@genai-fi/nanogpt 0.2.5 → 0.2.7

package/dist/layers/TiedEmbedding.js

@@ -1,4 +1,6 @@
-import { o as p, c as a, b as V, E as u, C as X, R as Y, d as A, B as Z, S as ee, f as te, g as se, h as ne, I as re, L as ae, N as ue, P as ie, i as oe, j as le, k as ce, l as pe, n as B, t as he, T as L, m as fe, p as me, q as de, r as q, u as P, v as U, _ as H } from "../index-DcaSvB38.js";
+import { o as h, c as u, x as B, E as c, B as V, y as X, D as Y, I as Z, F as ee, N as te, H as se, J as ne, K as re, O as ae, Q as ue, f as L, w as ie, T as A, m as oe, U as le, t as ce, k as C, V as P, v as U, _ as H } from "../index-D1SlunD-.js";
+import { s as pe, r as f } from "../sum-02UQ5Eaq.js";
+import { c as he } from "../complex-D6Bq1XDf.js";
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -15,57 +17,13 @@ import { o as p, c as a, b as V, E as u, C as X, R as Y, d as A, B as Z, S as ee
  * limitations under the License.
  * =============================================================================
  */
-function ge(t, e) {
-  const s = a(t, "real", "complex"), n = a(e, "imag", "complex");
-  V(s.shape, n.shape, `real and imag shapes, ${s.shape} and ${n.shape}, must match in call to tf.complex().`);
-  const r = { real: s, imag: n };
-  return u.runKernel(X, r);
+function fe(t, e, s = !1, n = !1) {
+  let r = u(t, "a", "matMul"), i = u(e, "b", "matMul");
+  [r, i] = B(r, i);
+  const o = { a: r, b: i }, p = { transposeA: s, transposeB: n };
+  return c.runKernel(V, o, p);
 }
-const $e = /* @__PURE__ */ p({ complex_: ge });
-/**
- * @license
- * Copyright 2020 Google LLC. All Rights Reserved.
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- * =============================================================================
- */
-function xe(t, e) {
-  const n = { x: a(t, "x", "reshape", "string_or_numeric") }, r = { shape: e };
-  return u.runKernel(Y, n, r);
-}
-const f = /* @__PURE__ */ p({ reshape_: xe });
-/**
- * @license
- * Copyright 2020 Google LLC. All Rights Reserved.
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- * =============================================================================
- */
-function ke(t, e, s = !1, n = !1) {
-  let r = a(t, "a", "matMul"), i = a(e, "b", "matMul");
-  [r, i] = A(r, i);
-  const l = { a: r, b: i }, h = { transposeA: s, transposeB: n };
-  return u.runKernel(Z, l, h);
-}
-const d = /* @__PURE__ */ p({ matMul_: ke });
+const m = /* @__PURE__ */ h({ matMul_: fe });
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -82,11 +40,11 @@ const d = /* @__PURE__ */ p({ matMul_: ke });
  * limitations under the License.
  * =============================================================================
  */
-function be(t) {
-  const s = { x: a(t, "x", "sigmoid", "float32") };
-  return u.runKernel(ee, s);
+function de(t) {
+  const s = { x: u(t, "x", "sigmoid", "float32") };
+  return c.runKernel(X, s);
 }
-const De = /* @__PURE__ */ p({ sigmoid_: be });
+const me = /* @__PURE__ */ h({ sigmoid_: de });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -103,34 +61,11 @@ const De = /* @__PURE__ */ p({ sigmoid_: be });
  * limitations under the License.
  * =============================================================================
  */
-function Se(t) {
-  const s = { x: a(t, "x", "elu", "float32") };
-  return u.runKernel(te, s);
-}
-const ye = /* @__PURE__ */ p({ elu_: Se });
-/**
- * @license
- * Copyright 2018 Google LLC. All Rights Reserved.
- * Licensed under the Apache License, Version 2.0 (the "License");
- * you may not use this file except in compliance with the License.
