConstraint = rt(t.biasConstraint), this.dropout = ve([1, qt([0, t.dropout == null ? 0 : t.dropout])]), this.recurrentDropout = ve([
       1,
       qt([0, t.recurrentDropout == null ? 0 : t.recurrentDropout])
     ]), this.dropoutFunc = t.dropoutFunc, this.stateSize = this.units, this.dropoutMask = null, this.recurrentDropoutMask = null;
@@ -10726,7 +10726,7 @@ class Cs extends In {
   constructor(t) {
     if (super(t), this.DEFAULT_ACTIVATION = "tanh", this.DEFAULT_RECURRENT_ACTIVATION = "hardSigmoid", this.DEFAULT_KERNEL_INITIALIZER = "glorotNormal", this.DEFAULT_RECURRENT_INITIALIZER = "orthogonal", this.DEFAULT_BIAS_INITIALIZER = "zeros", t.resetAfter)
       throw new d("GRUCell does not support reset_after parameter set to true.");
-    this.units = t.units, ot(this.units, "units"), this.activation = Qt(t.activation === void 0 ? this.DEFAULT_ACTIVATION : t.activation), this.recurrentActivation = Qt(t.recurrentActivation === void 0 ? this.DEFAULT_RECURRENT_ACTIVATION : t.recurrentActivation), this.useBias = t.useBias == null ? !0 : t.useBias, this.kernelInitializer = J(t.kernelInitializer || this.DEFAULT_KERNEL_INITIALIZER), this.recurrentInitializer = J(t.recurrentInitializer || this.DEFAULT_RECURRENT_INITIALIZER), this.biasInitializer = J(t.biasInitializer || this.DEFAULT_BIAS_INITIALIZER), this.kernelRegularizer = X(t.kernelRegularizer), this.recurrentRegularizer = X(t.recurrentRegularizer), this.biasRegularizer = X(t.biasRegularizer), this.kernelConstraint = rt(t.kernelConstraint), this.recurrentConstraint = rt(t.recurrentConstraint), this.biasConstraint = rt(t.biasConstraint), this.dropout = Ne([1, qt([0, t.dropout == null ? 0 : t.dropout])]), this.recurrentDropout = Ne([
+    this.units = t.units, ot(this.units, "units"), this.activation = Qt(t.activation === void 0 ? this.DEFAULT_ACTIVATION : t.activation), this.recurrentActivation = Qt(t.recurrentActivation === void 0 ? this.DEFAULT_RECURRENT_ACTIVATION : t.recurrentActivation), this.useBias = t.useBias == null ? !0 : t.useBias, this.kernelInitializer = J(t.kernelInitializer || this.DEFAULT_KERNEL_INITIALIZER), this.recurrentInitializer = J(t.recurrentInitializer || this.DEFAULT_RECURRENT_INITIALIZER), this.biasInitializer = J(t.biasInitializer || this.DEFAULT_BIAS_INITIALIZER), this.kernelRegularizer = X(t.kernelRegularizer), this.recurrentRegularizer = X(t.recurrentRegularizer), this.biasRegularizer = X(t.biasRegularizer), this.kernelConstraint = rt(t.kernelConstraint), this.recurrentConstraint = rt(t.recurrentConstraint), this.biasConstraint = rt(t.biasConstraint), this.dropout = ve([1, qt([0, t.dropout == null ? 0 : t.dropout])]), this.recurrentDropout = ve([
       1,
       qt([0, t.recurrentDropout == null ? 0 : t.recurrentDropout])
     ]), this.dropoutFunc = t.dropoutFunc, this.implementation = t.implementation, this.stateSize = this.units, this.dropoutMask = null, this.recurrentDropoutMask = null;
@@ -10760,10 +10760,10 @@ class Cs extends In {
       0 < this.dropout && this.dropout < 1 && (t = w(t, r[0]));
       let c = St(t, this.kernel.read());
       this.useBias && (c = zt(c, this.bias.read())), 0 < this.recurrentDropout && this.recurrentDropout < 1 && (i = w(i, a[0]));
