conv_op.cc 15.4 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"
C
chengduoZH 已提交
16 17 18 19

namespace paddle {
namespace operators {

C
chengduoZH 已提交
20
void ConvOp::InferShape(framework::InferShapeContext* ctx) const {
C
chengduoZH 已提交
21
  PADDLE_ENFORCE(ctx->HasInput("Input"),
C
chengduoZH 已提交
22
                 "Input(Input) of ConvOp should not be null.");
C
chengduoZH 已提交
23
  PADDLE_ENFORCE(ctx->HasInput("Filter"),
C
chengduoZH 已提交
24
                 "Input(Filter) of ConvOp should not be null.");
C
chengduoZH 已提交
25
  PADDLE_ENFORCE(ctx->HasOutput("Output"),
C
chengduoZH 已提交
26
                 "Output(Output) of ConvOp should not be null.");
C
chengduoZH 已提交
27 28 29 30 31 32

  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
33
  std::vector<int> dilations = ctx->Attrs().Get<std::vector<int>>("dilations");
C
chengduoZH 已提交
34

C
chengduoZH 已提交
35 36
  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
37 38 39 40 41 42 43 44 45
  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
46

Y
Yang Yu 已提交
47
  PADDLE_ENFORCE_EQ(in_dims[1], filter_dims[1] * groups,
C
chengduoZH 已提交
48
                    "The number of input channels should be equal to filter "
C
chengduoZH 已提交
49
                    "channels * groups.");
F
fengjiayi 已提交
50

C
chengduoZH 已提交
51
  PADDLE_ENFORCE_EQ(
Y
Yang Yu 已提交
52
      filter_dims[0] % groups, 0,
C
chengduoZH 已提交
53 54 55
      "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
56
  for (size_t i = 0; i < strides.size(); ++i) {
C
chengduoZH 已提交
57 58 59 60 61 62
    PADDLE_ENFORCE(in_dims[i + 2] + 2 * paddings[i] -
                           (dilations[i] * (filter_dims[i + 2] - 1) + 1) >
                       0,
                   "Due to the settings of paddings, filter_dims and "
                   "dilations, the output size is less than 0, please check "
                   "again.");
C
chengduoZH 已提交
63
    output_shape.push_back(OutputSize(in_dims[i + 2], filter_dims[i + 2],
C
chengduoZH 已提交
64
                                      dilations[i], paddings[i], strides[i]));
C
chengduoZH 已提交
65
  }
66
  ctx->SetOutputDim("Output", framework::make_ddim(output_shape));
Y
Yang Yu 已提交
67
  ctx->ShareLoD("Input", "Output");
C
chengduoZH 已提交
68 69
}

70 71 72
framework::OpKernelType ConvOp::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
  bool use_cudnn = ctx.Attr<bool>("use_cudnn");
C
chengduoZH 已提交
73
  use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
C
chengduoZH 已提交
74 75 76 77 78 79
#ifdef PADDLE_WITH_CUDA
  if (platform::is_gpu_place(ctx.GetPlace())) {
    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
    use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
  }
#endif
80 81 82 83 84 85 86 87 88 89 90 91 92 93
  framework::LibraryType library_;
  if (use_cudnn) {
    library_ = framework::LibraryType::kCUDNN;
  } else {
    library_ = framework::LibraryType::kPlain;
  }

  std::string data_format = ctx.Attr<std::string>("data_format");
  framework::DataLayout layout_ = framework::StringToDataLayout(data_format);
  return framework::OpKernelType(
      framework::ToDataType(ctx.Input<Tensor>("Input")->type()), ctx.GetPlace(),
      layout_, library_);
}