- * You may obtain a copy of the License at
- *
- * http://www.apache.org/licenses/LICENSE-2.0
- *
- * Unless required by applicable law or agreed to in writing, software
- * distributed under the License is distributed on an "AS IS" BASIS,
- * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
- * See the License for the specific language governing permissions and
- * limitations under the License.
- * =============================================================================
- */
-function _e(t, e = null, s = !1) {
-  let n = a(t, "x", "sum");
-  n.dtype === "bool" && (n = se(n, "int32"));
-  const r = { x: n }, i = { axis: e, keepDims: s };
-  return u.runKernel(ne, r, i);
+function ge(t) {
+  const s = { x: u(t, "x", "elu", "float32") };
+  return c.runKernel(Y, s);
 }
-const Me = /* @__PURE__ */ p({ sum_: _e });
+const $e = /* @__PURE__ */ h({ elu_: ge });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -147,11 +82,11 @@ const Me = /* @__PURE__ */ p({ sum_: _e });
  * limitations under the License.
  * =============================================================================
  */
-function Ke(t) {
-  const s = { input: a(t, "input", "imag") };
-  return u.runKernel(re, s);
+function xe(t) {
+  const s = { input: u(t, "input", "imag") };
+  return c.runKernel(Z, s);
 }
-const we = /* @__PURE__ */ p({ imag_: Ke });
+const ke = /* @__PURE__ */ h({ imag_: xe });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -168,11 +103,11 @@ const we = /* @__PURE__ */ p({ imag_: Ke });
  * limitations under the License.
  * =============================================================================
  */
-function We(t, e = 0.2) {
-  const n = { x: a(t, "x", "leakyRelu") }, r = { alpha: e };
-  return u.runKernel(ae, n, r);
+function De(t, e = 0.2) {
+  const n = { x: u(t, "x", "leakyRelu") }, r = { alpha: e };
+  return c.runKernel(ee, n, r);
 }
-const ze = /* @__PURE__ */ p({ leakyRelu_: We });
+const be = /* @__PURE__ */ h({ leakyRelu_: De });
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -189,11 +124,11 @@ const ze = /* @__PURE__ */ p({ leakyRelu_: We });
  * limitations under the License.
  * =============================================================================
  */
-function Ee(t) {
-  const s = { x: a(t, "x", "neg") };
-  return u.runKernel(ue, s);
+function ye(t) {
+  const s = { x: u(t, "x", "neg") };
+  return c.runKernel(te, s);
 }
-const Oe = /* @__PURE__ */ p({ neg_: Ee });
+const Se = /* @__PURE__ */ h({ neg_: ye });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -210,11 +145,11 @@ const Oe = /* @__PURE__ */ p({ neg_: Ee });
  * limitations under the License.
  * =============================================================================
  */
-function Re(t, e) {
-  const s = a(t, "x", "prelu"), n = a(e, "alpha", "prelu"), r = { x: s, alpha: n };
-  return u.runKernel(ie, r);
+function Me(t, e) {
+  const s = u(t, "x", "prelu"), n = u(e, "alpha", "prelu"), r = { x: s, alpha: n };
+  return c.runKernel(se, r);
 }
-const Fe = /* @__PURE__ */ p({ prelu_: Re });
+const Ke = /* @__PURE__ */ h({ prelu_: Me });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -231,11 +166,11 @@ const Fe = /* @__PURE__ */ p({ prelu_: Re });
  * limitations under the License.
  * =============================================================================
  */
-function Le(t) {
-  const s = { input: a(t, "input", "real") };
-  return u.runKernel(oe, s);
+function _e(t) {
+  const s = { input: u(t, "input", "real") };
+  return c.runKernel(ne, s);
 }
-const Ae = /* @__PURE__ */ p({ real_: Le });
