conv_op.cc 15.9 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
C
chengduoZH 已提交
2

L
Luo Tao 已提交
3 4 5
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
C
chengduoZH 已提交
6

L
Luo Tao 已提交
7
    http://www.apache.org/licenses/LICENSE-2.0
C
chengduoZH 已提交
8

L
Luo Tao 已提交
9 10 11 12 13
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. */
C
chengduoZH 已提交
14

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/conv_op.h"
Y
Update  
Yi Wang 已提交
16 17 18 19

#include <string>
#include <vector>

20 21 22 23 24 25
#ifdef PADDLE_WITH_CUDA
#include "paddle/fluid/platform/cudnn_helper.h"
#endif
#ifdef PADDLE_WITH_MKLDNN
#include "paddle/fluid/platform/mkldnn_helper.h"
#endif
C
chengduoZH 已提交
26 27 28 29

namespace paddle {
namespace operators {

C
chengduoZH 已提交
30
void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
C
chengduoZH 已提交
31
  PADDLE_ENFORCE(ctx->HasInput("Input"),
C
chengduoZH 已提交
32
                 "Input(Input) of ConvOp should not be null.");
C
chengduoZH 已提交
33
  PADDLE_ENFORCE(ctx->HasInput("Filter"),
C
chengduoZH 已提交
34
                 "Input(Filter) of ConvOp should not be null.");
C
chengduoZH 已提交
35
  PADDLE_ENFORCE(ctx->HasOutput("Output"),
C
chengduoZH 已提交
36
                 "Output(Output) of ConvOp should not be null.");
C
chengduoZH 已提交
37 38 39 40 41 42

  auto in_dims = ctx->GetInputDim("Input");
  auto filter_dims = ctx->GetInputDim("Filter");
  std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
  std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
  int groups = ctx->Attrs().Get<int>("groups");
C
chengduoZH 已提交
43
  std::vector<int> dilations = ctx->Attrs().Get<std::vector<int>>("dilations");
C
chengduoZH 已提交
44

C
chengduoZH 已提交
45 46
  PADDLE_ENFORCE(in_dims.size() == 4 || in_dims.size() == 5,
                 "Conv intput should be 4-D or 5-D tensor.");
C
chengduoZH 已提交
47 48 49 50 51 52 53 54 55
  PADDLE_ENFORCE_EQ(
      in_dims.size(), filter_dims.size(),
      "Conv input dimension and filter dimension should be the same.");
  PADDLE_ENFORCE(
      in_dims.size() - strides.size() == 2U,
      "Conv input dimension and strides dimension should be consistent.");
  PADDLE_ENFORCE_EQ(
      paddings.size(), strides.size(),
      "Conv paddings dimension and Conv strides dimension should be the same.");
F
fengjiayi 已提交
56

Y
Yang Yu 已提交
57
  PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[1] * groups,
C
chengduoZH 已提交
58
                    "The number of input channels should be equal to filter "
C
chengduoZH 已提交
59
                    "channels * groups.");
F
fengjiayi 已提交
60

C
chengduoZH 已提交
61
  PADDLE_ENFORCE_EQ(
Y
Yang Yu 已提交
62
      filter_dims[0] % groups, 0,
C
chengduoZH 已提交
63 64 65
      "The number of output channels should be divided by groups.");

  std::vector<int64_t> output_shape({in_dims[0], filter_dims[0]});
C
chengduoZH 已提交
66
  for (size_t i = 0; i < strides.size(); ++i) {
Y
Yang Yang 已提交
67 68 69
    output_shape.push_back(ConvOutputSize(in_dims[i + 2], filter_dims[i + 2],
                                          dilations[i], paddings[i],
                                          strides[i]));
C
chengduoZH 已提交
70
  }
71
  ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
Y
Yang Yu 已提交
72
  ctx->ShareLoD("Input", "Output");
C
chengduoZH 已提交
73 74
}

