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
316
317
318
319
320
321
322
323
324
325
326
327
328
329
330
331
332
333
334
335
336
337
338
339
340
341
342
343
344
345
346
347
348
349
350
351
352
353
354
355
356
357
358
359
360
361
362
363
364
365
366
367
368
369
370
371
372
373
374
375
376
377
378
379
380
381
382
383
384
385
386
387
388
389
390
391
392
393
394
395
396
397
398
399
400
401
402
403
404
405
406
407
408
409
410
411
412
413
414
415
416
417
418
419
420
421
422
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserve.
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. */
#pragma once
#include "paddle/framework/eigen.h"
#include "paddle/framework/op_registry.h"
#include "paddle/operators/math/depthwise_conv.h"
#include "paddle/operators/math/im2col.h"
#include "paddle/operators/math/math_function.h"
#include "paddle/operators/math/vol2col.h"
namespace paddle {
namespace operators {
using Tensor = framework::Tensor;
// Base convolution operator definations for other conv
// like operators to reuse the implementation.
inline int OutputSize(int input_size, int filter_size, int dilation,
int padding, int stride) {
const int dkernel = dilation * (filter_size - 1) + 1;
const int output_size = (input_size + 2 * padding - dkernel) / stride + 1;
return output_size;
}
inline bool IsExpand(std::vector<int64_t>& filter_dim,
std::vector<int>& strides, std::vector<int>& paddings,
std::vector<int>& dilations) {
bool filter_1 = true, strides_1 = true, padding_0 = true, dilation_1 = true;
for (size_t j = 0; j < strides.size(); ++j) {
filter_1 = filter_1 && (static_cast<int>(filter_dim[j + 2]) == 1);
strides_1 = strides_1 && (strides[j] == 1);
padding_0 = padding_0 && (paddings[j] == 0);
dilation_1 = dilation_1 && (dilations[j] == 1);
}
return !(filter_1 && strides_1 && padding_0 && dilation_1);
}
// Define Op classes in .h file so that other conv
// operator implementations can reuse the code.
class Conv2DOpMaker : public framework::OpProtoAndCheckerMaker {
public:
Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker);
};
class Conv3DOpMaker : public framework::OpProtoAndCheckerMaker {
public:
Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker);
};
class ConvOp : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override;
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override;
};
class ConvOpGrad : public framework::OperatorWithKernel {
public:
using framework::OperatorWithKernel::OperatorWithKernel;
void InferShape(framework::InferShapeContext* ctx) const override;
protected:
framework::OpKernelType GetExpectedKernelType(
const framework::ExecutionContext& ctx) const override;
};
template <typename DeviceContext, typename T>
class GemmConvKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
// The filter will be reshaped in the calculations,
// so here use an assignment operation,
// that avoids modifying the variable in the Scope.
Tensor filter = *context.Input<Tensor>("Filter");
Tensor* output = context.Output<Tensor>("Output");
output->mutable_data<T>(context.GetPlace());
int groups = context.Attr<int>("groups");
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
const int batch_size = static_cast<int>(input->dims()[0]);
// filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w}
std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
// output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w}
std::vector<int64_t> output_shape_vec(framework::vectorize(output->dims()));
// use col_shape in the im2col calculation
// col_shape_vec: {i_c/g, k_h, k_w, o_h, o_w} or {i_c/g, k_d, k_h, k_w, o_d,
// o_h, o_w}
size_t data_dim = filter_shape_vec.size() - 2;
std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
col_shape_vec[0] = input->dims()[1] / groups;
for (size_t j = 0; j < data_dim; ++j) {
col_shape_vec[j + 1] = filter_shape_vec[j + 2];
col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
}
framework::DDim col_shape(framework::make_ddim(col_shape_vec));
// use col_matrix_shape in the gemm calculation
// size: (i_c/g * k_h * k_w, o_h * o_w) or (i_c/g * k_d * k_h * k_w, o_d *
// o_h * o_w)
framework::DDim col_matrix_shape =
framework::flatten_to_2d(col_shape, data_dim + 1);
bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations);
Tensor col;
// col_matrix shares the same piece of data with col,
// but will be reshaped into a two-dimensional matrix shape
// to call the matrix multiplication interface.