+const we = /* @__PURE__ */ h({ real_: _e });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -252,11 +187,11 @@ const Ae = /* @__PURE__ */ p({ real_: Le });
  * limitations under the License.
  * =============================================================================
  */
-function Be(t) {
-  const s = { x: a(t, "x", "relu") };
-  return u.runKernel(le, s);
+function We(t) {
+  const s = { x: u(t, "x", "relu") };
+  return c.runKernel(re, s);
 }
-const Te = /* @__PURE__ */ p({ relu_: Be });
+const ze = /* @__PURE__ */ h({ relu_: We });
 /**
  * @license
  * Copyright 2020 Google LLC. All Rights Reserved.
@@ -273,11 +208,11 @@ const Te = /* @__PURE__ */ p({ relu_: Be });
  * limitations under the License.
  * =============================================================================
  */
-function Ne(t) {
-  const s = { x: a(t, "x", "relu6") };
-  return u.runKernel(ce, s);
+function Ee(t) {
+  const s = { x: u(t, "x", "relu6") };
+  return c.runKernel(ae, s);
 }
-const ve = /* @__PURE__ */ p({ relu6_: Ne });
+const Oe = /* @__PURE__ */ h({ relu6_: Ee });
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -294,11 +229,11 @@ const ve = /* @__PURE__ */ p({ relu6_: Ne });
  * limitations under the License.
  * =============================================================================
  */
-function Ce(t, e = 0) {
-  const n = { x: a(t, "x", "step") }, r = { alpha: e };
-  return u.runKernel(pe, n, r);
+function Fe(t, e = 0) {
+  const n = { x: u(t, "x", "step") }, r = { alpha: e };
+  return c.runKernel(ue, n, r);
 }
-const Ge = /* @__PURE__ */ p({ step_: Ce });
+const Re = /* @__PURE__ */ h({ step_: Fe });
 /**
  * @license
  * Copyright 2018 Google LLC. All Rights Reserved.
@@ -315,19 +250,19 @@ const Ge = /* @__PURE__ */ p({ step_: Ce });
  * limitations under the License.
  * =============================================================================
  */
-function Ie(t, e, s) {
-  const n = a(t, "x", "transpose");
-  if (e == null && (e = n.shape.map((l, h) => h).reverse()), B(n.rank === e.length, () => `Error in transpose: rank of input ${n.rank} must match length of perm ${e}.`), e.forEach((l) => {
-    B(l >= 0 && l < n.rank, () => `All entries in 'perm' must be between 0 and ${n.rank - 1} but got ${e}`);
+function Ae(t, e, s) {
+  const n = u(t, "x", "transpose");
+  if (e == null && (e = n.shape.map((o, p) => p).reverse()), L(n.rank === e.length, () => `Error in transpose: rank of input ${n.rank} must match length of perm ${e}.`), e.forEach((o) => {
+    L(o >= 0 && o < n.rank, () => `All entries in 'perm' must be between 0 and ${n.rank - 1} but got ${e}`);
   }), n.rank <= 1)
     return n.clone();
   const r = { x: n }, i = { perm: e };
-  return n.dtype === "complex64" ? he(() => {
-    let l = Ae(n), h = we(n);
-    return l = u.runKernel(L, { x: l }, i), h = u.runKernel(L, { x: h }, i), s && (h = Oe(h)), $e(l, h);
-  }) : u.runKernel(L, r, i);
+  return n.dtype === "complex64" ? ie(() => {
+    let o = we(n), p = ke(n);
+    return o = c.runKernel(A, { x: o }, i), p = c.runKernel(A, { x: p }, i), s && (p = Se(p)), he(o, p);
+  }) : c.runKernel(A, r, i);
 }
-const je = /* @__PURE__ */ p({ transpose_: Ie });
+const Be = /* @__PURE__ */ h({ transpose_: Ae });
 /**
  * @license
  * Copyright 2019 Google LLC. All Rights Reserved.
@@ -344,36 +279,36 @@ const je = /* @__PURE__ */ p({ transpose_: Ie });
  * limitations under the License.
  * =============================================================================
  */
-function qe(t, e, s) {
+function Le(t, e, s) {
   if (s == null || s === "linear")
     return t;