-      const h = this.recurrentKernel.read(), [p, f] = Kt(h, [2 * this.units, this.units], h.rank - 1), g = St(i, p), [b, m, v] = Kt(c, 3, c.rank - 1), [y, C] = Kt(g, 2, g.rank - 1);
+      const h = this.recurrentKernel.read(), [p, f] = Kt(h, [2 * this.units, this.units], h.rank - 1), g = St(i, p), [b, m, N] = Kt(c, 3, c.rank - 1), [y, C] = Kt(g, 2, g.rank - 1);
       o = this.recurrentActivation.apply($(b, y)), l = this.recurrentActivation.apply($(m, C));
-      const N = St(w(l, i), f);
-      u = this.activation.apply($(v, N));
+      const v = St(w(l, i), f);
+      u = this.activation.apply($(N, v));
       const I = $(w(o, i), w($(1, pt(o)), u));
       return [I, I];
     });
@@ -10814,7 +10814,7 @@ ma.className = "GRU";
 S(ma);
 class Dn extends In {
   constructor(t) {
-    super(t), this.DEFAULT_ACTIVATION = "tanh", this.DEFAULT_RECURRENT_ACTIVATION = "hardSigmoid", this.DEFAULT_KERNEL_INITIALIZER = "glorotNormal", this.DEFAULT_RECURRENT_INITIALIZER = "orthogonal", this.DEFAULT_BIAS_INITIALIZER = "zeros", this.units = t.units, ot(this.units, "units"), this.activation = Qt(t.activation === void 0 ? this.DEFAULT_ACTIVATION : t.activation), this.recurrentActivation = Qt(t.recurrentActivation === void 0 ? this.DEFAULT_RECURRENT_ACTIVATION : t.recurrentActivation), this.useBias = t.useBias == null ? !0 : t.useBias, this.kernelInitializer = J(t.kernelInitializer || this.DEFAULT_KERNEL_INITIALIZER), this.recurrentInitializer = J(t.recurrentInitializer || this.DEFAULT_RECURRENT_INITIALIZER), this.biasInitializer = J(t.biasInitializer || this.DEFAULT_BIAS_INITIALIZER), this.unitForgetBias = t.unitForgetBias, this.kernelRegularizer = X(t.kernelRegularizer), this.recurrentRegularizer = X(t.recurrentRegularizer), this.biasRegularizer = X(t.biasRegularizer), this.kernelConstraint = rt(t.kernelConstraint), this.recurrentConstraint = rt(t.recurrentConstraint), this.biasConstraint = rt(t.biasConstraint), this.dropout = Ne([1, qt([0, t.dropout == null ? 0 : t.dropout])]), this.recurrentDropout = Ne([
+    super(t), this.DEFAULT_ACTIVATION = "tanh", this.DEFAULT_RECURRENT_ACTIVATION = "hardSigmoid", this.DEFAULT_KERNEL_INITIALIZER = "glorotNormal", this.DEFAULT_RECURRENT_INITIALIZER = "orthogonal", this.DEFAULT_BIAS_INITIALIZER = "zeros", this.units = t.units, ot(this.units, "units"), this.activation = Qt(t.activation === void 0 ? this.DEFAULT_ACTIVATION : t.activation), this.recurrentActivation = Qt(t.recurrentActivation === void 0 ? this.DEFAULT_RECURRENT_ACTIVATION : t.recurrentActivation), this.useBias = t.useBias == null ? !0 : t.useBias, this.kernelInitializer = J(t.kernelInitializer || this.DEFAULT_KERNEL_INITIALIZER), this.recurrentInitializer = J(t.recurrentInitializer || this.DEFAULT_RECURRENT_INITIALIZER), this.biasInitializer = J(t.biasInitializer || this.DEFAULT_BIAS_INITIALIZER), this.unitForgetBias = t.unitForgetBias, this.kernelRegularizer = X(t.kernelRegularizer), this.recurrentRegularizer = X(t.recurrentRegularizer), this.biasRegularizer = X(t.biasRegularizer), this.kernelConstraint = rt(t.kernelConstraint), this.recurrentConstraint = rt(t.recurrentConstraint), this.biasConstraint = rt(t.biasConstraint), this.dropout = ve([1, qt([0, t.dropout == null ? 0 : t.dropout])]), this.recurrentDropout = ve([
