This repository has been archived by the owner on Aug 15, 2019. It is now read-only.
-
Notifications
You must be signed in to change notification settings - Fork 953
/
gradients.ts
378 lines (365 loc) · 12.2 KB
/
gradients.ts
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
112
113
114
115
116
117
118
119
120
121
122
123
124
125
126
127
128
129
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
/**
* @license
* Copyright 2018 Google Inc. 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.
* =============================================================================
*/
import {CustomGradientFunc, ScopeFn} from './engine';
import {ENV} from './environment';
import {Scalar, Tensor, Variable} from './tensor';
import {NamedTensorMap, TensorContainer} from './tensor_types';
import * as util from './util';
/**
* Create a new gradient scope. Similar to scope, but forces all inner scopes
* to not clean up so that gradient operations can be used inside of this
* scope.
* @param nameOrScopeFn The name of the scope, or the function to execute.
* If a name is provided, the 2nd argument should be the function.
* If a name is provided, and debug mode is on, the timing and the memory
* usage of the function will be tracked and displayed on the console
* using the provided name.
* @param scopeFn The function to execute.
*/
function gradScope<T extends TensorContainer>(
nameOrScopeFn: string|ScopeFn<T>, scopeFn?: ScopeFn<T>): T {
return ENV.engine.tidy(nameOrScopeFn, scopeFn, true /* gradScope */);
}
/**
* Provided `f(x)`, returns another function `g(x, dy?)`, which gives the
* gradient of `f(x)` with respect to `x`.
*
* If `dy` is provided, the gradient of `f(x).mul(dy).sum()` with respect to
* `x` is computed instead. `f(x)` must take a single tensor `x` and return a
* single tensor `y`. If `f()` takes multiple inputs, use `tf.grads` instead.
*
* ```js
* // f(x) = x ^ 2
* const f = x => x.square();
* // f'(x) = 2x
* const g = tf.grad(f);
*
* const x = tf.tensor1d([2, 3]);
* g(x).print();
* ```
*
* ```js
* // f(x) = x ^ 3
* const f = x => x.pow(tf.scalar(3, 'int32'));
* // f'(x) = 3x ^ 2
* const g = tf.grad(f);
* // f''(x) = 6x
* const gg = tf.grad(g);
*
* const x = tf.tensor1d([2, 3]);
* gg(x).print();
* ```
*
* @param f The function f(x), to compute gradient for.
*/
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function grad<I extends Tensor, O extends Tensor>(f: (x: I) => O): (
x: I, dy?: O) => I {
util.assert(
util.isFunction(f), () => 'The f passed in grad(f) must be a function');
return (x: I, dy?: O): I => {
util.assert(
x instanceof Tensor,
() => 'The x passed in grad(f)(x) must be a tensor');
util.assert(
dy == null || dy instanceof Tensor,
() => 'The dy passed in grad(f)(x, dy) must be a tensor');
return ENV.engine.tidy(() => {
const {value, grads} = ENV.engine.gradients(() => f(x), [x], dy);
if (dy != null) {
util.assertShapesMatch(
value.shape, dy.shape,
'The shape of dy passed in grad(f)(x, dy) must match the shape ' +
'returned by f(x)');
}
checkGrads(grads);
return grads[0] as I;
});
};
}
/**
* Provided `f(x1, x2,...)`, returns another function `g([x1, x2,...], dy?)`,
* which gives an array of gradients of `f()` with respect to each input
* [`x1`,`x2`,...].
*
* If `dy` is passed when calling `g()`, the gradient of
* `f(x1,...).mul(dy).sum()` with respect to each input is computed instead.
* The provided `f` must take one or more tensors and return a single tensor
* `y`. If `f()` takes a single input, we recommend using `tf.grad` instead.
*
* ```js
* // f(a, b) = a * b
* const f = (a, b) => a.mul(b);
* // df / da = b, df / db = a
* const g = tf.grads(f);
*
* const a = tf.tensor1d([2, 3]);
* const b = tf.tensor1d([-2, -3]);
* const [da, db] = g([a, b]);
* console.log('da');
* da.print();
* console.log('db');
* db.print();
* ```
*
* @param f The function `f(x1, x2,...)` to compute gradients for.