75 76
framework::OpKernelType ConvOp::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
77
  framework::LibraryType library{framework::LibraryType::kPlain};
C
chengduoZH 已提交
78
#ifdef PADDLE_WITH_CUDA
79
  if (platform::CanCUDNNBeUsed(ctx)) {
80
    library = framework::LibraryType::kCUDNN;
C
chengduoZH 已提交
81 82
  }
#endif
83
#ifdef PADDLE_WITH_MKLDNN
84
  if (library == framework::LibraryType::kPlain &&
85
      platform::CanMKLDNNBeUsed(ctx)) {
86
    library = framework::LibraryType::kMKLDNN;
87
  }
88
#endif
89

K
Kexin Zhao 已提交
90 91 92 93 94 95 96 97
  auto input_data_type =
      framework::ToDataType(ctx.Input<Tensor>("Input")->type());
  auto filter_data_type =
      framework::ToDataType(ctx.Input<Tensor>("Filter")->type());
  PADDLE_ENFORCE_EQ(input_data_type, filter_data_type,
                    "input and filter data type should be consistent");

  if (input_data_type == framework::proto::VarType::FP16) {
98
    PADDLE_ENFORCE_EQ(library, framework::LibraryType::kCUDNN,
K
Kexin Zhao 已提交
99 100 101
                      "float16 can only be used when CUDNN is used");
  }

102
  std::string data_format = ctx.Attr<std::string>("data_format");
103
  // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
104 105 106
  framework::DataLayout layout = framework::StringToDataLayout(data_format);
  return framework::OpKernelType(input_data_type, ctx.GetPlace(), layout,
                                 library);
107 108
}

109
Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
C
chengduoZH 已提交
110 111 112
    : OpProtoAndCheckerMaker(proto, op_checker) {
  AddInput(
      "Input",
C
fix doc  
chengduoZH 已提交
113 114 115 116
      "(Tensor) The input tensor of convolution operator. "
      "The format of input tensor is NCHW, where N is batch size, C is the "
      "number of channels, H is the height of the feature, "
      "and W is the width of the feature.");
C
chengduoZH 已提交
117
  AddInput("Filter",
C
fix doc  
chengduoZH 已提交
118
           "(Tensor) The filter tensor of convolution operator. "
C
chengduoZH 已提交
119 120
           "The format of the filter tensor is MCHW, where M is the number of "
           "output image channels, C is the number of input image channels, "
C
fix doc  
chengduoZH 已提交
121 122
           "H is the height of the filter, and W is the width of the filter. "
           "If the groups attribute is greater than 1, C equals the number of "
C
chengduoZH 已提交
123 124
           "input image channels divided by the groups.");
  AddOutput("Output",
C
fix doc  
chengduoZH 已提交
125 126
            "(Tensor) The output tensor of convolution operator. "
            "The format of output tensor is also NCHW.");
C
chengduoZH 已提交
127 128 129 130
  AddAttr<std::vector<int>>("strides",
                            "(vector<int> default:{1, 1}), the "
                            "strides(h_stride, w_stride) of "
                            "convolution operator.")
C
chengduoZH 已提交
131
      .SetDefault({1, 1});
C
chengduoZH 已提交
132 133 134 135
  AddAttr<std::vector<int>>("paddings",
                            "(vector<int> default:{0, 0}), the "
                            "paddings(h_pad, w_pad) of "
                            "convolution operator.")
C
chengduoZH 已提交
136 137 138
      .SetDefault({0, 0});
  AddAttr<int>(
      "groups",
C
chengduoZH 已提交
139
      "(int default:1), the groups number of the convolution operator. "
C
fix doc  
chengduoZH 已提交
140 141 142 143
      "According to grouped convolution in Alex Krizhevsky's Deep CNN paper: "
      "when group=2, the first half of the filters is only connected to the "
      "first half of the input channels, while the second half of the filters "
      "is only connected to the second half of the input channels.")
C
chengduoZH 已提交
144
      .SetDefault(1);
C
chengduoZH 已提交
145
  AddAttr<std::vector<int>>("dilations",
C
chengduoZH 已提交
146 147
                            "(vector<int> default:{1, 1}), the "
                            "dilations(h_dilation, w_dilation) of "
C
chengduoZH 已提交
148
                            "convolution operator.")
C
chengduoZH 已提交
149
      .SetDefault({1, 1});
150 151 152 153
  AddAttr<bool>(
      "use_cudnn",
      "(bool, default false) Only used in cudnn kernel, need install cudnn")
      .SetDefault(false);
154 155 156
  AddAttr<bool>("use_mkldnn",
                "(bool, default false) Only used in mkldnn kernel")
      .SetDefault(false);
157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172
  AddAttr<std::string>(
      "data_format",
      "(string, default NCHW) Only used in "
      "An optional string from: \"NHWC\", \"NCHW\". "
      "Defaults to \"NHWC\". Specify the data format of the output data, "
      "the input will be transformed automatically. ")
      .SetDefault("AnyLayout");
  // TODO(dzhwinter): need to registered layout transform function
  AddAttr<int>("workspace_size_MB",
               "Only used in cudnn kernel. Need set use_cudnn to true."
               "workspace size for cudnn, in MB, "
               "workspace is a section of GPU memory which will be "
               "allocated/freed each time the operator runs, larger "
               "workspace size can increase performance but also requires "
               "better hardware. This size should be chosen carefully.")
      .SetDefault(4096);
C
chengduoZH 已提交
173
  AddComment(R"DOC(
C
fix doc  
chengduoZH 已提交
174 175
Convolution Operator.

