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
|
#ifndef CAFFE2_OPERATORS_POOL_OP_H_
#define CAFFE2_OPERATORS_POOL_OP_H_
#include <vector>
#include "caffe2/core/common_omp.h"
#include "caffe2/core/context.h"
#include "caffe2/core/logging.h"
#include "caffe2/core/operator.h"
#include "caffe2/operators/conv_pool_op_base.h"
namespace caffe2 {
template <typename T, class Context, class Functor>
class PoolOp final : public ConvPoolOpBase<Context> {
public:
USE_CONV_POOL_BASE_FUNCTIONS(Context);
template <class... Args>
explicit PoolOp(Args&&... args)
: ConvPoolOpBase<Context>(std::forward<Args>(args)...), functor_(*this) {
const int kernel_size = kernel_.size();
for (const auto i : c10::irange(kernel_size)) {
CAFFE_ENFORCE_EQ(
dilation_[i], 1, "Pooling op does not support dilation right now.");
}
if (!global_pooling_) {
for (const auto i : c10::irange(kernel_size)) {
CAFFE_ENFORCE(
pads_[i] < kernel_[i] && pads_[i + kernel_size] < kernel_[i],
"Pad should be smaller than kernel.");
}
}
}
~PoolOp() = default;
bool RunOnDeviceWithOrderNCHW() override {
const auto& X = Input(0);
auto* Y = Output(0);
const int N = X.dim32(0);
const int C = X.dim32(1);
ConvPoolOpBase<Context>::SetOutputSize(X, Y, C);
const T* X_data = X.template data<T>();
T* Y_data = Y->template mutable_data<T>();
if (N == 0) {
return true;
}
if (global_pooling_) {
const int HxW = X.numel() / (N * C);
return functor_.template GlobalPoolingForward<T, StorageOrder::NCHW>(
N, C, HxW, X_data, Y_data, &context_);
}
const std::vector<int> X_HW_dims = GetDims(X);
const std::vector<int> Y_HW_dims = GetDims(*Y);
return functor_.template Forward<T, StorageOrder::NCHW>(
N,
C,
X_HW_dims,
Y_HW_dims,
kernel_,
dilation_,
stride_,
pads_,
X.template data<T>(),
Y->template mutable_data<T>(),
&context_);
}
bool RunOnDeviceWithOrderNHWC() override {
const auto& X = Input(0);
auto* Y = Output(0);
const int ndim = X.dim();
const int N = X.dim32(0);
const int C = X.dim32(ndim - 1);
ConvPoolOpBase<Context>::SetOutputSize(X, Y, C);
const T* X_data = X.template data<T>();
T* Y_data = Y->template mutable_data<T>();
if (N == 0) {
return true;
}
if (global_pooling_) {
const int HxW = X.numel() / (N * C);
return functor_.template GlobalPoolingForward<T, StorageOrder::NHWC>(
N, C, HxW, X_data, Y_data, &context_);
}
const std::vector<int> X_HW_dims = GetDims(X);
const std::vector<int> Y_HW_dims = GetDims(*Y);
return functor_.template Forward<T, StorageOrder::NHWC>(
N,
C,
X_HW_dims,
Y_HW_dims,
kernel_,
dilation_,
stride_,
pads_,
X.template data<T>(),
Y->template mutable_data<T>(),
&context_);
}
private:
const Functor functor_;
};
template <typename T, class Context, class Functor>
class PoolGradientOp final : public ConvPoolOpBase<Context> {
public:
USE_CONV_POOL_BASE_FUNCTIONS(Context);
template <class... Args>
explicit PoolGradientOp(Args&&... args)
: ConvPoolOpBase<Context>(std::forward<Args>(args)...), functor_(*this) {}
~PoolGradientOp() = default;
bool RunOnDeviceWithOrderNCHW() override {
const auto& X = Input(0);
const auto& Y = Input(1);
const auto& dY = Input(2);
auto* dX = Output(0, X.sizes(), at::dtype<T>());
const int N = X.dim32(0);
const int C = X.dim32(1);
const std::vector<int> X_HW_dims = GetDims(X);
const std::vector<int> Y_HW_dims = GetDims(Y);
ConvPoolOpBase<Context>::ComputePads(X_HW_dims);
const T* dY_data = dY.template data<T>();
const T* X_data = X.template data<T>();
const T* Y_data = Y.template data<T>();
T* dX_data = dX->template mutable_data<T>();
if (N == 0) {
return true;
}
if (global_pooling_) {
const int HxW = X.numel() / (N * C);
