im2sequence_op.cc 5.4 KB
Newer Older
1
/* Copyright (c) 2016 PaddlePaddle Authors. All Rights Reserved.
G
gongweibao 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/im2sequence_op.h"
16
#include <vector>
G
gongweibao 已提交
17 18 19 20

namespace paddle {
namespace operators {

21
class Im2SequenceOp : public framework::OperatorWithKernel {
G
gongweibao 已提交
22 23 24 25 26
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
  void InferShape(framework::InferShapeContext* ctx) const override {
G
gongweibao 已提交
27
    PADDLE_ENFORCE(ctx->HasInput("X"),
28
                   "Input(X) of Im2SequenceOp should not be null.");
G
gongweibao 已提交
29
    PADDLE_ENFORCE(ctx->HasOutput("Out"),
30
                   "Output(Out) of Im2SequenceOp op should not be null.");
G
gongweibao 已提交
31

G
gongweibao 已提交
32
    auto in_dim = ctx->GetInputDim("X");
G
gongweibao 已提交
33
    PADDLE_ENFORCE_EQ(in_dim.size(), 4,
W
wanghaoshuang 已提交
34
                      "Input(X) format must be 4D tensor, eg., NCHW.");
G
gongweibao 已提交
35

W
wanghaoshuang 已提交
36 37 38
    auto kernels = ctx->Attrs().Get<std::vector<int>>("kernels");
    auto strides = ctx->Attrs().Get<std::vector<int>>("strides");
    auto paddings = ctx->Attrs().Get<std::vector<int>>("paddings");
G
gongweibao 已提交
39

W
wanghaoshuang 已提交
40 41
    int batch_size = in_dim[0];
    int img_channels = in_dim[1];
G
gongweibao 已提交
42 43
    int img_height = in_dim[2];
    int img_width = in_dim[3];
G
gongweibao 已提交
44

Y
Yang Yang 已提交
45 46 47 48
    int output_height = Im2SeqOutputSize(img_height, kernels[0], paddings[0],
                                         paddings[2], strides[0]);
    int output_width = Im2SeqOutputSize(img_width, kernels[1], paddings[1],
                                        paddings[3], strides[1]);
G
gongweibao 已提交
49

W
wanghaoshuang 已提交
50
    ctx->SetOutputDim("Out", {batch_size * output_height * output_width,
W
wanghaoshuang 已提交
51
                              img_channels * kernels[0] * kernels[1]});
G
gongweibao 已提交
52 53 54
  }
};

55
class Im2SequenceOpMaker : public framework::OpProtoAndCheckerMaker {
G
gongweibao 已提交
56
 public:
57
  Im2SequenceOpMaker(OpProto* proto, OpAttrChecker* op_checker)
G
gongweibao 已提交
58
      : OpProtoAndCheckerMaker(proto, op_checker) {
W
wanghaoshuang 已提交
59
    AddInput("X",
W
wanghaoshuang 已提交
60
             "(Tensor) The input tensor has NCHW format."
W
wanghaoshuang 已提交
61 62 63 64
             "N: batch size"
             "C: channels"
             "H: height"
             "W: width");
W
wanghaoshuang 已提交
65
    AddOutput("Out", "(LodTensor) The output data of im2sequence op,");
W
wanghaoshuang 已提交
66 67
    AddAttr<std::vector<int>>("kernels",
                              "(vector<int>), the "
W
wanghaoshuang 已提交
68 69 70 71 72
                              "kernels(kernel_height, kernel_width)");
    AddAttr<std::vector<int>>("strides",
                              "(vector<int> default:{1, 1}), the "
                              "strides(h_stride, w_stride)")
        .SetDefault({1, 1});
W
wanghaoshuang 已提交
73 74 75 76
    AddAttr<std::vector<int>>("paddings",
                              "(vector<int> default:{0, 0, 0, 0}), the "
                              "paddings(up_pad, left_pad, down_pad, right_pad)")
        .SetDefault({0, 0, 0, 0});
G
gongweibao 已提交
77
    AddComment(R"DOC(
W
wanghaoshuang 已提交
78 79 80 81
This op uses kernels to scan images and converts these images to sequences.
After expanding, The number of time steps are output_height * output_width
and the dimension of each time step is kernel_height * kernel_width * channels,
in which:
W
wanghaoshuang 已提交
82 83

output_height =
W
wanghaoshuang 已提交
84
    1 + (padding_height + padding_down + img_height - kernel_height + stride_height - 1) /
W
wanghaoshuang 已提交
85 86
            stride_height;
output_width =
W
wanghaoshuang 已提交
87
    1 + (padding_left + padding+right + img_width - kernel_width + stride_width - 1) /
W
wanghaoshuang 已提交
88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112
            stride_width;

This op can be used after convolution neural network, and before recurrent neural network.

Given:

x = [[[[ 6.  2.  1.]
       [ 8.  3.  5.]
       [ 0.  2.  6.]]

      [[ 2.  4.  4.]
       [ 6.  3.  0.]
       [ 6.  4.  7.]]]

     [[[ 6.  7.  1.]
       [ 5.  7.  9.]
       [ 2.  4.  8.]]

      [[ 1.  2.  1.]
       [ 1.  3.  5.]
       [ 9.  0.  8.]]]]
x.dims = {2, 2, 3, 3}

And:

W
wanghaoshuang 已提交
113 114 115
kernels = [2, 2]
strides = [1, 1]
paddings = [0, 0, 0, 0]
W
wanghaoshuang 已提交
116 117 118 119 120 121 122 123 124 125 126 127 128 129

Then:

output.data = [[ 6.  2.  8.  3.  2.  4.  6.  3.]
               [ 2.  1.  3.  5.  4.  4.  3.  0.]
               [ 8.  3.  0.  2.  6.  3.  6.  4.]
               [ 3.  5.  2.  6.  3.  0.  4.  7.]
               [ 6.  7.  5.  7.  1.  2.  1.  3.]
               [ 7.  1.  7.  9.  2.  1.  3.  5.]
               [ 5.  7.  2.  4.  1.  3.  9.  0.]
               [ 7.  9.  4.  8.  3.  5.  0.  8.]]
output.dims = {8, 9}
output.lod = [[0, 4, 8]]

G
gongweibao 已提交
130 131 132 133
)DOC");
  }
};

134
class Im2SequenceGradOp : public framework::OperatorWithKernel {
G
gongweibao 已提交
135 136 137 138
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;

 protected:
G
add gpu  
gongweibao 已提交
139 140 141
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("X"), "Input(X) should not be null");
    PADDLE_ENFORCE(ctx->HasInput(framework::GradVarName("Out")),
G
gongweibao 已提交
142 143
                   "Input(Out@GRAD) shouldn't be null.");
    ctx->SetOutputDim(framework::GradVarName("X"), ctx->GetInputDim("X"));
G
add gpu  
gongweibao 已提交
144
  }
G
gongweibao 已提交
145 146 147 148 149 150
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
151 152
REGISTER_OP(im2sequence, ops::Im2SequenceOp, ops::Im2SequenceOpMaker,
            im2sequence_grad, ops::Im2SequenceGradOp);
G
gongweibao 已提交
153
REGISTER_OP_CPU_KERNEL(
154 155
    im2sequence,
    ops::Im2SequenceKernel<paddle::platform::CPUDeviceContext, float>);
G
gongweibao 已提交
156
REGISTER_OP_CPU_KERNEL(
157 158
    im2sequence_grad,
    ops::Im2SequenceGradKernel<paddle::platform::CPUDeviceContext, float>);