C
chengduoZH 已提交
176
The convolution operation calculates the output based on the input, filter
C
chengduoZH 已提交
177
and strides, paddings, dilations, groups parameters. The size of each dimension of the
C
chengduoZH 已提交
178
parameters is checked in the infer-shape.
C
chengduoZH 已提交
179
Input(Input) and Output(Output) are in NCHW format. Where N is batch
C
fix doc  
chengduoZH 已提交
180
size, C is the number of channels, H is the height of the feature, and W is
C
chengduoZH 已提交
181 182 183 184 185 186
the width of the feature.
Filters(Input) is MCHW format. Where M is the number of output image channels, C is
the number of input image channels, H is the height of the filter, and W
is the width of the filter.
Parameters(strides, paddings, dilations) are two elements. These two elements represent
height and width, respectively.
C
chengduoZH 已提交
187 188 189 190
The input(X) size and output(Out) size may be different.

Example:
  Input:
C
chengduoZH 已提交
191 192
       Input shape: $(N, C_{in}, H_{in}, W_{in})$
       Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
C
chengduoZH 已提交
193
  Output:
C
chengduoZH 已提交
194 195 196 197 198 199
       Output shape: $(N, C_{out}, H_{out}, W_{out})$
  Where
$$
       H_{out}= \frac{(H_{in} + 2 * paddings[0] - (dilations[0] * (H_f - 1) + 1))}{strides[0]}+ 1 \\
       W_{out}= \frac{(W_{in} + 2 * paddings[1] - (dilations[1] * (W_f - 1) + 1))}{strides[1]}+ 1
$$
C
chengduoZH 已提交
200
)DOC");
C
chengduoZH 已提交
201 202
}

