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
|
/* Copyright 2015 The TensorFlow Authors. 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.
==============================================================================*/
#if GOOGLE_CUDA
#define EIGEN_USE_GPU
#include "tensorflow/core/framework/register_types.h"
#include "tensorflow/core/kernels/training_ops.h"
namespace tensorflow {
typedef Eigen::GpuDevice GPUDevice;
namespace functor {
template <typename T>
struct ApplyGradientDescent<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstFlat grad) {
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
var.device(d) -= lr.reshape(single).broadcast(bcast) * grad;
}
};
template <typename T>
struct ApplyAdagrad<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::Flat accum,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstFlat grad) {
accum.device(d) += grad.square();
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
var.device(d) -= lr.reshape(single).broadcast(bcast) * grad * accum.rsqrt();
}
};
template <typename T>
struct ApplyAdadelta<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::Flat accum,
typename TTypes<T>::Flat accum_update,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstScalar rho,
typename TTypes<T>::ConstScalar epsilon,
typename TTypes<T>::ConstFlat grad) {
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
accum.device(d) = accum * rho.reshape(single).broadcast(bcast) +
grad.square() * (grad.constant(T(1)) -
rho.reshape(single).broadcast(bcast));
const auto update =
(accum_update + epsilon.reshape(single).broadcast(bcast)).sqrt() *
(accum + epsilon.reshape(single).broadcast(bcast)).rsqrt() * grad;
var.device(d) -= update * lr.reshape(single).broadcast(bcast);
accum_update.device(d) =
accum_update * rho.reshape(single).broadcast(bcast) +
update.square() *
(grad.constant(T(1)) - rho.reshape(single).broadcast(bcast));
}
};
template <typename T>
struct ApplyMomentum<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::Flat accum,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstFlat grad,
typename TTypes<T>::ConstScalar momentum, bool use_nesterov) {
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
accum.device(d) = accum * momentum.reshape(single).broadcast(bcast) + grad;
if (use_nesterov) {
var.device(d) -= grad * lr.reshape(single).broadcast(bcast) +
accum * momentum.reshape(single).broadcast(bcast) *
lr.reshape(single).broadcast(bcast);
} else {
var.device(d) -= lr.reshape(single).broadcast(bcast) * accum;
}
}
};
template <typename T>
struct ApplyAdam<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::Flat m, typename TTypes<T>::Flat v,
typename TTypes<T>::ConstScalar beta1_power,
typename TTypes<T>::ConstScalar beta2_power,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstScalar beta1,
typename TTypes<T>::ConstScalar beta2,
typename TTypes<T>::ConstScalar epsilon,
typename TTypes<T>::ConstFlat grad, bool use_nesterov) {
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
const auto one = static_cast<T>(1.0);
m.device(d) =
m + (beta1.constant(one) - beta1).reshape(single).broadcast(bcast) *
(grad - m);
v.device(d) =
v + (beta2.constant(one) - beta2).reshape(single).broadcast(bcast) *
(grad.square() - v);
if (use_nesterov) {
var.device(d) -=
(lr * (beta2_power.constant(one) - beta2_power).sqrt() /
(beta1_power.constant(one) - beta1_power))
.reshape(single)
.broadcast(bcast) *
(m * beta1.reshape(single).broadcast(bcast) +
(beta1.constant(one) - beta1).reshape(single).broadcast(bcast) *
grad) /
(epsilon.reshape(single).broadcast(bcast) + v.sqrt());
} else {
var.device(d) -= (lr * (beta2_power.constant(one) - beta2_power).sqrt() /
(beta1_power.constant(one) - beta1_power))
.reshape(single)
.broadcast(bcast) *
m /
(epsilon.reshape(single).broadcast(bcast) + v.sqrt());
}
}
};
template <typename T>
struct ApplyRMSProp<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::Flat ms, typename TTypes<T>::Flat mom,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstScalar rho,
typename TTypes<T>::ConstScalar momentum,
typename TTypes<T>::ConstScalar epsilon,
typename TTypes<T>::ConstFlat grad) {
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
const auto one = static_cast<T>(1.0);
ms.device(d) =
ms + (rho.constant(one) - rho).reshape(single).broadcast(bcast) *
(grad.square() - ms);
mom.device(d) =
mom * momentum.reshape(single).broadcast(bcast) +
lr.reshape(single).broadcast(bcast) * grad /
((epsilon.reshape(single).broadcast(bcast) + ms).sqrt());
var.device(d) -= mom;
}
};
template <typename T>
struct ApplyCenteredRMSProp<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::Flat mg, typename TTypes<T>::Flat ms,
typename TTypes<T>::Flat mom,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstScalar rho,
typename TTypes<T>::ConstScalar momentum,