94
Conv2DOpMaker::Conv2DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
C
chengduoZH 已提交
95 96 97
    : OpProtoAndCheckerMaker(proto, op_checker) {
  AddInput(
      "Input",
C
fix doc  
chengduoZH 已提交
98 99 100 101
      "(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
102
  AddInput("Filter",
C
fix doc  
chengduoZH 已提交
103
           "(Tensor) The filter tensor of convolution operator. "
C
chengduoZH 已提交
104 105
           "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
106 107
           "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
108 109
           "input image channels divided by the groups.");
  AddOutput("Output",
C
fix doc  
chengduoZH 已提交
110 111
            "(Tensor) The output tensor of convolution operator. "
            "The format of output tensor is also NCHW.");
C
chengduoZH 已提交
112 113 114 115
  AddAttr<std::vector<int>>("strides",
                            "(vector<int> default:{1, 1}), the "
                            "strides(h_stride, w_stride) of "
                            "convolution operator.")
C
chengduoZH 已提交
116
      .SetDefault({1, 1});
C
chengduoZH 已提交
117 118 119 120
  AddAttr<std::vector<int>>("paddings",
                            "(vector<int> default:{0, 0}), the "
                            "paddings(h_pad, w_pad) of "
                            "convolution operator.")
C
chengduoZH 已提交
121 122 123
      .SetDefault({0, 0});
  AddAttr<int>(
      "groups",
C
chengduoZH 已提交
124
      "(int default:1), the groups number of the convolution operator. "
C
fix doc  
chengduoZH 已提交
125 126 127 128
      "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
129
      .SetDefault(1);
C
chengduoZH 已提交
130
  AddAttr<std::vector<int>>("dilations",
C
chengduoZH 已提交
131 132
                            "(vector<int> default:{1, 1}), the "
                            "dilations(h_dilation, w_dilation) of "
C
chengduoZH 已提交
133
                            "convolution operator.")
C
chengduoZH 已提交
134
      .SetDefault({1, 1});
135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154
  AddAttr<bool>(
      "use_cudnn",
      "(bool, default false) Only used in cudnn kernel, need install cudnn")
      .SetDefault(false);
  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
155
  AddComment(R"DOC(
C
fix doc  
chengduoZH 已提交
156 157
Convolution Operator.

C
chengduoZH 已提交
158
The convolution operation calculates the output based on the input, filter
C
chengduoZH 已提交
159
and strides, paddings, dilations, groups parameters. The size of each dimension of the
C
chengduoZH 已提交
160
parameters is checked in the infer-shape.
C
chengduoZH 已提交
161
Input(Input) and Output(Output) are in NCHW format. Where N is batch
C
fix doc  
chengduoZH 已提交
162
size, C is the number of channels, H is the height of the feature, and W is
C
chengduoZH 已提交
163 164 165 166 167 168
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
169 170 171 172
The input(X) size and output(Out) size may be different.

Example:
  Input:
C
chengduoZH 已提交
173 174
       Input shape: $(N, C_{in}, H_{in}, W_{in})$
       Filter shape: $(C_{out}, C_{in}, H_f, W_f)$
C
chengduoZH 已提交
175
  Output:
C
chengduoZH 已提交
176 177 178 179 180 181
       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
182
)DOC");
C
chengduoZH 已提交
183 184
}

185
Conv3DOpMaker::Conv3DOpMaker(OpProto* proto, OpAttrChecker* op_checker)
C
chengduoZH 已提交
186 187 188
    : OpProtoAndCheckerMaker(proto, op_checker) {
  AddInput(
      "Input",
C
fix doc  
chengduoZH 已提交
189
      "(Tensor) The input tensor of convolution operator. "
C
chengduoZH 已提交
190
      "The format of input tensor is NCDHW. Where N is batch size, C is the "
C
fix doc  
chengduoZH 已提交
191 192 193
      "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
194
  AddInput("Filter",
C
fix doc  
chengduoZH 已提交
195
           "(Tensor) The filter tensor of convolution operator. "
C
chengduoZH 已提交
196 197
           "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
198 199 200
           "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
201 202
           "input image channels divided by the groups.");
  AddOutput("Output",
C
fix doc  
chengduoZH 已提交
203
            "(Tensor) The output tensor of convolution operator."
C
chengduoZH 已提交
204
            "The format of output tensor is also NCDHW.");
C
chengduoZH 已提交
205 206 207 208
  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
209
      .SetDefault({1, 1, 1});
C
chengduoZH 已提交
210 211 212 213
  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
214 215 216
      .SetDefault({0, 0, 0});
  AddAttr<int>(
      "groups",
C
chengduoZH 已提交
217
      "(int default:1), the groups number of the convolution operator. "
C
fix doc  
chengduoZH 已提交
218 219 220 221
      "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
222
      .SetDefault(1);
C
chengduoZH 已提交
223
  AddAttr<std::vector<int>>("dilations",
C
chengduoZH 已提交
224 225
                            "(vector<int> default:{1, 1, 1}), the "
                            "dilations(d_dilation, h_dilation, w_dilation) of "
C
chengduoZH 已提交
226
                            "convolution operator.")
C
chengduoZH 已提交
227
      .SetDefault({1, 1, 1});
228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246
  AddAttr<bool>(
      "use_cudnn",
      "(bool, default false) Only used in cudnn kernel, need install cudnn")
      .SetDefault(false);
  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
247