203
Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
C
chengduoZH 已提交
204 205 206
    : OpProtoAndCheckerMaker(proto, op_checker) {
  AddInput(
      "Input",
C
fix doc  
chengduoZH 已提交
207
      "(Tensor) The input tensor of convolution operator. "
C
chengduoZH 已提交
208
      "The format of input tensor is NCDHW. Where N is batch size, C is the "
C
fix doc  
chengduoZH 已提交
209 210 211
      "number of channels, D is the depth of the feature, H is the height of "
      "the feature, "
      "and W is the width of the feature.");
C
chengduoZH 已提交
212
  AddInput("Filter",
C
fix doc  
chengduoZH 已提交
213
           "(Tensor) The filter tensor of convolution operator. "
C
chengduoZH 已提交
214 215
           "The format of the filter tensor is MCDHW, where M is the number of "
           "output image channels, C is the number of input image channels, "
C
fix doc  
chengduoZH 已提交
216 217 218
           "D is the depth of the filter, H is the height of the filter, and W "
           "is the width of the filter."
           "If the groups attribute is greater than 1, C equals the number of "
C
chengduoZH 已提交
219 220
           "input image channels divided by the groups.");
  AddOutput("Output",
C
fix doc  
chengduoZH 已提交
221
            "(Tensor) The output tensor of convolution operator."
C
chengduoZH 已提交
222
            "The format of output tensor is also NCDHW.");
C
chengduoZH 已提交
223 224 225 226
  AddAttr<std::vector<int>>("strides",
                            "(vector<int>, default:{1, 1, 1}), the "
                            "strides(d_stride, h_stride, w_stride) of "
                            "convolution operator.")
C
chengduoZH 已提交
227
      .SetDefault({1, 1, 1});
C
chengduoZH 已提交
228 229 230 231
  AddAttr<std::vector<int>>("paddings",
                            "(vector<int>, default:{0, 0, 0}), the "
                            "paddings(d_pad, h_pad, w_pad) of convolution "
                            "operator.")
C
chengduoZH 已提交
232 233 234
      .SetDefault({0, 0, 0});
  AddAttr<int>(
      "groups",
C
chengduoZH 已提交
235
      "(int default:1), the groups number of the convolution operator. "
C
fix doc  
chengduoZH 已提交
236 237 238 239
      "According to grouped convolution in Alex Krizhevsky's Deep CNN paper: "
      "when group=2, the first half of the filters is only connected to the "
      "first half of the input channels, while the second half of the filters "
      "is only connected to the second half of the input channels.")
C
chengduoZH 已提交
240
      .SetDefault(1);
C
chengduoZH 已提交
241
  AddAttr<std::vector<int>>("dilations",
C
chengduoZH 已提交
242 243
                            "(vector<int> default:{1, 1, 1}), the "
                            "dilations(d_dilation, h_dilation, w_dilation) of "
C
chengduoZH 已提交
244
                            "convolution operator.")
C
chengduoZH 已提交
245
      .SetDefault({1, 1, 1});
246 247 248 249
  AddAttr<bool>(
      "use_cudnn",
      "(bool, default false) Only used in cudnn kernel, need install cudnn")
      .SetDefault(false);
250 251 252
  AddAttr<bool>("use_mkldnn",
                "(bool, default false) Only used in mkldnn kernel")
      .SetDefault(false);
253 254 255 256 257 258 259 260 261 262 263 264 265 266 267
  AddAttr<std::string>(
      "data_format",
      "(string, default NCHW) Only used in "
      "An optional string from: \"NHWC\", \"NCHW\". "
      "Defaults to \"NHWC\". Specify the data format of the output data, "
      "the input will be transformed automatically. ")
      .SetDefault("AnyLayout");
  // TODO(dzhwinter): need to registered layout transform function
  AddAttr<int>("workspace_size_MB",
               "Only used in cudnn kernel. workspace size for cudnn, in MB, "
               "workspace is a section of GPU memory which will be "
               "allocated/freed each time the operator runs, larger "
               "workspace size can increase performance but also requires "
               "better hardware. This size should be chosen carefully.")
      .SetDefault(4096);
C
fix doc  
chengduoZH 已提交
268

C
chengduoZH 已提交
269
  AddComment(R"DOC(
C
fix doc  
chengduoZH 已提交
270 271
Convolution3D Operator.

C
chengduoZH 已提交
272
The convolution operation calculates the output based on the input, filter
C
chengduoZH 已提交
273
and strides, paddings, dilations, groups parameters. The size of each dimension of the
C
chengduoZH 已提交
274
parameters is checked in the infer-shape.
C
chengduoZH 已提交
275
Input(Input) and output(Output) are in NCDHW format, where N is batch
C
fix doc  
chengduoZH 已提交
276
size, C is the number of channels,D is the depth of the feature, H is the height of
C
chengduoZH 已提交
277 278 279 280 281 282
the feature, and W is the width of the feature.
Filters(Input) is MCDHW format, where M is the number of output image channels,
C is the number of input image channels, D is the depth of the filter,
H is the height of the filter, and W is the width of the filter.
Parameters(strides, paddings, dilations) are three elements. These three elements
represent depth, height and width, respectively.
C
fix doc  
chengduoZH 已提交
283 284 285 286
The input(X) size and output(Out) size may be different.