   if (s === "relu")
-    return fe(t, Ge(e));
+    return oe(t, Re(e));
   throw new Error(`Cannot compute gradient for fused activation ${s}.`);
 }
-function Pe(t, e) {
+function Te(t, e) {
   let s = e;
-  const n = me(t.shape, e.shape);
-  return n.length > 0 && (s = Me(s, n)), f(s, t.shape);
+  const n = le(t.shape, e.shape);
+  return n.length > 0 && (s = pe(s, n)), f(s, t.shape);
 }
-function Ue(t, e, s, n) {
+function Ne(t, e, s, n) {
   if (e === "linear")
     return t;
   if (e === "relu")
-    return Te(t);
+    return ze(t);
   if (e === "elu")
-    return ye(t);
+    return $e(t);
   if (e === "relu6")
-    return ve(t);
+    return Oe(t);
   if (e === "prelu")
-    return Fe(t, s);
+    return Ke(t, s);
   if (e === "leakyrelu")
-    return ze(t, n);
+    return be(t, n);
   if (e === "sigmoid")
-    return De(t);
+    return me(t);
   throw new Error(`Unknown fused activation ${e}.`);
 }
-const He = (t, e) => !(t > 0) || e === "linear";
+const ve = (t, e) => !(t > 0) || e === "linear";
 /**
  * @license
  * Copyright 2019 Google LLC. All Rights Reserved.
@@ -390,49 +325,49 @@ const He = (t, e) => !(t > 0) || e === "linear";
  * limitations under the License.
  * =============================================================================
  */
-function Je({ a: t, b: e, transposeA: s = !1, transposeB: n = !1, bias: r, activation: i = "linear", preluActivationWeights: l, leakyreluAlpha: h = 0.2 }) {
-  if (He(u.state.gradientDepth, i) === !1) {
-    let x = d(t, e, s, n);
-    return r != null && (x = de(x, r)), Ue(x, i, l, h);
+function Ge({ a: t, b: e, transposeA: s = !1, transposeB: n = !1, bias: r, activation: i = "linear", preluActivationWeights: o, leakyreluAlpha: p = 0.2 }) {
+  if (ve(c.state.gradientDepth, i) === !1) {
+    let x = m(t, e, s, n);
+    return r != null && (x = ce(x, r)), Ne(x, i, o, p);
   }
-  let o = a(t, "a", "fused matMul"), c = a(e, "b", "fused matMul");
-  [o, c] = A(o, c);
-  const b = s ? o.shape[o.rank - 2] : o.shape[o.rank - 1], D = n ? c.shape[c.rank - 1] : c.shape[c.rank - 2], w = s ? o.shape[o.rank - 1] : o.shape[o.rank - 2], W = n ? c.shape[c.rank - 2] : c.shape[c.rank - 1], T = o.shape.slice(0, -2), S = c.shape.slice(0, -2), N = q(T), v = q(S);
-  B(b === D, () => `Error in fused matMul: inner shapes (${b}) and (${D}) of Tensors with shapes ${o.shape} and ${c.shape} and transposeA=${s} and transposeB=${n} must match.`);
-  const O = P(o.shape.slice(0, -2), c.shape.slice(0, -2)).concat([w, W]), R = s ? f(o, [N, b, w]) : f(o, [N, w, b]), F = n ? f(c, [v, W, D]) : f(c, [v, D, W]);
-  let y;
-  r != null && (y = a(r, "bias", "fused matMul"), [y] = A(y, o), P(O, y.shape));
-  let C;
-  l != null && (C = a(l, "prelu weights", "fused matMul"));
-  const G = (x, K) => {
-    const [g, $, k, z] = K, m = qe(f(x, k.shape), k, i);
-    let _, M;
-    if (!s && !n ? (_ = d(m, $, !1, !0), M = d(g, m, !0, !1)) : !s && n ? (_ = d(m, $, !1, !1), M = d(m, g, !0, !1)) : s && !n ? (_ = d($, m, !1, !0), M = d(g, m, !1, !1)) : (_ = d($, m, !0, !0), M = d(m, g, !0, !0)), r != null) {
-      const Q = Pe(z, m);
-      return [_, M, Q];
+  let a = u(t, "a", "fused matMul"), l = u(e, "b", "fused matMul");
+  [a, l] = B(a, l);
+  const D = s ? a.shape[a.rank - 2] : a.shape[a.rank - 1], b = n ? l.shape[l.rank - 1] : l.shape[l.rank - 2], w = s ? a.shape[a.rank - 1] : a.shape[a.rank - 2], W = n ? l.shape[l.rank - 2] : l.shape[l.rank - 1], T = a.shape.slice(0, -2), y = l.shape.slice(0, -2), N = C(T), v = C(y);
+  L(D === b, () => `Error in fused matMul: inner shapes (${D}) and (${b}) of Tensors with shapes ${a.shape} and ${l.shape} and transposeA=${s} and transposeB=${n} must match.`);
