-
Notifications
You must be signed in to change notification settings - Fork 1.9k
/
optimizer_constructors.ts
189 lines (181 loc) · 7.11 KB
/
optimizer_constructors.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
/**
* @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.
* =============================================================================
*/
import {AdadeltaOptimizer} from './adadelta_optimizer';
import {AdagradOptimizer} from './adagrad_optimizer';
import {AdamOptimizer} from './adam_optimizer';
import {AdamaxOptimizer} from './adamax_optimizer';
import {MomentumOptimizer} from './momentum_optimizer';
import {RMSPropOptimizer} from './rmsprop_optimizer';
import {SGDOptimizer} from './sgd_optimizer';
export class OptimizerConstructors {
/**
* Constructs a `tf.SGDOptimizer` that uses stochastic gradient descent.
*
* ```js
* // Fit a quadratic function by learning the coefficients a, b, c.
* const xs = tf.tensor1d([0, 1, 2, 3]);
* const ys = tf.tensor1d([1.1, 5.9, 16.8, 33.9]);
*
* const a = tf.scalar(Math.random()).variable();
* const b = tf.scalar(Math.random()).variable();
* const c = tf.scalar(Math.random()).variable();
*
* // y = a * x^2 + b * x + c.
* const f = x => a.mul(x.square()).add(b.mul(x)).add(c);
* const loss = (pred, label) => pred.sub(label).square().mean();
*
* const learningRate = 0.01;
* const optimizer = tf.train.sgd(learningRate);
*
* // Train the model.
* for (let i = 0; i < 10; i++) {
* optimizer.minimize(() => loss(f(xs), ys));
* }
*
* // Make predictions.
* console.log(
* `a: ${a.dataSync()}, b: ${b.dataSync()}, c: ${c.dataSync()}`);
* const preds = f(xs).dataSync();
* preds.forEach((pred, i) => {
* console.log(`x: ${i}, pred: ${pred}`);
* });
* ```
*
* @param learningRate The learning rate to use for the SGD algorithm.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static sgd(learningRate: number): SGDOptimizer {
return new SGDOptimizer(learningRate);
}
/**
* Constructs a `tf.MomentumOptimizer` that uses momentum gradient
* descent.
*
* See
* [http://proceedings.mlr.press/v28/sutskever13.pdf](
* http://proceedings.mlr.press/v28/sutskever13.pdf)
*
* @param learningRate The learning rate to use for the Momentum gradient
* descent algorithm.
* @param momentum The momentum to use for the momentum gradient descent
* algorithm.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static momentum(learningRate: number, momentum: number, useNesterov = false):
MomentumOptimizer {
return new MomentumOptimizer(learningRate, momentum, useNesterov);
}
/**
* Constructs a `tf.RMSPropOptimizer` that uses RMSProp gradient
* descent. This implementation uses plain momentum and is not centered
* version of RMSProp.
*
* See
* [http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf](
* http://www.cs.toronto.edu/~tijmen/csc321/slides/lecture_slides_lec6.pdf)
*
* @param learningRate The learning rate to use for the RMSProp gradient
* descent algorithm.
* @param decay The discounting factor for the history/coming gradient.
* @param momentum The momentum to use for the RMSProp gradient descent
* algorithm.
* @param epsilon Small value to avoid zero denominator.
* @param centered If true, gradients are normalized by the estimated
* variance of the gradient.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static rmsprop(
learningRate: number, decay = .9, momentum = 0.0, epsilon: number = null,
centered = false): RMSPropOptimizer {
return new RMSPropOptimizer(
learningRate, decay, momentum, epsilon, centered);
}
/**
* Constructs a `tf.AdamOptimizer` that uses the Adam algorithm.
* See [https://arxiv.org/abs/1412.6980](https://arxiv.org/abs/1412.6980)
*
* @param learningRate The learning rate to use for the Adam gradient
* descent algorithm.
* @param beta1 The exponential decay rate for the 1st moment estimates.
* @param beta2 The exponential decay rate for the 2nd moment estimates.
* @param epsilon A small constant for numerical stability.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static adam(
learningRate = 0.001, beta1 = 0.9, beta2 = 0.999,
epsilon: number = null): AdamOptimizer {
return new AdamOptimizer(learningRate, beta1, beta2, epsilon);
}
/**
* Constructs a `tf.AdadeltaOptimizer` that uses the Adadelta algorithm.
* See [https://arxiv.org/abs/1212.5701](https://arxiv.org/abs/1212.5701)
*
* @param learningRate The learning rate to use for the Adadelta gradient
* descent algorithm.
* @param rho The learning rate decay over each update.
* @param epsilon A constant epsilon used to better condition the grad
* update.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static adadelta(learningRate = .001, rho = .95, epsilon: number = null):
AdadeltaOptimizer {
return new AdadeltaOptimizer(learningRate, rho, epsilon);
}
/**
* Constructs a `tf.AdamaxOptimizer` that uses the Adamax algorithm.
* See [https://arxiv.org/abs/1412.6980](https://arxiv.org/abs/1412.6980)
*
* @param learningRate The learning rate to use for the Adamax gradient
* descent algorithm.
* @param beta1 The exponential decay rate for the 1st moment estimates.
* @param beta2 The exponential decay rate for the 2nd moment estimates.
* @param epsilon A small constant for numerical stability.
* @param decay The learning rate decay over each update.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static adamax(
learningRate = 0.002, beta1 = 0.9, beta2 = 0.999, epsilon: number = null,
decay = 0.0): AdamaxOptimizer {
return new AdamaxOptimizer(learningRate, beta1, beta2, epsilon, decay);
}
/**
* Constructs a `tf.AdagradOptimizer` that uses the Adagrad algorithm.
* See
* [http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf](
* http://www.jmlr.org/papers/volume12/duchi11a/duchi11a.pdf)
* or
* [http://ruder.io/optimizing-gradient-descent/index.html#adagrad](
* http://ruder.io/optimizing-gradient-descent/index.html#adagrad)
*
* @param learningRate The learning rate to use for the Adagrad gradient
* descent algorithm.
* @param initialAccumulatorValue Starting value for the accumulators, must be
* positive.
*
* @doc {heading: 'Training', subheading: 'Optimizers', namespace: 'train'}
*/
static adagrad(learningRate: number, initialAccumulatorValue = 0.1):
AdagradOptimizer {
return new AdagradOptimizer(learningRate, initialAccumulatorValue);
}
}