       1,
       qt([0, t.recurrentDropout == null ? 0 : t.recurrentDropout])
     ]), this.dropoutFunc = t.dropoutFunc, this.implementation = t.implementation, this.stateSize = [this.units, this.units], this.dropoutMask = null, this.recurrentDropoutMask = null;
@@ -10869,8 +10869,8 @@ class Dn extends In {
       0 < this.recurrentDropout && this.recurrentDropout < 1 && (i = w(i, o[0])), p = $(p, St(i, this.recurrentKernel.read())), this.useBias && (p = zt(p, this.bias.read()));
       const [f, g, b, m] = Kt(p, 4, p.rank - 1);
       l = this.recurrentActivation.apply(f), u = this.recurrentActivation.apply(g), c = $(w(u, r), w(l, this.activation.apply(b))), h = this.recurrentActivation.apply(m);
-      const v = w(h, this.activation.apply(c));
-      return [v, v, c];
+      const N = w(h, this.activation.apply(c));
+      return [N, N, c];
     });
   }
   getConfig() {
@@ -11153,12 +11153,12 @@ class Ds extends Dn {
         dropoutFunc: this.dropoutFunc
       }));
       const g = this.recurrentDropoutMask;
-      let b = u(r, g, 0), m = u(r, g, 1), v = u(r, g, 2), y = u(r, g, 3);
-      const C = 3, [N, I, z, _] = Kt(this.kernel.read(), o, C), [T, E, R, q] = this.useBias ? Kt(this.bias.read(), o) : [null, null, null, null];
-      c = this.inputConv(c, N, T, this.padding), h = this.inputConv(h, I, E, this.padding), p = this.inputConv(p, z, R, this.padding), f = this.inputConv(f, _, q, this.padding);
+      let b = u(r, g, 0), m = u(r, g, 1), N = u(r, g, 2), y = u(r, g, 3);
+      const C = 3, [v, I, z, _] = Kt(this.kernel.read(), o, C), [T, E, R, q] = this.useBias ? Kt(this.bias.read(), o) : [null, null, null, null];
+      c = this.inputConv(c, v, T, this.padding), h = this.inputConv(h, I, E, this.padding), p = this.inputConv(p, z, R, this.padding), f = this.inputConv(f, _, q, this.padding);
       const [bt, ie, re, xt] = Kt(this.recurrentKernel.read(), o, C);
-      b = this.recurrentConv(b, bt), m = this.recurrentConv(m, ie), v = this.recurrentConv(v, re), y = this.recurrentConv(y, xt);
-      const Tt = this.recurrentActivation.apply($(c, b)), me = this.recurrentActivation.apply($(h, m)), ze = $(w(me, a), w(Tt, this.activation.apply($(p, v)))), Ts = w(this.recurrentActivation.apply($(f, y)), this.activation.apply(ze));
+      b = this.recurrentConv(b, bt), m = this.recurrentConv(m, ie), N = this.recurrentConv(N, re), y = this.recurrentConv(y, xt);
+      const Tt = this.recurrentActivation.apply($(c, b)), me = this.recurrentActivation.apply($(h, m)), ze = $(w(me, a), w(Tt, this.activation.apply($(p, N)))), Ts = w(this.recurrentActivation.apply($(f, y)), this.activation.apply(ze));
       return [Ts, Ts, ze];
     });
   }
@@ -11329,7 +11329,7 @@ class xa extends W {
 }
 xa.className = "Flatten";
 S(xa);
-class Na extends W {
+class va extends W {
   constructor(t) {
     super(t), this.supportsMasking = !0, this.activation = Qt(t.activation);
   }
@@ -11345,9 +11345,9 @@ class Na extends W {
     return Object.assign(t, e), t;
   }
 }
-Na.className = "Activation";
-S(Na);
-class va extends W {
+va.className = "Activation";