+  const O = P(a.shape.slice(0, -2), l.shape.slice(0, -2)).concat([w, W]), F = s ? f(a, [N, D, w]) : f(a, [N, w, D]), R = n ? f(l, [v, W, b]) : f(l, [v, b, W]);
+  let S;
+  r != null && (S = u(r, "bias", "fused matMul"), [S] = B(S, a), P(O, S.shape));
+  let G;
+  o != null && (G = u(o, "prelu weights", "fused matMul"));
+  const I = (x, _) => {
+    const [g, $, k, z] = _, d = Le(f(x, k.shape), k, i);
+    let M, K;
+    if (!s && !n ? (M = m(d, $, !1, !0), K = m(g, d, !0, !1)) : !s && n ? (M = m(d, $, !1, !1), K = m(d, g, !0, !1)) : s && !n ? (M = m($, d, !1, !0), K = m(g, d, !1, !1)) : (M = m($, d, !0, !0), K = m(d, g, !0, !0)), r != null) {
+      const Q = Te(z, d);
+      return [M, K, Q];
     } else
-      return [_, M];
-  }, I = {
-    a: R,
-    b: F,
-    bias: y,
-    preluActivationWeights: C
-  }, j = { transposeA: s, transposeB: n, activation: i, leakyreluAlpha: h };
-  return r == null ? U((K, g, $) => {
+      return [M, K];
+  }, j = {
+    a: F,
+    b: R,
+    bias: S,
+    preluActivationWeights: G
+  }, q = { transposeA: s, transposeB: n, activation: i, leakyreluAlpha: p };
+  return r == null ? U((_, g, $) => {
     const k = (
       // tslint:disable-next-line: no-unnecessary-type-assertion
-      u.runKernel(H, I, j)
+      c.runKernel(H, j, q)
     );
-    return $([K, g, k]), { value: f(k, O), gradFunc: G };
-  })(R, F) : U((K, g, $, k) => {
+    return $([_, g, k]), { value: f(k, O), gradFunc: I };
+  })(F, R) : U((_, g, $, k) => {
     const z = (
       // tslint:disable-next-line: no-unnecessary-type-assertion
-      u.runKernel(H, I, j)
+      c.runKernel(H, j, q)
     );
-    return k([K, g, z, $]), { value: f(z, O), gradFunc: G };
-  })(R, F, y);
+    return k([_, g, z, $]), { value: f(z, O), gradFunc: I };
+  })(F, R, S);
 }
-const J = /* @__PURE__ */ p({ fusedMatMul_: Je });
+const J = /* @__PURE__ */ h({ fusedMatMul_: Ge });
 /**
  * @license
  * Copyright 2018 Google LLC
@@ -456,7 +391,7 @@ class E extends Error {
  * https://opensource.org/licenses/MIT.
  * =============================================================================
  */
-function Qe(t, e, s, n) {
+function Ie(t, e, s, n) {
   if (t.rank < 2 || e.rank < 2)
     throw new E(`dot requires both inputs to be rank >= 2 but got x shape = ${t.shape} and y shape = ${e.shape}`);
   if (e.rank >= 3) {
@@ -476,9 +411,9 @@ function Qe(t, e, s, n) {
   {
     const r = t.shape.slice(), i = r.pop();
     t = f(t, [-1, i]);
-    const l = e.shape.slice(), h = l.pop(), o = l.pop(), c = [...l, h], b = Array.from({ length: e.rank }, (T, S) => S === 0 ? e.rank - 2 : S <= e.rank - 2 ? S - 1 : S);
-    e = f(je(e, b), [o, -1]);
-    const D = [...r, ...c];
+    const o = e.shape.slice(), p = o.pop(), a = o.pop(), l = [...o, p], D = Array.from({ length: e.rank }, (T, y) => y === 0 ? e.rank - 2 : y <= e.rank - 2 ? y - 1 : y);
+    e = f(Be(e, D), [a, -1]);
+    const b = [...r, ...l];
     return f(J({
       a: t,
       b: e,
@@ -486,10 +421,10 @@ function Qe(t, e, s, n) {
       transposeB: !1,
       bias: null,
       activation: s
-    }), D);
+    }), b);
   }
 }
-class Ye {
+class Ue {
   vocabSize;
   embedDim;
   tf;
@@ -512,7 +447,7 @@ class Ye {
     return this.tf.gather(this.tiedWeights, e, 0);
   }
   project(e) {
-    return Qe(e, this.tiedWeights.transpose());
+    return Ie(e, this.tiedWeights.transpose());
   }
   getWeights() {
     return [this.tiedWeights];
@@ -531,5 +466,5 @@ class Ye {
   }
 }
 export {
-  Ye as default
+  Ue as default
 };

