unfold_op.cc 7.6 KB
Newer Older
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
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
 *
 * 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. */

#include "paddle/fluid/operators/unfold_op.h"

namespace paddle {
namespace operators {

class UnfoldOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X",
             "Tensor, "
             "the input of unfold op. "
             "The format of X is [N, C_in, H, W], "
             "where N is the batch size, C_in is the input channels, "
             "H is the height and W is the width");
    AddOutput(
        "Y",
        "Tensor, "
        "the output of unfold op. "
        "The format of Y is [N, C_in*filter_height*filter_width, "
        "output_height*output_width], where N is the batch size, "
        "C_in is the input channels of X, filter_height and filter_width is "
        "height and width of the filtering kernel, output_height and "
        "output_width "
        "is the calculated height and width of output feature map.");
    AddAttr<std::vector<int>>(
        "kernel_sizes",
        "vector<int>, the kernel sizes of the convolution operator.");
    AddAttr<std::vector<int>>(
        "strides", "vector<int>, the strides of the convolution operator.");
    AddAttr<std::vector<int>>(
        "paddings",
        "vector<int>, the paddings applied to pad the feature map.");
    AddAttr<std::vector<int>>(
        "dilations", "vector<int>, the dilations of the convolution operator.");
    AddComment(R"DOC(
**Unfold Operator**

This Operator is used to extract sliding local blocks from a batched input tensor, also known
as im2col when operated on batched 2D image tensor. For each block under the convolution filter,
T
tianshuo78520a 已提交
54
all element will be rearranged as a column. While the convolution filter sliding over the input
55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78
feature map, a series of such columns will be formed. 
    )DOC");
  }
};

class UnfoldOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("X"),
                   "Input(X) of UnfoldOp should not be null");
    PADDLE_ENFORCE(ctx->HasOutput("Y"),
                   "Output(Y) of UnfoldOp should not be null");
    auto in_dims = ctx->GetInputDim("X");
    std::vector<int> kernel_sizes =
        ctx->Attrs().Get<std::vector<int>>("kernel_sizes");
    std::vector<int> strides = ctx->Attrs().Get<std::vector<int>>("strides");
    std::vector<int> paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
    std::vector<int> dilations =
        ctx->Attrs().Get<std::vector<int>>("dilations");

    // Only [N, C, H, W] input supported now
    PADDLE_ENFORCE(
        in_dims.size() == 4,
T
tianshuo78520a 已提交
79
        "Input should be 4-D tensor of format [N, C, H, W], but get %u",
80 81 82 83 84 85 86 87 88
        in_dims.size());
    PADDLE_ENFORCE(
        in_dims.size() - kernel_sizes.size() == 2U,
        "The dims of X should be larger than that of kernel_sizes "
        "by a number of 2, due to the batch size and input channel dim. "
        "But recieved dims(X:%u) - dims(kernel_sizes:%u) != 2",
        in_dims.size(), kernel_sizes.size());
    PADDLE_ENFORCE_EQ(
        strides.size(), kernel_sizes.size(),
T
tianshuo78520a 已提交
89
        "The dims of strides should be the same with that of kernel_sizes. "
90 91 92 93 94 95 96 97 98
        "But recieved dims(strides: %u) != dims(kernel_sizes: %u).",
        strides.size(), kernel_sizes.size());
    PADDLE_ENFORCE_EQ(
        paddings.size(), 2 * strides.size(),
        "The dims of paddings should be 2 times of that of strides. "
        "But recieved dims(paddings: %u) != 2*dims(strides: %u).",
        paddings.size(), strides.size());
    PADDLE_ENFORCE_EQ(
        strides.size(), dilations.size(),
T
tianshuo78520a 已提交
99
        "The dims of strides should be the same with that of dilations. "
100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122
        "But recieved dims(strides: %u) != dims(dilations: %u).",
        strides.size(), dilations.size());

    std::vector<int> out_dims;
    out_dims.push_back(in_dims[0]);

    int output_channels = in_dims[1] * kernel_sizes[0] * kernel_sizes[1];
    out_dims.push_back(output_channels);

    int output_height =
        CalcOutputSize(in_dims[2], kernel_sizes[0], dilations[0], paddings[0],
                       paddings[2], strides[0]);
    int output_width = CalcOutputSize(in_dims[3], kernel_sizes[1], dilations[1],
                                      paddings[1], paddings[3], strides[1]);
    int output_col_length = output_height * output_width;
    out_dims.push_back(output_col_length);

    ctx->SetOutputDim("Y", framework::make_ddim(out_dims));
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
123 124 125
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "X"),
        ctx.device_context());
126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144
  }
};

class UnfoldGradOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Y")),
                   "The gradient of Y should not be null");
    PADDLE_ENFORCE(ctx->HasInput("X"), "The input X should not be null");
    PADDLE_ENFORCE(ctx->HasOutput(framework::GradVarName("X")),
                   "The gradient of X should not be null");
    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
145 146 147
    return framework::OpKernelType(OperatorWithKernel::IndicateVarDataType(
                                       ctx, framework::GradVarName("Y")),
                                   ctx.device_context());
148 149 150
  }
};

H
hong 已提交
151 152
template <typename T>
class UnfoldGradMaker : public framework::SingleGradOpMaker<T> {
153
 public:
H
hong 已提交
154
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
155 156

 protected:
H
hong 已提交
157 158
  std::unique_ptr<T> Apply() const override {
    std::unique_ptr<T> op(new T());
159
    op->SetType("unfold_grad");
H
hong 已提交
160 161 162 163
    op->SetInput(framework::GradVarName("Y"), this->OutputGrad("Y"));
    op->SetInput("X", this->Input("X"));
    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    op->SetAttrMap(this->Attrs());
164 165 166 167 168 169 170 171 172 173 174 175
    return op;
  }
};

DECLARE_NO_NEED_BUFFER_VARS_INFERENCE(UnfoldGradOpNoNeedBufferVarsInference,
                                      "X");

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OPERATOR(unfold, ops::UnfoldOp, ops::UnfoldOpMaker,
H
hong 已提交
176 177
                  ops::UnfoldGradMaker<paddle::framework::OpDesc>,
                  ops::UnfoldGradMaker<paddle::imperative::OpBase>);
178 179 180 181 182 183 184 185 186 187
REGISTER_OPERATOR(unfold_grad, ops::UnfoldGradOp,
                  ops::UnfoldGradOpNoNeedBufferVarsInference);

REGISTER_OP_CPU_KERNEL(
    unfold, ops::UnfoldOpKernel<paddle::platform::CPUDeviceContext, float>,
    ops::UnfoldOpKernel<paddle::platform::CPUDeviceContext, double>);
REGISTER_OP_CPU_KERNEL(
    unfold_grad,
    ops::UnfoldGradOpKernel<paddle::platform::CPUDeviceContext, float>,
    ops::UnfoldGradOpKernel<paddle::platform::CPUDeviceContext, double>);