roi_pool_op.cc 6.7 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"
W
wanghaox 已提交
16 17 18 19

namespace paddle {
namespace operators {

W
wanghaox 已提交
20
using Tensor = framework::Tensor;
21
using LoDTensor = framework::LoDTensor;
W
wanghaox 已提交
22

W
wanghaox 已提交
23
class ROIPoolOp : public framework::OperatorWithKernel {
W
wanghaox 已提交
24 25 26 27 28
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("X"),
W
wanghaox 已提交
29 30 31
                   "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
32
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
W
wanghaox 已提交
33
                   "Output(Out) of ROIPoolOp should not be null.");
W
wanghaox 已提交
34
    PADDLE_ENFORCE(ctx->HasOutput("Argmax"),
W
wanghaox 已提交
35
                   "Output(Argmax) of ROIPoolOp should not be null.");
W
wanghaox 已提交
36
    auto input_dims = ctx->GetInputDim("X");
W
wanghaox 已提交
37 38 39 40 41
    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,
42
                   "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
W
wopeizl 已提交
43
                   "given as [[x1, y1, x2, y2], ...].");
W
wanghaox 已提交
44
    PADDLE_ENFORCE(rois_dims[1] == kROISize,
45
                   "ROIs should be a 2-D LoDTensor of shape (num_rois, 4)"
W
wopeizl 已提交
46
                   "given as [[x1, y1, x2, y2], ...].");
W
wanghaox 已提交
47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66

    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);
67
  }
W
wanghaox 已提交
68 69

 protected:
70
  framework::OpKernelType GetExpectedKernelType(
W
wanghaox 已提交
71 72 73 74 75 76 77
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
        ctx.device_context());
  }
};

W
wanghaox 已提交
78
class ROIPoolGradOp : public framework::OperatorWithKernel {
W
wanghaox 已提交
79 80 81 82 83 84 85 86 87 88 89 90
 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:
91
  framework::OpKernelType GetExpectedKernelType(
W
wanghaox 已提交
92 93 94 95 96 97 98
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(
        framework::ToDataType(ctx.Input<framework::Tensor>("X")->type()),
        ctx.device_context());
  }
};

W
wanghaox 已提交
99
class ROIPoolOpMaker : public framework::OpProtoAndCheckerMaker {
W
wanghaox 已提交
100
 public:
Y
Yu Yang 已提交
101
  void Make() override {
W
wanghaox 已提交
102 103
    AddInput("X",
             "(Tensor), "
W
wanghaox 已提交
104 105 106 107 108 109
             "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",
110
             "(LoDTensor), "
W
wanghaox 已提交
111
             "ROIs (Regions of Interest) to pool over. "
112
             "should be a 2-D LoDTensor of shape (num_rois, 4)"
W
wopeizl 已提交
113
             "given as [[x1, y1, x2, y2], ...]. "
W
wanghaox 已提交
114 115 116
             "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
117 118
    AddOutput("Out",
              "(Tensor), "
W
wanghaox 已提交
119 120
              "The output of ROIPoolOp is a 4-D tensor with shape "
              "(num_rois, channels, pooled_h, pooled_w).");
W
wanghaox 已提交
121 122 123 124
    AddOutput("Argmax",
              "(Tensor), "
              "Argmaxes corresponding to indices in X used "
              "for gradient computation. Only output "
P
peizhilin 已提交
125
              "if arg \"is_test\" is false.")
126
        .AsIntermediate();
W
wanghaox 已提交
127
    AddAttr<float>("spatial_scale",
W
wanghaox 已提交
128 129 130 131
                   "(float, default 1.0), "
                   "Multiplicative spatial scale factor "
                   "to translate ROI coords from their input scale "
                   "to the scale used when pooling.")
132
        .SetDefault(1.0);
W
wanghaox 已提交
133
    AddAttr<int>("pooled_height",
W
wanghaox 已提交
134 135
                 "(int, default 1), "
                 "The pooled output height.")
136
        .SetDefault(1);
W
wanghaox 已提交
137
    AddAttr<int>("pooled_width",
W
wanghaox 已提交
138 139
                 "(int, default 1), "
                 "The pooled output width.")
140
        .SetDefault(1);
W
wanghaox 已提交
141
    AddComment(R"DOC(
Y
yi.wu 已提交
142
**ROIPool Operator**
W
wanghaox 已提交
143

Y
yi.wu 已提交
144 145 146 147 148
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
149

Y
yi.wu 已提交
150 151
1. Dividing each region proposal into equal-sized sections with
   the pooled_width and pooled_height
Y
update  
yi.wu 已提交
152

Y
yi.wu 已提交
153
2. Finding the largest value in each section
Y
update  
yi.wu 已提交
154

Y
yi.wu 已提交
155 156
3. Copying these max values to the output buffer

W
wanghaox 已提交
157 158 159 160 161 162 163 164 165 166
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");
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
Y
Yang Yang 已提交
167
REGISTER_OPERATOR(roi_pool, ops::ROIPoolOp, ops::ROIPoolOpMaker,
168 169
                  paddle::framework::DefaultGradOpDescMaker<true>);
REGISTER_OPERATOR(roi_pool_grad, ops::ROIPoolGradOp);
W
wanghaox 已提交
170
REGISTER_OP_CPU_KERNEL(
Q
QI JUN 已提交
171 172 173
    roi_pool,
    ops::CPUROIPoolOpKernel<paddle::platform::CPUDeviceContext, float>,
    ops::CPUROIPoolOpKernel<paddle::platform::CPUDeviceContext, double>);
W
wanghaox 已提交
174 175
REGISTER_OP_CPU_KERNEL(
    roi_pool_grad,
Q
QI JUN 已提交
176
    ops::CPUROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, float>,
J
jerrywgz 已提交
177
    ops::CPUROIPoolGradOpKernel<paddle::platform::CPUDeviceContext, double>);