return functor_.template GlobalPoolingBackward<T, StorageOrder::NCHW>(
N, C, HxW, dY_data, X_data, Y_data, dX_data, &context_);
}
return functor_.template Backward<T, StorageOrder::NCHW>(
N,
C,
X_HW_dims,
Y_HW_dims,
kernel_,
dilation_,
stride_,
pads_,
dY_data,
X_data,
Y_data,
dX_data,
&context_);
}
bool RunOnDeviceWithOrderNHWC() override {
const auto& X = Input(0);
const auto& Y = Input(1);
const auto& dY = Input(2);
auto* dX = Output(0, X.sizes(), at::dtype<T>());
const int ndim = X.dim();
const int N = X.dim32(0);
const int C = X.dim32(ndim - 1);
const std::vector<int> X_HW_dims = GetDims(X);
const std::vector<int> Y_HW_dims = GetDims(Y);
ConvPoolOpBase<Context>::ComputePads(X_HW_dims);
const T* dY_data = dY.template data<T>();
const T* X_data = X.template data<T>();
const T* Y_data = Y.template data<T>();
T* dX_data = dX->template mutable_data<T>();
if (N == 0) {
return true;
}
if (global_pooling_) {
const int HxW = X.numel() / (N * C);
return functor_.template GlobalPoolingBackward<T, StorageOrder::NHWC>(
N, C, HxW, dY_data, X_data, Y_data, dX_data, &context_);
}
return functor_.template Backward<T, StorageOrder::NHWC>(
N,
C,
X_HW_dims,
Y_HW_dims,
kernel_,
dilation_,
stride_,
pads_,
dY_data,
X_data,
Y_data,
dX_data,
&context_);
}
private:
const Functor functor_;
};
template <class Context>
struct AveragePoolFunctor {
explicit AveragePoolFunctor(const OperatorBase& op)
: count_include_pad(
op.template GetSingleArgument<bool>("count_include_pad", false)) {}
template <typename T, StorageOrder kOrder>
bool GlobalPoolingForward(
int N,
int C,
int HxW,
const T* X,
T* Y,
Context* context) const;
template <typename T, StorageOrder kOrder>
bool Forward(
int N,
int C,
const std::vector<int>& X_dims,
const std::vector<int>& Y_dims,
const std::vector<int>& kernel,
const std::vector<int>& dilation,
const std::vector<int>& stride,
const std::vector<int>& pads,
const T* X,
T* Y,
Context* context) const;
template <typename T, StorageOrder kOrder>
bool GlobalPoolingBackward(
int N,
int C,
int HxW,
const T* dY,
const T* X,
const T* Y,
T* dX,
Context* context) const;
template <typename T, StorageOrder kOrder>
bool Backward(
int N,
int C,
const std::vector<int>& X_dims,
const std::vector<int>& Y_dims,
const std::vector<int>& kernel,
const std::vector<int>& dilation,
const std::vector<int>& stride,
const std::vector<int>& pads,
const T* dY,
const T* X,
const T* Y,
T* dX,
Context* context) const;
const bool count_include_pad;
Tensor ones{Context::GetDeviceType()};
};
template <class Context>
struct MaxPoolFunctor {
explicit MaxPoolFunctor(const OperatorBase& /* op */) {}
template <typename T, StorageOrder kOrder>
bool GlobalPoolingForward(
int N,
int C,
int HxW,
const T* X,
T* Y,
Context* context) const;
template <typename T, StorageOrder kOrder>
bool Forward(
int N,
int C,
const std::vector<int>& X_dims,
const std::vector<int>& Y_dims,
const std::vector<int>& kernel,
const std::vector<int>& dilation,
const std::vector<int>& stride,
const std::vector<int>& pads,
const T* X,
T* Y,
Context* context) const;
template <typename T, StorageOrder kOrder>
bool GlobalPoolingBackward(
int N,
int C,
int HxW,
const T* dY,
const T* X,
const T* Y,
T* dX,
Context* context) const;
template <typename T, StorageOrder kOrder>
bool Backward(
int N,
int C,
const std::vector<int>& X_dims,
const std::vector<int>& Y_dims,
const std::vector<int>& kernel,
const std::vector<int>& dilation,
const std::vector<int>& stride,
const std::vector<int>& pads,
const T* dY,
const T* X,
const T* Y,
T* dX,
Context* context) const;
};
} // namespace caffe2
#endif // CAFFE2_OPERATORS_POOL_OP_H_
|