package/dist/main.d.ts

@@ -6,3 +6,5 @@ export { default as loadTextData } from './data/textLoader';
 export type { ITrainerOptions } from './Trainer';
 export type { IGenerateOptions } from './Generator';
 export type { TrainingLogEntry } from './NanoGPTModel';
+export type { GPTConfig } from './config';
+export { estimateParameterCount, estimateMemoryUsage, estimateTrainingMemoryUsage, estimateResources, validateConfig, } from './utilities/parameters';

package/dist/main.js

@@ -1,12 +1,20 @@
-import { default as o } from "./NanoGPTModel.js";
-import { default as t } from "./TeachableLLM.js";
-import { default as l } from "./tokeniser/CharTokeniser.js";
-import { default as s } from "./utilities/waitForModel.js";
-import { default as m } from "./data/textLoader.js";
+import { default as r } from "./NanoGPTModel.js";
+import { default as s } from "./TeachableLLM.js";
+import { default as i } from "./tokeniser/CharTokeniser.js";
+import { default as d } from "./utilities/waitForModel.js";
+import { default as u } from "./data/textLoader.js";
+import { estimateMemoryUsage as n, estimateParameterCount as T, estimateResources as g, estimateTrainingMemoryUsage as M, validateConfig as C } from "./utilities/parameters.js";
+import "./ops/scatterSub.js";
+import "./ops/gatherSub.js";
 export {
-  l as CharTokeniser,
-  o as NanoGPT,
-  t as TeachableLLM,
-  m as loadTextData,
-  s as waitForModel
+  i as CharTokeniser,
+  r as NanoGPT,
+  s as TeachableLLM,
+  n as estimateMemoryUsage,
+  T as estimateParameterCount,
+  g as estimateResources,
+  M as estimateTrainingMemoryUsage,
+  u as loadTextData,
+  C as validateConfig,
+  d as waitForModel
 };

