/
Swish.php
284 lines (245 loc) · 6.55 KB
/
Swish.php
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
<?php
namespace Rubix\ML\NeuralNet\Layers;
use Tensor\Matrix;
use Rubix\ML\Deferred;
use Rubix\ML\NeuralNet\Parameter;
use Rubix\ML\NeuralNet\Optimizers\Optimizer;
use Rubix\ML\NeuralNet\Initializers\Constant;
use Rubix\ML\NeuralNet\Initializers\Initializer;
use Rubix\ML\NeuralNet\ActivationFunctions\Sigmoid;
use Rubix\ML\Exceptions\RuntimeException;
use Generator;
/**
* Swish
*
* Swish is a parametric activation layer that utilizes smooth rectified activation functions. The trainable
* *beta* parameter allows each activation function in the layer to tailor its output to the training set by
* interpolating between the linear function and ReLU.
*
* [1] P. Ramachandran et al. (2017). Swish: A Self-gated Activation Function.
* [2] P. Ramachandran et al. (2017). Searching for Activation Functions.
*
* @category Machine Learning
* @package Rubix/ML
* @author Andrew DalPino
*/
class Swish implements Hidden, Parametric
{
/**
* The initializer of the beta parameter.
*
* @var Initializer
*/
protected Initializer $initializer;
/**
* The sigmoid activation function.
*
* @var Sigmoid
*/
protected Sigmoid $sigmoid;
/**
* The width of the layer.
*
* @var positive-int|null
*/
protected ?int $width = null;
/**
* The parameterized scaling factors.
*
* @var Parameter|null
*/
protected ?Parameter $beta = null;
/**
* The memoized input matrix.
*
* @var Matrix|null
*/
protected ?Matrix $input = null;
/**
* The memorized activation matrix.
*
* @var Matrix|null
*/
protected ?Matrix $output = null;
/**
* @param Initializer|null $initializer
*/
public function __construct(?Initializer $initializer = null)
{
$this->initializer = $initializer ?? new Constant(1.0);
$this->sigmoid = new Sigmoid();
}
/**
* Return the width of the layer.
*
* @internal
*
* @throws RuntimeException
* @return positive-int
*/
public function width() : int
{
if ($this->width === null) {
throw new RuntimeException('Layer has not been initialized.');
}
return $this->width;
}
/**
* Initialize the layer with the fan in from the previous layer and return
* the fan out for this layer.
*
* @internal
*
* @param positive-int $fanIn
* @return positive-int
*/
public function initialize(int $fanIn) : int
{
$fanOut = $fanIn;
$beta = $this->initializer->initialize(1, $fanOut)->columnAsVector(0);
$this->width = $fanOut;
$this->beta = new Parameter($beta);
return $fanOut;
}
/**
* Compute a forward pass through the layer.
*
* @internal
*
* @param Matrix $input
* @return Matrix
*/
public function forward(Matrix $input) : Matrix
{
$output = $this->activate($input);
$this->input = $input;
$this->output = $output;
return $output;
}
/**
* Compute an inferential pass through the layer.
*
* @internal
*
* @param Matrix $input
* @return Matrix
*/
public function infer(Matrix $input) : Matrix
{
return $this->activate($input);
}
/**
* Calculate the gradient and update the parameters of the layer.
*
* @internal
*
* @param Deferred $prevGradient
* @param Optimizer $optimizer
* @throws RuntimeException
* @return Deferred
*/
public function back(Deferred $prevGradient, Optimizer $optimizer) : Deferred
{
if (!$this->beta) {
throw new RuntimeException('Layer has not been initialized.');
}
if (!$this->input or !$this->output) {
throw new RuntimeException('Must perform forward pass'
. ' before backpropagating.');
}
$dOut = $prevGradient();
$dIn = $this->input;
$dBeta = $dOut->multiply($dIn)->sum();
$this->beta->update($dBeta, $optimizer);
$input = $this->input;
$output = $this->output;
$this->input = $this->output = null;
return new Deferred([$this, 'gradient'], [$input, $output, $dOut]);
}
/**
* Calculate the gradient for the previous layer.
*
* @internal
*
* @param Matrix $input
* @param Matrix $output
* @param Matrix $dOut
* @return Matrix
*/
public function gradient($input, $output, $dOut) : Matrix
{
return $this->differentiate($input, $output)
->multiply($dOut);
}
/**
* Return the parameters of the layer.
*
* @internal
*
* @throws \RuntimeException
* @return \Generator<\Rubix\ML\NeuralNet\Parameter>
*/
public function parameters() : Generator
{
if (!$this->beta) {
throw new RuntimeException('Layer has not been initialized.');
}
yield 'beta' => $this->beta;
}
/**
* Restore the parameters in the layer from an associative array.
*
* @internal
*
* @param \Rubix\ML\NeuralNet\Parameter[] $parameters
*/
public function restore(array $parameters) : void
{
$this->beta = $parameters['beta'];
}
/**
* Compute the Swish activation function and return a matrix.
*
* @param Matrix $input
* @throws RuntimeException
* @return Matrix
*/
protected function activate(Matrix $input) : Matrix
{
if (!$this->beta) {
throw new RuntimeException('Layer has not been initialized.');
}
$zHat = $input->multiply($this->beta->param());
return $this->sigmoid->activate($zHat)
->multiply($input);
}
/**
* Calculate the derivative of the activation function at a given output.
*
* @param Matrix $input
* @param Matrix $output
* @throws RuntimeException
* @return Matrix
*/
protected function differentiate(Matrix $input, Matrix $output) : Matrix
{
if (!$this->beta) {
throw new RuntimeException('Layer has not been initialized.');
}
$ones = Matrix::ones(...$output->shape());
return $output->divide($input)
->multiply($ones->subtract($output))
->add($output);
}
/**
* Return the string representation of the object.
*
* @internal
*
* @return string
*/
public function __toString() : string
{
return "Swish (initializer: {$this->initializer})";
}
}