Tensor col_matrix;
if (is_expand) {
col.mutable_data<T>(col_shape, context.GetPlace());
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
filter.Resize(filter_matrix_shape);
framework::DDim output_matrix_shape = {
output->dims()[1],
output->numel() / (output->dims()[0] * output->dims()[1])};
// convolution operator: im2col(or vol2col) + gemm
int in_step = static_cast<int>(input->dims()[1]) / groups;
int out_step = static_cast<int>(output->dims()[1]) / groups;
math::Vol2ColFunctor<DeviceContext, T> vol2col;
math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
auto& dev_ctx = context.template device_context<DeviceContext>();
for (int i = 0; i < batch_size; i++) {
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
Tensor out_batch = output->Slice(i, i + 1).Resize(output_matrix_shape);
for (int g = 0; g < groups; g++) {
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
if (!is_expand) {
col.ShareDataWith(in_slice);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
} else if (data_dim == 2U) {
// im2col
im2col(dev_ctx, in_slice, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&col);
} else if (data_dim == 3U) {
// vol2col
vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col);
}
// gemm
Tensor out_slice = out_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
math::matmul<DeviceContext, T>(dev_ctx, filter_slice, false, col_matrix,
false, T(1.0), &out_slice, T(0.0));
}
}
}
};
template <typename DeviceContext, typename T>
class GemmConvGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
const Tensor* output_grad =
context.Input<Tensor>(framework::GradVarName("Output"));
Tensor* input_grad =
context.Output<Tensor>(framework::GradVarName("Input"));
Tensor* filter_grad =
context.Output<Tensor>(framework::GradVarName("Filter"));
// The filter and filter_grad will be reshaped in the calculations,
// so here use an assignment operation,
// that avoids modifying the variable in the Scope.
Tensor filter = *context.Input<Tensor>("Filter");
if (!input_grad && !filter_grad) return;
int groups = context.Attr<int>("groups");
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
const int batch_size = static_cast<int>(input->dims()[0]);
// filter_shape_vec: {k_o, k_i, k_h, k_w} or {k_o, k_i, k_d, k_h, k_w}
std::vector<int64_t> filter_shape_vec(framework::vectorize(filter.dims()));
// output_shape_vec: {o_n, o_c, o_h, o_w} or {o_n, o_c, o_d, o_h, o_w}
std::vector<int64_t> output_shape_vec(
framework::vectorize(output_grad->dims()));
// use col_shape in the im2col calculation
// col_shape_vec: {i_c/g, k_h, k_w, o_h, o_w} or {i_c/g, k_d, k_h, k_w, o_d,
// o_h, o_w}
size_t data_dim = filter_shape_vec.size() - 2;
std::vector<int64_t> col_shape_vec(1 + 2 * data_dim);
col_shape_vec[0] = input->dims()[1] / groups;
for (size_t j = 0; j < data_dim; ++j) {
col_shape_vec[j + 1] = filter_shape_vec[j + 2];
col_shape_vec[j + 1 + data_dim] = output_shape_vec[j + 2];
}
framework::DDim col_shape(framework::make_ddim(col_shape_vec));
// use col_matrix_shape in the gemm calculation
// size: (i_c/g * k_h * k_w, o_h * o_w)
// or
// (i_c/g * k_d * k_h * k_w, o_d * o_h * o_w)
framework::DDim col_matrix_shape =
framework::flatten_to_2d(col_shape, data_dim + 1);
framework::DDim input_shape = framework::slice_ddim(
input->dims(), 1, static_cast<int>(input->dims().size()));
framework::DDim filter_matrix_shape = {filter.dims()[0],
filter.numel() / filter.dims()[0]};
filter.Resize(filter_matrix_shape);
framework::DDim output_matrix_shape = {
output_grad->dims()[1],
output_grad->numel() /
(output_grad->dims()[0] * output_grad->dims()[1])};
// convolution backward input operator: gemm + col2im(or col2vol)
// convolution backward weight operator: im2col(or vol2col) + gemm
int in_step = static_cast<int>(input->dims()[1]) / groups;
int out_step = static_cast<int>(output_grad->dims()[1]) / groups;
bool is_expand = IsExpand(filter_shape_vec, strides, paddings, dilations);
Tensor col;
// col_matrix shares the same piece of data with col,
// but will be reshaped into a two-dimensional matrix shape
// to call the matrix multiplication interface.
Tensor col_matrix;
if (is_expand) {
col.mutable_data<T>(col_shape, context.GetPlace());
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
}
math::SetConstant<DeviceContext, T> set_zero;
auto& dev_ctx = context.template device_context<DeviceContext>();
if (input_grad) {
input_grad->mutable_data<T>(context.GetPlace());
// if is_expand is false, the operation of set_zero is unnecessary,
// because math::matmul will reset input_grad.