C
chengduoZH 已提交
248
  AddComment(R"DOC(
C
fix doc  
chengduoZH 已提交
249 250
Convolution3D Operator.

C
chengduoZH 已提交
251
The convolution operation calculates the output based on the input, filter
C
chengduoZH 已提交
252
and strides, paddings, dilations, groups parameters. The size of each dimension of the
C
chengduoZH 已提交
253
parameters is checked in the infer-shape.
C
chengduoZH 已提交
254
Input(Input) and output(Output) are in NCDHW format, where N is batch
C
fix doc  
chengduoZH 已提交
255
size, C is the number of channels,D is the depth of the feature, H is the height of
C
chengduoZH 已提交
256 257 258 259 260 261
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
262 263 264 265
The input(X) size and output(Out) size may be different.

Example:
  Input:
C
chengduoZH 已提交
266 267
       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
268
  Output:
C
chengduoZH 已提交
269 270 271 272 273 274 275
       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
276 277 278
)DOC");
}

C
chengduoZH 已提交
279 280 281 282 283 284 285 286 287 288 289
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);
  }
}

290 291 292
framework::OpKernelType ConvOpGrad::GetExpectedKernelType(
    const framework::ExecutionContext& ctx) const {
  bool use_cudnn = ctx.Attr<bool>("use_cudnn");
293
  use_cudnn &= platform::is_gpu_place(ctx.GetPlace());
C
chengduoZH 已提交
294 295 296 297 298 299 300
#ifdef PADDLE_WITH_CUDA
  if (platform::is_gpu_place(ctx.GetPlace())) {
    auto& dev_ctx = ctx.template device_context<platform::CUDADeviceContext>();
    use_cudnn &= dev_ctx.cudnn_handle() != nullptr;
  }
#endif

301 302 303 304 305 306 307 308 309 310 311 312 313 314
  framework::LibraryType library_;
  if (use_cudnn) {
    library_ = framework::LibraryType::kCUDNN;
  } else {
    library_ = framework::LibraryType::kPlain;
  }

  std::string data_format = ctx.Attr<std::string>("data_format");
  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
315 316 317 318
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
C
chengduoZH 已提交
319 320
REGISTER_OP(conv2d, ops::ConvOp, ops::Conv2DOpMaker, conv2d_grad,
            ops::ConvOpGrad);
321 322

// depthwise convolution op
323 324
REGISTER_OP(depthwise_conv2d, ops::ConvOp, ops::Conv2DOpMaker,
            depthwise_conv2d_grad, ops::ConvOpGrad);
C
chengduoZH 已提交
325 326 327
REGISTER_OP(conv3d, ops::ConvOp, ops::Conv3DOpMaker, conv3d_grad,
            ops::ConvOpGrad);

328 329
// depthwise conv kernel
// TODO(xingzhaolong): neon kernel for mobile
Z
zlx 已提交
330
REGISTER_OP_CPU_KERNEL(
331
    depthwise_conv2d,
X
xzl 已提交
332 333 334 335
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvKernel<paddle::platform::CPUDeviceContext, double>);

REGISTER_OP_CPU_KERNEL(
336
    depthwise_conv2d_grad,
X
xzl 已提交
337 338
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, float>,
    ops::GemmConvGradKernel<paddle::platform::CPUDeviceContext, double>);
Z
zlx 已提交
339

C
chengduoZH 已提交
340
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
341 342 343 344 345 346
    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
347 348

REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
349 350 351 352 353 354
    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>);