+S(va);
+class Na extends W {
   constructor(t) {
     super(t), this.n = t.n, this.inputSpec = [{ ndim: 2 }];
   }
@@ -11364,8 +11364,8 @@ class va extends W {
     return Object.assign(t, e), t;
   }
 }
-va.className = "RepeatVector";
-S(va);
+Na.className = "RepeatVector";
+S(Na);
 class Sa extends W {
   constructor(t) {
     super(t), this.targetShape = t.targetShape;
@@ -12110,16 +12110,16 @@ class _a extends W {
       c.sort();
       const h = !Ft(c, It(0, a).slice(0, a - 1)), p = () => {
         if (h) {
-          const y = A(this.movingMean.read(), u), C = A(this.movingVariance.read(), u), N = this.center ? A(this.beta.read(), u) : null, I = this.scale ? A(this.gamma.read(), u) : null;
-          return _e(i, y, C, N, I, this.epsilon);
+          const y = A(this.movingMean.read(), u), C = A(this.movingVariance.read(), u), v = this.center ? A(this.beta.read(), u) : null, I = this.scale ? A(this.gamma.read(), u) : null;
+          return _e(i, y, C, v, I, this.epsilon);
         } else
           return _e(i, this.movingMean.read(), this.movingVariance.read(), this.beta == null ? null : this.beta.read(), this.gamma == null ? null : this.gamma.read(), this.epsilon);
       };
       if (!n)
         return p();
-      const [f, g, b] = tm(i, this.gamma.read(), this.beta.read(), o, this.epsilon), m = (y, C, N) => {
+      const [f, g, b] = tm(i, this.gamma.read(), this.beta.read(), o, this.epsilon), m = (y, C, v) => {
         x(() => {
-          const I = 1 - N, z = y.read(), _ = w(V(z, C), I);
+          const I = 1 - v, z = y.read(), _ = w(V(z, C), I);
           y.write(V(z, _));
         });
       };
@@ -12496,7 +12496,7 @@ class Qa extends Xa {
   call(t, e) {
     return x(() => {
       const n = O(t);
-      return ve(n, 1);
+      return Ne(n, 1);
     });
   }
 }
@@ -12531,7 +12531,7 @@ class no extends to {
   call(t, e) {
     return x(() => {
       const n = O(t);
-      return this.dataFormat === "channelsLast" ? ve(n, [1, 2]) : ve(n, [2, 3]);
+      return this.dataFormat === "channelsLast" ? Ne(n, [1, 2]) : Ne(n, [2, 3]);
     });
   }
 }
@@ -12805,8 +12805,8 @@ class oo extends W {
       t.rank === 3 ? (c = !0, u = kn([t])) : u = t;
       for (let I = 0; I < u.shape[0]; I++)
         m.push(b);
-      const v = Ku(m, [m.length, 4]), y = Hu(0, m.length, 1, "int32"), N = rm(u, v, y, [i, r], "nearest");
-      return c ? Lt(O(en(N)), l) : Lt(N, l);
+      const N = Ku(m, [m.length, 4]), y = Hu(0, m.length, 1, "int32"), v = rm(u, N, y, [i, r], "nearest");
+      return c ? Lt(O(en(v)), l) : Lt(v, l);
     });
   }
   upsize(t, e, n, i) {
@@ -12896,7 +12896,7 @@ class lo extends W {
               Received countWeights=${e.countWeights}`);
         n = O(e.countWeights);
       }
-      const i = ve(t), r = qu(t), a = Gt(this.numTokens, i).bufferSync().get(0), o = Ue(r, 0).bufferSync().get(0);
+      const i = Ne(t), r = qu(t), a = Gt(this.numTokens, i).bufferSync().get(0), o = Ue(r, 0).bufferSync().get(0);
       if (!(a && o))
         throw new d(`Input values must be between 0 < values <= numTokens with numTokens=${this.numTokens}`);
       return am(t, this.outputMode, this.numTokens, n);