package/dist/ops/gatherSub.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs';
+export declare function gatherSub(values: Tensor, labels: Tensor, logits: Tensor): Tensor;

package/dist/ops/gatherSub.js

@@ -0,0 +1,66 @@
+import { engine as l } from "@tensorflow/tfjs";
+import { o as g, c as i, E as b, G as d, r as c, a as h } from "../index-D1SlunD-.js";
+import { r as p, s as f } from "../stack-DB2YLlAs.js";
+/**
+ * @license
+ * Copyright 2018 Google LLC. All Rights Reserved.
+ * Licensed under the Apache License, Version 2.0 (the "License");
+ * you may not use this file except in compliance with the License.
+ * You may obtain a copy of the License at
+ *
+ * http://www.apache.org/licenses/LICENSE-2.0
+ *
+ * Unless required by applicable law or agreed to in writing, software
+ * distributed under the License is distributed on an "AS IS" BASIS,
+ * WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied.
+ * See the License for the specific language governing permissions and
+ * limitations under the License.
+ * =============================================================================
+ */
+function m(e, t) {
+  const n = i(t, "indices", "gatherND", "int32"), r = { params: i(e, "x", "gatherND", "string_or_numeric"), indices: n };
+  return b.runKernel(d, r);
+}
+const N = /* @__PURE__ */ g({ gatherND_: m });
+class S {
+  variableNames = ["labels", "logits", "values"];
+  outputShape;
+  userCode;
+  constructor(t) {
+    this.outputShape = [t], this.userCode = `
+      void main() {
+        int coords = getOutputCoords();
+        int index = int(getLabelsAtOutCoords());
+        float val = getValuesAtOutCoords();
+        float logit = getLogits(coords, index);
+        setOutput(val - logit);
+      }
+    `;
+  }
+}
+function k(e) {
+  const { logits: t, labels: n, values: s } = e.inputs, r = e.backend, o = n.shape[0], a = new S(o);
+  return r.runWebGLProgram(a, [n, t, s], "float32");
+}
+const G = {
+  kernelName: "EfficientGatherSub",
+  backendName: "webgl",
+  kernelFunc: k
+};
+c(G);
+function v(e) {
+  const { values: t, labels: n, logits: s } = e.inputs, r = n.shape[0], o = p(0, r, 1, "int32"), a = f([o, n], 1), u = N(s, a);
+  return h(t, u);
+}
+const C = {
+  kernelName: "EfficientGatherSub",
+  backendName: "cpu",
+  kernelFunc: v
+};
+c(C);
+function K(e, t, n) {
+  return l().runKernel("EfficientGatherSub", { logits: n, labels: t, values: e }, {});
+}
+export {
+  K as gatherSub
+};

package/dist/ops/node/sparseCrossEntropy.d.ts

@@ -0,0 +1 @@
+export {};

package/dist/ops/node/sparseCrossEntropy.js

@@ -0,0 +1,11 @@
+import { r as o } from "../../index-D1SlunD-.js";
+function r(e) {
+  const { logits: t, labels: n } = e.inputs;
+  return e.backend.executeMultipleOutputs("SparseSoftmaxCrossEntropyWithLogits", [], [t, n], 2);
+}
+const s = {
+  kernelName: "NativeSparseSoftmaxCrossEntropy",
+  backendName: "tensorflow",
+  kernelFunc: r
+};
+o(s);

package/dist/ops/scatterSub.d.ts

@@ -0,0 +1,2 @@
+import { Tensor } from '@tensorflow/tfjs';
+export declare function scatterSub(probabilities: Tensor, labels: Tensor, scale: Tensor): Tensor;