if (is_expand) {
set_zero(dev_ctx, input_grad, static_cast<T>(0));
}
math::Col2VolFunctor<DeviceContext, T> col2vol;
math::Col2ImFunctor<math::ColFormat::kCFO, DeviceContext, T> col2im;
for (int i = 0; i < batch_size; i++) {
Tensor out_grad_batch =
output_grad->Slice(i, i + 1).Resize(output_matrix_shape);
Tensor in_grad_batch = input_grad->Slice(i, i + 1).Resize(input_shape);
for (int g = 0; g < groups; g++) {
// gemm
Tensor out_grad_slice =
out_grad_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor filter_slice = filter.Slice(g * out_step, (g + 1) * out_step);
Tensor in_grad_slice =
in_grad_batch.Slice(g * in_step, (g + 1) * in_step);
if (!is_expand) {
col_matrix.ShareDataWith(in_grad_slice);
col_matrix.Resize(col_matrix_shape);
}
math::matmul<DeviceContext, T>(dev_ctx, filter_slice, true,
out_grad_slice, false, T(1.0),
&col_matrix, T(0.0));
if (is_expand && data_dim == 2U) {
col2im(dev_ctx, col, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&in_grad_slice);
} else if (is_expand && data_dim == 3U) {
col2vol(dev_ctx, col, dilations, strides, paddings, &in_grad_slice);
}
}
}
}
if (filter_grad) {
filter_grad->mutable_data<T>(context.GetPlace());
Tensor filter_grad_ = *filter_grad;
filter_grad_.Resize(filter_matrix_shape);
set_zero(dev_ctx, filter_grad, static_cast<T>(0));
math::Im2ColFunctor<math::ColFormat::kCFO, DeviceContext, T> im2col;
math::Vol2ColFunctor<DeviceContext, T> vol2col;
for (int i = 0; i < batch_size; i++) {
Tensor out_grad_batch =
output_grad->Slice(i, i + 1).Resize(output_matrix_shape);
Tensor in_batch = input->Slice(i, i + 1).Resize(input_shape);
for (int g = 0; g < groups; g++) {
// im2col
Tensor out_grad_slice =
out_grad_batch.Slice(g * out_step, (g + 1) * out_step);
Tensor in_slice = in_batch.Slice(g * in_step, (g + 1) * in_step);
if (!is_expand) {
col.ShareDataWith(in_slice);
col_matrix.ShareDataWith(col);
col_matrix.Resize(col_matrix_shape);
} else if (data_dim == 2U) {
im2col(dev_ctx, in_slice, dilations, strides,
std::vector<int>{paddings[0], paddings[1], paddings[0],
paddings[1]},
&col);
} else if (data_dim == 3U) {
vol2col(dev_ctx, in_slice, dilations, strides, paddings, &col);
}
// gemm
Tensor filter_grad_slice =
filter_grad_.Slice(g * out_step, (g + 1) * out_step);
math::matmul<DeviceContext, T>(dev_ctx, out_grad_slice, false,
col_matrix, true, T(1.0),
&filter_grad_slice, T(1.0));
}
}
}
}
};
template <typename DeviceContext, typename T>
class DepthwiseConvKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
Tensor filter = *context.Input<Tensor>("Filter");
Tensor* output = context.Output<Tensor>("Output");
output->mutable_data<T>(context.GetPlace());
PADDLE_ENFORCE_EQ(
output->dims()[1] % input->dims()[1], 0,
"The output channels must be a multiple of the input channels");
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
math::DepthwiseConvFunctor<DeviceContext, T> depthwiseConv;
auto& dev_ctx = context.template device_context<DeviceContext>();
depthwiseConv(dev_ctx, *input, filter, strides, paddings, output);
}
};
template <typename DeviceContext, typename T>
class DepthwiseConvGradKernel : public framework::OpKernel<T> {
public:
void Compute(const framework::ExecutionContext& context) const override {
const Tensor* input = context.Input<Tensor>("Input");
const Tensor* output_grad =
context.Input<Tensor>(framework::GradVarName("Output"));
Tensor* input_grad =
context.Output<Tensor>(framework::GradVarName("Input"));
Tensor* filter_grad =
context.Output<Tensor>(framework::GradVarName("Filter"));
Tensor filter = *context.Input<Tensor>("Filter");
if (!input_grad && !filter_grad) return;
std::vector<int> strides = context.Attr<std::vector<int>>("strides");
std::vector<int> paddings = context.Attr<std::vector<int>>("paddings");
std::vector<int> dilations = context.Attr<std::vector<int>>("dilations");
math::SetConstant<DeviceContext, T> set_zero;
auto& dev_ctx = context.template device_context<DeviceContext>();
math::DepthwiseConvInputGradFunctor<DeviceContext, T>
depthwiseConvInputGrad;
math::DepthwiseConvFilterGradFunctor<DeviceContext, T>
depthwiseConvFilterGrad;
if (input_grad) {
input_grad->mutable_data<T>(context.GetPlace());
set_zero(dev_ctx, input_grad, static_cast<T>(0));
depthwiseConvInputGrad(dev_ctx, *input, filter, *output_grad, strides,
paddings, input_grad);
}
if (filter_grad) {
filter_grad->mutable_data<T>(context.GetPlace());
set_zero(dev_ctx, filter_grad, static_cast<T>(0));
depthwiseConvFilterGrad(dev_ctx, *input, *output_grad, strides, paddings,
filter_grad);
}
}
};
} // namespace operators
} // namespace paddle