conv2d_op.cc 4.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14
/* 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. */

武毅 已提交
15
#include "paddle/operators/conv2d_op.h"
16 17 18 19

namespace paddle {
namespace operators {

武毅 已提交
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
void Conv2DOp::InferShape(framework::InferShapeContext* ctx) const {
  PADDLE_ENFORCE(ctx->HasInput("Input"),
                 "Input(Input) of Conv2DOp should not be null.");
  PADDLE_ENFORCE(ctx->HasInput("Filter"),
                 "Input(Filter) of Conv2DOp should not be null.");
  PADDLE_ENFORCE(ctx->HasOutput("Output"),
                 "Output(Output) of Conv2DOp should not be null.");

  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");
  int input_channels = in_dims[1];
  int output_channels = filter_dims[0];

  PADDLE_ENFORCE_EQ(in_dims.size(), 4, "Conv2DOp input should be 4-D.");
  PADDLE_ENFORCE_EQ(filter_dims.size(), 4, "Conv2DOp filter should be 4-D.");
  PADDLE_ENFORCE_EQ(input_channels, filter_dims[1] * groups,
                    "The number of input channels should be equal to filter "
                    "channels * groups.");
  PADDLE_ENFORCE_EQ(
      output_channels % groups, 0,
      "The number of output channels should be divided by groups.");

  auto output_height =
      OutputSize(in_dims[2], filter_dims[2], paddings[0], strides[0]);
  auto output_width =
      OutputSize(in_dims[3], filter_dims[3], paddings[1], strides[1]);
  ctx->SetOutputDim("Output",
                    {in_dims[0], filter_dims[0], output_height, output_width});
51 52
}

武毅 已提交
53 54 55 56 57 58
Conv2DOpMaker::Conv2DOpMaker(framework::OpProto* proto,
                             framework::OpAttrChecker* op_checker)
    : OpProtoAndCheckerMaker(proto, op_checker) {
  AddInput(
      "Input",
      "The input tensor of convolution operator. "
59 60 61
      "The format of input tensor is NCHW, where N is batch size, C is the "
      "number of channels, H is the height of the image, "
      "and W is the width of the image.");
武毅 已提交
62
  AddInput("Filter",
63
           "The filter tensor of convolution operator. "
武毅 已提交
64 65
           "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, "
66 67
           "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 "
武毅 已提交
68 69
           "input image channels divided by the groups.");
  AddOutput("Output",
70
            "The output tensor of convolution operator. "
武毅 已提交
71 72 73 74 75 76 77
            "The format of output tensor is also NCHW.");
  AddAttr<std::vector<int>>("strides", "strides of convolution operator.")
      .SetDefault({1, 1});
  AddAttr<std::vector<int>>("paddings", "paddings of convolution operator.")
      .SetDefault({0, 0});
  AddAttr<int>(
      "groups",
78 79 80 81 82
      "Group size of convolution operator. "
      "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.")
武毅 已提交
83 84
      .SetDefault(1);
  AddComment(R"DOC(
85 86 87 88 89 90
Convolution Operator.

The convolution operation calculates the output based on the input, filter, 
strides, paddings, and groups parameters. The size of each dimension of the
parameters is checked in the infer-shape method.

91
)DOC");
武毅 已提交
92
}
93

武毅 已提交
94 95 96 97 98
void Conv2DOpGrad::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);
99
  }
武毅 已提交
100 101 102 103
  if (ctx->HasOutput(framework::GradVarName("Filter"))) {
    ctx->SetOutputDim(framework::GradVarName("Filter"), filter_dims);
  }
}
104 105 106 107 108

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
109
REGISTER_OP(conv2d, ops::Conv2DOp, ops::Conv2DOpMaker, conv2d_grad,
110 111 112
            ops::Conv2DOpGrad);

REGISTER_OP_CPU_KERNEL(
H
hedaoyuan 已提交
113
    conv2d, ops::GemmConv2DKernel<paddle::platform::CPUPlace, float>);
H
hedaoyuan 已提交
114
REGISTER_OP_CPU_KERNEL(
H
hedaoyuan 已提交
115
    conv2d_grad, ops::GemmConvGrad2DKernel<paddle::platform::CPUPlace, float>);