Example:
  Input:
C
chengduoZH 已提交
287 288
       Input shape: $(N, C_{in}, D_{in}, H_{in}, W_{in})$
       Filter shape: $(C_{out}, C_{in}, D_f, H_f, W_f)$
C
fix doc  
chengduoZH 已提交
289
  Output:
C
chengduoZH 已提交
290 291 292 293 294 295 296
       Output shape: $(N, C_{out}, D_{out}, H_{out}, W_{out})$
  Where
  $$
       D_{out}= \frac{(D_{in} + 2 * paddings[0] - (dilations[0] * (D_f - 1) + 1))}{ strides[0]}+ 1 \\
       H_{out}= \frac{(H_{in} + 2 * paddings[1] - (dilations[1] * (H_f - 1) + 1))}{ strides[1]}+ 1 \\
       W_{out}= \frac{(W_{in} + 2 * paddings[2] - (dilations[2] * (W_f - 1) + 1))}{ strides[2]}+ 1
  $$
C
chengduoZH 已提交
297 298 299
)DOC");
}

C
chengduoZH 已提交
300 301 302 303 304 305 306 307 308 309 310
void ConvOpGrad::InferShape(framework::InferShapeContext* ctx) const {
  auto in_dims = ctx->GetInputDim("Input");
  auto filter_dims = ctx->GetInputDim("Filter");
  if (ctx->HasOutput(framework::GradVarName("Input"))) {
    ctx->SetOutputDim(framework::GradVarName("Input"), in_dims);
  }
  if (ctx->HasOutput(framework::GradVarName("Filter"))) {
    ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims);
  }
}

311 312
framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
313
  framework::LibraryType library_{framework::LibraryType::kPlain};
C
chengduoZH 已提交
314
#ifdef PADDLE_WITH_CUDA
315 316
  if (platform::CanCUDNNBeUsed(ctx)) {
    library_ = framework::LibraryType::kCUDNN;
C
chengduoZH 已提交
317 318
  }
#endif
319 320 321 322
#ifdef PADDLE_WITH_MKLDNN
  if (library_ == framework::LibraryType::kPlain &&
      platform::CanMKLDNNBeUsed(ctx)) {
    library_ = framework::LibraryType::kMKLDNN;
323
  }
324
#endif
325 326

  std::string data_format = ctx.Attr<std::string>("data_format");
327
  // TODO(pzelazko-intel): enable MKLDNN layout when it's ready
328 329 330 331 332 333
  framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
  return framework::OpKernelType(
      framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
      layout_, library_);
}

C
chengduoZH 已提交
334 335 336 337
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
C
chengduoZH 已提交
338 339
REGISTER_OP(conv2d, ops::ConvOp, ops::Conv2DOpMaker, conv2d_grad,
            ops::ConvOpGrad);
340 341

// depthwise convolution op
342 343
REGISTER_OP(depthwise_conv2d, ops::ConvOp, ops::Conv2DOpMaker,
            depthwise_conv2d_grad, ops::ConvOpGrad);
C
chengduoZH 已提交
344 345 346
REGISTER_OP(conv3d, ops::ConvOp, ops::Conv3DOpMaker, conv3d_grad,
            ops::ConvOpGrad);

347 348
// depthwise conv kernel
// TODO(xingzhaolong): neon kernel for mobile
Z
zlx 已提交
349
REGISTER_OP_CPU_KERNEL(
350
    depthwise_conv2d,
X
xzl 已提交
351 352 353 354
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);

REGISTER_OP_CPU_KERNEL(
355
    depthwise_conv2d_grad,
X
xzl 已提交
356 357
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
Z
zlx 已提交
358

C
chengduoZH 已提交
359
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
360 361 362 363 364 365
    conv2d, ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
    conv2d_grad,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
C
chengduoZH 已提交
366 367

REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
368 369 370 371 372 373
    conv3d, ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
    conv3d_grad,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);