typename TTypes<T>::ConstScalar epsilon,
typename TTypes<T>::ConstFlat grad) {
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
const auto one = static_cast<T>(1.0);
const auto one_minus_rho =
(rho.constant(one) - rho).reshape(single).broadcast(bcast);
ms.device(d) = ms + one_minus_rho * (grad.square() - ms);
mg.device(d) = mg + one_minus_rho * (grad - mg);
auto denom = (ms - mg.square()) + epsilon.reshape(single).broadcast(bcast);
mom.device(d) = mom * momentum.reshape(single).broadcast(bcast) +
lr.reshape(single).broadcast(bcast) * grad / denom.sqrt();
var.device(d) -= mom;
}
};
template <typename T>
struct ApplyAddSign<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::Flat m,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstScalar alpha,
typename TTypes<T>::ConstScalar sign_decay,
typename TTypes<T>::ConstScalar beta,
typename TTypes<T>::ConstFlat grad) {
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
// The following is the GPU equivalent of the CPU version:
// m.device(d) = m * beta() + grad * (static_cast<T>(1) - beta());
const auto one = static_cast<T>(1.0);
auto beta_bcast = beta.reshape(single).broadcast(bcast);
auto one_minus_beta =
(beta.constant(one) - beta).reshape(single).broadcast(bcast);
m.device(d) = m * beta_bcast + grad * one_minus_beta;
// The following is the GPU equivalent of the CPU version:
// var.device(d) -= lr() * (alpha() + sign_decay() * sign_gm) * grad;
auto sign_gm = grad.sign() * m.sign();
auto lr_bcast = lr.reshape(single).broadcast(bcast);
auto alpha_bcast = alpha.reshape(single).broadcast(bcast);
auto sign_decay_bcast = sign_decay.reshape(single).broadcast(bcast);
var.device(d) -=
lr_bcast * (alpha_bcast + sign_decay_bcast * sign_gm) * grad;
}
};
template <typename T>
struct ApplyPowerSign<GPUDevice, T> {
void operator()(const GPUDevice& d, typename TTypes<T>::Flat var,
typename TTypes<T>::Flat m,
typename TTypes<T>::ConstScalar lr,
typename TTypes<T>::ConstScalar logbase,
typename TTypes<T>::ConstScalar sign_decay,
typename TTypes<T>::ConstScalar beta,
typename TTypes<T>::ConstFlat grad) {
Eigen::array<typename TTypes<T>::Tensor::Index, 1> bcast;
bcast[0] = grad.dimension(0);
Eigen::Sizes<1> single;
// The following is the GPU equivalent of the CPU version:
// m.device(d) = m * beta() + grad * (static_cast<T>(1) - beta());
const auto one = static_cast<T>(1.0);
auto beta_bcast = beta.reshape(single).broadcast(bcast);
auto one_minus_beta =
(beta.constant(one) - beta).reshape(single).broadcast(bcast);
m.device(d) = m * beta_bcast + grad * one_minus_beta;
// The following is the GPU equivalent of the CPU version:
// auto grad_scale = (logbase() * sign_decay() * sign_gm).exp();
// var.device(d) -= lr() * grad_scale * grad;
auto sign_gm = grad.sign() * m.sign();
auto lr_bcast = lr.reshape(single).broadcast(bcast);
auto logbase_bcast = logbase.reshape(single).broadcast(bcast);
auto sign_decay_bcast = sign_decay.reshape(single).broadcast(bcast);
auto grad_scale = (logbase_bcast * sign_decay_bcast * sign_gm).exp();
var.device(d) -= lr_bcast * grad_scale * grad;
}
};
} // namespace functor
template struct functor::ApplyGradientDescent<GPUDevice, Eigen::half>;
template struct functor::ApplyGradientDescent<GPUDevice, float>;
template struct functor::ApplyGradientDescent<GPUDevice, double>;
template struct functor::ApplyAdagrad<GPUDevice, Eigen::half>;
template struct functor::ApplyAdagrad<GPUDevice, float>;
template struct functor::ApplyAdagrad<GPUDevice, double>;
template struct functor::ApplyAdadelta<GPUDevice, Eigen::half>;
template struct functor::ApplyAdadelta<GPUDevice, float>;
template struct functor::ApplyAdadelta<GPUDevice, double>;
template struct functor::ApplyMomentum<GPUDevice, Eigen::half>;
template struct functor::ApplyMomentum<GPUDevice, float>;
template struct functor::ApplyMomentum<GPUDevice, double>;
template struct functor::ApplyAdam<GPUDevice, Eigen::half>;
template struct functor::ApplyAdam<GPUDevice, float>;
template struct functor::ApplyAdam<GPUDevice, double>;
template struct functor::ApplyRMSProp<GPUDevice, Eigen::half>;
template struct functor::ApplyRMSProp<GPUDevice, float>;
template struct functor::ApplyRMSProp<GPUDevice, double>;
template struct functor::ApplyCenteredRMSProp<GPUDevice, Eigen::half>;
template struct functor::ApplyCenteredRMSProp<GPUDevice, float>;
template struct functor::ApplyCenteredRMSProp<GPUDevice, double>;
template struct functor::ApplyAddSign<GPUDevice, Eigen::half>;
template struct functor::ApplyAddSign<GPUDevice, float>;
template struct functor::ApplyAddSign<GPUDevice, double>;
template struct functor::ApplyPowerSign<GPUDevice, Eigen::half>;
template struct functor::ApplyPowerSign<GPUDevice, float>;
template struct functor::ApplyPowerSign<GPUDevice, double>;
} // end namespace tensorflow
#endif // GOOGLE_CUDA
|
