This repository has been archived by the owner on Aug 15, 2019.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
/
relu_ops.ts
177 lines (161 loc) · 5.22 KB
/
relu_ops.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
/**
* @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 {ENV} from '../environment';
import {Tensor} from '../tensor';
import {convertToTensor} from '../tensor_util_env';
import {TensorLike} from '../types';
import {maximum} from './binary_ops';
import {getReductionAxes} from './broadcast_util';
import {where} from './logical_ops';
import {op} from './operation';
import {SELU_SCALE, SELU_SCALEALPHA} from './selu_util';
import {scalar, zerosLike} from './tensor_ops';
/**
* Computes rectified linear element-wise: `max(x, 0)`.
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
*
* x.relu().print(); // or tf.relu(x)
* ```
* @param x The input tensor. If the dtype is `bool`, the output dtype will be
* `int32'.
*/
/** @doc {heading: 'Operations', subheading: 'Basic math'} */
function relu_<T extends Tensor>(x: T|TensorLike): T {
const $x = convertToTensor(x, 'x', 'relu');
if ($x.dtype === 'bool') {
return $x.toInt();
}
const grad = (dy: T) => {
const stepRes = $x.step();
return {$x: () => dy.mulStrict(stepRes.toFloat())};
};
return ENV.engine.runKernel(backend => backend.relu($x), {$x}, grad);
}
/**
* Computes exponential linear element-wise: `x > 0 ? e ^ x - 1 : 0`.
*
* ```js
* const x = tf.tensor1d([-1, 1, -3, 2]);
*
* x.elu().print(); // or tf.elu(x)
* ```
* @param x The input tensor.
*/
/** @doc {heading: 'Operations', subheading: 'Basic math'} */
function elu_<T extends Tensor>(x: T|TensorLike): T {
const $x = convertToTensor(x, 'x', 'elu');
const grad = (dy: T, saved: Tensor[]) => {
const [y] = saved;
return {
$x: () =>
ENV.engine.runKernel(backend => backend.eluDer(dy, y), {dy, y}) as T
};
};
return ENV.engine.runKernel(
(backend, save) => save(backend.elu($x)), {$x}, grad);
}
/**
* Computes scaled exponential linear element-wise.
*
* `x < 0 ? scale * alpha * (exp(x) - 1) : x`
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
*
* x.selu().print(); // or tf.selu(x)
* ```
* @param x The input tensor.
*/
/** @doc {heading: 'Operations', subheading: 'Basic math'} */
function selu_<T extends Tensor>(x: T|TensorLike): T {
const $x = convertToTensor(x, 'x', 'selu');
const grad = (dy: T) => {
return {
$x: () => {
const mask = $x.greater(scalar(0));
const scaleAlpha = scalar(SELU_SCALEALPHA);
const scale = scalar(SELU_SCALE);
const greaterThanZeroDer = dy.mul(scale);
const lessEqualZeroDer = dy.mul(scaleAlpha).mul($x.toFloat().exp());
return where(mask, greaterThanZeroDer, lessEqualZeroDer) as T;
}
};
};
return ENV.engine.runKernel(backend => backend.selu($x), {$x}, grad);
}
/**
* Computes leaky rectified linear element-wise.
*
* See
* [http://web.stanford.edu/~awni/papers/relu_hybrid_icml2013_final.pdf](
* http://web.stanford.edu/~awni/papers/relu_hybrid_icml2013_final.pdf)
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
*
* x.leakyRelu(0.1).print(); // or tf.leakyRelu(x, 0.1)
* ```
* @param x The input tensor.
* @param alpha The scaling factor for negative values, defaults to 0.2.
*/
/** @doc {heading: 'Operations', subheading: 'Basic math'} */
function leakyRelu_<T extends Tensor>(x: T|TensorLike, alpha = 0.2): T {
const $x = convertToTensor(x, 'x', 'leakyRelu');
return maximum(scalar(alpha).mul($x), $x);
}
/**
* Computes leaky rectified linear element-wise with parametric alphas.
*
* `x < 0 ? alpha * x : f(x) = x`
*
* ```js
* const x = tf.tensor1d([-1, 2, -3, 4]);
* const alpha = tf.scalar(0.1);
*
* x.prelu(alpha).print(); // or tf.prelu(x, alpha)
* ```
* @param x The input tensor.
* @param alpha Scaling factor for negative values.
*/
/** @doc {heading: 'Operations', subheading: 'Basic math'} */
function prelu_<T extends Tensor>(x: T|TensorLike, alpha: T|TensorLike): T {
const $x = convertToTensor(x, 'x', 'prelu');
const $alpha = convertToTensor(alpha, 'alpha', 'prelu');
const grad = (dy: Tensor) => {
const mask = $x.greater(0);
return {
$x: () => where(mask, dy, dy.mul($alpha)) as T,
$alpha: () => {
let res = where(mask, zerosLike(dy), dy.mul($x));
const reduceAxes = getReductionAxes($alpha.shape, dy.shape);
if (reduceAxes.length > 0) {
res = res.sum(reduceAxes);
}
return res.reshape($alpha.shape) as T;
}
};
};
return ENV.engine.runKernel(
backend => backend.prelu($x, $alpha), {$x, $alpha}, grad) as T;
}
export const elu = op({elu_});
export const leakyRelu = op({leakyRelu_});
export const prelu = op({prelu_});
export const relu = op({relu_});
export const selu = op({selu_});