roi_pool_op.cc 7.3 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
W
wanghaox 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14

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. */

Y
Yi Wang 已提交
15
#include "paddle/fluid/operators/roi_pool_op.h"
S
sneaxiy 已提交
16
#include <memory>
W
wanghaox 已提交
17 18 19 20

namespace paddle {
namespace operators {

W
wanghaox 已提交
21
using Tensor = framework::Tensor;
22
using LoDTensor = framework::LoDTensor;
W
wanghaox 已提交
23

W
wanghaox 已提交
24
class ROIPoolOp : public framework::OperatorWithKernel {
W
wanghaox 已提交
25 26 27 28 29
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("X"),
W
wanghaox 已提交
30 31 32
                   "Input(X) of ROIPoolOp should not be null.");
    PADDLE_ENFORCE(ctx->HasInput("ROIs"),
                   "Input(ROIs) of ROIPoolOp should not be null.");
W
wanghaox 已提交
33
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
W
wanghaox 已提交
34
                   "Output(Out) of ROIPoolOp should not be null.");
W
wanghaox 已提交
35
    PADDLE_ENFORCE(ctx->HasOutput("Argmax"),
W
wanghaox 已提交
36
                   "Output(Argmax) of ROIPoolOp should not be null.");
W
wanghaox 已提交
37
    auto input_dims = ctx->GetInputDim("X");
W
wanghaox 已提交
38 39 40 41 42
    auto rois_dims = ctx->GetInputDim("ROIs");

    PADDLE_ENFORCE(input_dims.size() == 4,
                   "The format of input tensor is NCHW.");
    PADDLE_ENFORCE(rois_dims.size() == 2,
43
                   "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
W
wopeizl 已提交
44
                   "given as [[x1, y1, x2, y2], ...].");
W
wanghaox 已提交
45
    PADDLE_ENFORCE(rois_dims[1] == kROISize,
46
                   "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
W
wopeizl 已提交
47
                   "given as [[x1, y1, x2, y2], ...].");
W
wanghaox 已提交
48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67

    int pooled_height = ctx->Attrs().Get<int>("pooled_height");
    int pooled_width = ctx->Attrs().Get<int>("pooled_width");
    float spatial_scale = ctx->Attrs().Get<float>("spatial_scale");

    PADDLE_ENFORCE_GT(pooled_height, 0,
                      "The pooled output height must greater than 0");
    PADDLE_ENFORCE_GT(pooled_width, 0,
                      "The pooled output width must greater than 0");
    PADDLE_ENFORCE_GT(spatial_scale, 0.0f,
                      "The spatial scale must greater than 0");

    auto out_dims = input_dims;
    out_dims[0] = rois_dims[0];
    out_dims[1] = input_dims[1];
    out_dims[2] = pooled_height;
    out_dims[3] = pooled_width;

    ctx->SetOutputDim("Out", out_dims);
    ctx->SetOutputDim("Argmax", out_dims);
68
  }
W
wanghaox 已提交
69 70

 protected:
71
  framework::OpKernelType GetExpectedKernelType(
W
wanghaox 已提交
72
      const framework::ExecutionContext& ctx) const override {
73 74 75
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "X"),
        ctx.device_context());
W
wanghaox 已提交
76 77 78
  }
};

W
wanghaox 已提交
79
class ROIPoolGradOp : public framework::OperatorWithKernel {
W
wanghaox 已提交
80 81 82 83 84 85 86 87 88 89 90 91
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
                   "The gradient of Out should not be null.");
    PADDLE_ENFORCE(ctx->HasOutputs(framework::GradVarName("X")),
                   "The gradient of X should not be null.");
    ctx->SetOutputsDim(framework::GradVarName("X"), ctx->GetInputsDim("X"));
  }

 protected:
92
  framework::OpKernelType GetExpectedKernelType(
W
wanghaox 已提交
93
      const framework::ExecutionContext& ctx) const override {
94 95 96
    return framework::OpKernelType(
        OperatorWithKernel::IndicateVarDataType(ctx, "X"),
        ctx.device_context());
W
wanghaox 已提交
97 98 99
  }
};