*/
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function grads<O extends Tensor>(f: (...args: Tensor[]) => O): (
args: Tensor[], dy?: O) => Tensor[] {
util.assert(
util.isFunction(f), () => 'The f passed in grads(f) must be a function');
return (args: Tensor[], dy?: O): Tensor[] => {
util.assert(
Array.isArray(args) && args.every(arg => arg instanceof Tensor),
() => 'The args passed in grads(f)(args) must be an array of tensors');
util.assert(
dy == null || dy instanceof Tensor,
() => 'The dy passed in grads(f)(args, dy) must be a tensor');
return ENV.engine.tidy(() => {
const {value, grads} = ENV.engine.gradients(() => f(...args), args, dy);
if (dy != null) {
util.assertShapesMatch(
value.shape, dy.shape,
'The shape of dy passed in grads(f)([x1,...], dy) must ' +
'match the shape returned by f([x1,...])');
}
checkGrads(grads);
return grads;
});
};
}
/**
* Like `tf.grad`, but also returns the value of `f()`. Useful when `f()`
* returns a metric you want to show.
*
* The result is a rich object with the following properties:
* - grad: The gradient of `f(x)` w.r.t `x` (result of `tf.grad`).
* - value: The value returned by `f(x)`.
*
* ```js
* // f(x) = x ^ 2
* const f = x => x.square();
* // f'(x) = 2x
* const g = tf.valueAndGrad(f);
*
* const x = tf.tensor1d([2, 3]);
* const {value, grad} = g(x);
*
* console.log('value');
* value.print();
* console.log('grad');
* grad.print();
* ```
*/
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function valueAndGrad<I extends Tensor, O extends Tensor>(f: (x: I) => O): (
x: I, dy?: O) => {
value: O;
grad: I;
} {
util.assert(
util.isFunction(f),
() => 'The f passed in valueAndGrad(f) must be a function');
return (x: I, dy?: O) => {
util.assert(
x instanceof Tensor,
() => 'The x passed in valueAndGrad(f)(x) must be a tensor');
util.assert(
dy == null || dy instanceof Tensor,
() => 'The dy passed in valueAndGrad(f)(x, dy) must be a tensor');
const {grads, value} = ENV.engine.gradients(() => f(x), [x], dy);
checkGrads(grads);
return {grad: grads[0] as I, value: value as O};
};
}
/**
* Like `tf.grads`, but returns also the value of `f()`. Useful when `f()`
* returns a metric you want to show.
*
* The result is a rich object with the following properties:
* - grads: The gradients of `f()` w.r.t each input (result of `tf.grads`).
* - value: The value returned by `f(x)`.
*
* ```js
* // f(a, b) = a * b
* const f = (a, b) => a.mul(b);
* // df/da = b, df/db = a
* const g = tf.valueAndGrads(f);
*
* const a = tf.tensor1d([2, 3]);
* const b = tf.tensor1d([-2, -3]);
* const {value, grads} = g([a, b]);
*
* const [da, db] = grads;
*
* console.log('value');
* value.print();
*
* console.log('da');
* da.print();
* console.log('db');
* db.print();
* ```
*/
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function valueAndGrads<O extends Tensor>(f: (...args: Tensor[]) => O): (
args: Tensor[], dy?: O) => {
grads: Tensor[];
value: O;
} {
util.assert(
util.isFunction(f),
() => 'The f passed in valueAndGrads(f) must be a function');
return (args: Tensor[], dy?: O) => {
util.assert(
Array.isArray(args) && args.every(arg => arg instanceof Tensor),
() => 'The args passed in valueAndGrads(f)(args) must be array of ' +
'tensors');
util.assert(
dy == null || dy instanceof Tensor,
() => 'The dy passed in valueAndGrads(f)(args, dy) must be a tensor');
const res = ENV.engine.gradients(() => f(...args), args, dy);
if (dy != null) {
util.assertShapesMatch(
res.value.shape, dy.shape,
'The shape of dy passed in valueAndGrads(f)([x1,...], dy) must ' +
'match the shape returned by f([x1,...])');
}
checkGrads(res.grads);
return res;
};
}
/**
* Computes and returns the gradient of f(x) with respect to the list of
* trainable variables provided by `varList`. If no list is provided, it
* defaults to all trainable variables.
*
* ```js
* const a = tf.variable(tf.tensor1d([3, 4]));
* const b = tf.variable(tf.tensor1d([5, 6]));
* const x = tf.tensor1d([1, 2]);
*
* // f(a, b) = a * x ^ 2 + b * x
* const f = () => a.mul(x.square()).add(b.mul(x)).sum();
* // df/da = x ^ 2, df/db = x
* const {value, grads} = tf.variableGrads(f);
*
* Object.keys(grads).forEach(varName => grads[varName].print());
* ```
*
* @param f The function to execute. f() should return a scalar.
* @param varList The list of trainable variables. Defaults to all variables.
*/
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function variableGrads(f: () => Scalar, varList?: Variable[]):
{value: Scalar, grads: NamedTensorMap} {
util.assert(
util.isFunction(f),
() => 'The f passed in variableGrads(f) must be a function');
util.assert(
varList == null ||
Array.isArray(varList) && varList.every(v => v instanceof Variable),
() =>
'The varList passed in variableGrads(f, varList) must be an array ' +
'of variables');
if (varList == null) {
// Get all of the trainable variables.
varList = [];
for (const varName in ENV.engine.registeredVariables) {
varList.push(ENV.engine.registeredVariables[varName]);
}
}
// Prune non-trainable variables.
const originalVarCount = varList.length;
varList = varList.filter(variable => variable.trainable);
util.assert(
varList.length > 0,
() =>
`variableGrads() expects at least one of the input variables to be ` +
`trainable, but none of the ${originalVarCount} variables is ` +
`trainable.`);
const allowNoGradients = true;
const {value, grads} =
ENV.engine.gradients(f, varList, null, allowNoGradients);
util.assert(
grads.some(g => g != null),
() => 'Cannot find a connection between any variable and the result of ' +
'the loss function y=f(x). Please make sure the operations that ' +
'use variables are inside the function f passed to minimize().');
util.assert(
value.rank === 0,
() => `The f passed in variableGrads(f) must return a scalar, but it ` +
`returned a rank-${value.rank} tensor`);
const namedGrads: NamedTensorMap = {};
varList.forEach((v, i) => {
if (grads[i] != null) {
namedGrads[v.name] = grads[i];
}
});
return {value, grads: namedGrads};
}
/**
* Overrides the gradient computation of a function `f`.
*
* Takes a function
* `f(...inputs) => {value: Tensor, gradFunc: dy => Tensor[]}` and returns
* another function `g(...inputs)` which takes the same inputs as `f`. When
* called, `g` returns `f().value`. In backward mode, custom gradients with
* respect to each input of `f` are computed using `f().gradFunc`.
*
* ```js
* const customOp = tf.customGrad(x => {
* // Override gradient of our custom x ^ 2 op to be dy * abs(x);
* return {value: x.square(), gradFunc: dy => [dy.mul(x.abs())]};
* });
*
* const x = tf.tensor1d([-1, -2, 3]);
* const dx = tf.grad(x => customOp(x));
*
* console.log(`f(x):`);
* customOp(x).print();
* console.log(`f'(x):`);
* dx(x).print();
* ```
*
* @param f The function to evaluate in forward mode, which should return
* `{value: Tensor, gradFunc: (dy) => Tensor[]}`, where `gradFunc` returns
* the custom gradients of `f` with respect to its inputs.
*/
/** @doc {heading: 'Training', subheading: 'Gradients'} */
function customGrad<T extends Tensor>(f: CustomGradientFunc<T>):
(...args: Tensor[]) => T {
return ENV.engine.customGrad(f);
}
function checkGrads(grads: Tensor[]) {
const numNullGradients = grads.filter(g => g == null).length;
if (numNullGradients > 0) {
throw new Error(
`Cannot compute gradient of y=f(x) with respect to x. Make sure that
the f you passed encloses all operations that lead from x to y.`);
}
}
export {
gradScope,
customGrad,
variableGrads,
valueAndGrad,
valueAndGrads,
grad,
grads,
};