W
wanghaox 已提交
100
class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker {
W
wanghaox 已提交
101
 public:
Y
Yu Yang 已提交
102
  void Make() override {
W
wanghaox 已提交
103 104
    AddInput("X",
             "(Tensor), "
W
wanghaox 已提交
105 106 107 108 109 110
             "the input of ROIPoolOp. "
             "The format of input tensor is NCHW. Where N is batch size, "
             "C is the number of input channels, "
             "H is the height of the feature, and "
             "W is the width of the feature.");
    AddInput("ROIs",
111
             "(LoDTensor), "
W
wanghaox 已提交
112
             "ROIs (Regions of Interest) to pool over. "
113
             "should be a 2-D LoDTensor of shape (num_rois, 4)"
W
wopeizl 已提交
114
             "given as [[x1, y1, x2, y2], ...]. "
W
wanghaox 已提交
115 116 117
             "Where batch_id is the id of the data, "
             "(x1, y1) is the top left coordinates, and "
             "(x2, y2) is the bottom right coordinates.");
W
wanghaox 已提交
118 119
    AddOutput("Out",
              "(Tensor), "
W
wanghaox 已提交
120 121
              "The output of ROIPoolOp is a 4-D tensor with shape "
              "(num_rois, channels, pooled_h, pooled_w).");
W
wanghaox 已提交
122 123 124 125
    AddOutput("Argmax",
              "(Tensor), "
              "Argmaxes corresponding to indices in X used "
              "for gradient computation. Only output "
P
peizhilin 已提交
126
              "if arg \"is_test\" is false.")
127
        .AsIntermediate();
W
wanghaox 已提交
128
    AddAttr<float>("spatial_scale",
W
wanghaox 已提交
129 130 131 132
                   "(float, default 1.0), "
                   "Multiplicative spatial scale factor "
                   "to translate ROI coords from their input scale "
                   "to the scale used when pooling.")
133
        .SetDefault(1.0);
W
wanghaox 已提交
134
    AddAttr<int>("pooled_height",
W
wanghaox 已提交
135 136
                 "(int, default 1), "
                 "The pooled output height.")
137
        .SetDefault(1);
W
wanghaox 已提交
138
    AddAttr<int>("pooled_width",
W
wanghaox 已提交
139 140
                 "(int, default 1), "
                 "The pooled output width.")
141
        .SetDefault(1);
W
wanghaox 已提交
142
    AddComment(R"DOC(
Y
yi.wu 已提交
143
**ROIPool Operator**
W
wanghaox 已提交
144

Y
yi.wu 已提交
145 146 147 148 149
Region of interest pooling (also known as RoI pooling) is to perform
is to perform max pooling on inputs of nonuniform sizes to obtain
fixed-size feature maps (e.g. 7*7).

The operator has three steps:
Y
yi.wu 已提交
150

Y
yi.wu 已提交
151 152
1. Dividing each region proposal into equal-sized sections with
   the pooled_width and pooled_height
Y
update  
yi.wu 已提交
153

Y
yi.wu 已提交
154
2. Finding the largest value in each section
Y
update  
yi.wu 已提交
155

Y
yi.wu 已提交
156 157
3. Copying these max values to the output buffer

W
wanghaox 已提交
158 159 160 161 162 163
ROI Pooling for Faster-RCNN. The link below is a further introduction: 
https://stackoverflow.com/questions/43430056/what-is-roi-layer-in-fast-rcnn
    )DOC");
  }
};

H
hong 已提交
164 165
template <typename T>
class ROIPoolGradMaker : public framework::SingleGradOpMaker<T> {
S
sneaxiy 已提交
166
 public:
H
hong 已提交
167
  using framework::SingleGradOpMaker<T>::SingleGradOpMaker;
S
sneaxiy 已提交
168 169

 protected:
170
  void Apply(GradOpPtr<T> op) const override {
S
sneaxiy 已提交
171
    op->SetType("roi_pool_grad");
H
hong 已提交
172 173 174 175 176 177
    op->SetInput("X", this->Input("X"));
    op->SetInput("ROIs", this->Input("ROIs"));
    op->SetInput("Argmax", this->Output("Argmax"));
    op->SetInput(framework::GradVarName("Out"), this->OutputGrad("Out"));
    op->SetOutput(framework::GradVarName("X"), this->InputGrad("X"));
    op->SetAttrMap(this->Attrs());
S
sneaxiy 已提交
178 179 180
  }
};

W
wanghaox 已提交
181 182 183 184
}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
Y
Yang Yang 已提交
185
REGISTER_OPERATOR(roi_pool, ops::ROIPoolOp, ops::ROIPoolOpMaker,
H
hong 已提交
186 187
                  ops::ROIPoolGradMaker<paddle::framework::OpDesc>,
                  ops::ROIPoolGradMaker<paddle::imperative::OpBase>);
188
REGISTER_OPERATOR(roi_pool_grad, ops::ROIPoolGradOp);
W
wanghaox 已提交
189
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
190 191 192
    roi_pool,
    ops::CPUROIPoolOpKernel<paddle::platform::CPUDeviceContext, float>,
    ops::CPUROIPoolOpKernel<paddle::platform::CPUDeviceContext, double>);
W
wanghaox 已提交
193 194
REGISTER_OP_CPU_KERNEL(
    roi_pool_grad,
Q
QI JUN 已提交
195
    ops::CPUROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, float>,
J
jerrywgz 已提交
196
    ops::CPUROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, double>);