yolo_box_op.cc 7.3 KB
Newer Older
D
dengkaipeng 已提交
1
/* Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserve.
D
dengkaipeng 已提交
2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25
   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/detection/yolo_box_op.h"
#include "paddle/fluid/framework/op_registry.h"

namespace paddle {
namespace operators {

using framework::Tensor;

class YoloBoxOp : public framework::OperatorWithKernel {
 public:
  using framework::OperatorWithKernel::OperatorWithKernel;
  void InferShape(framework::InferShapeContext* ctx) const override {
    PADDLE_ENFORCE(ctx->HasInput("X"),
                   "Input(X) of YoloBoxOp should not be null.");
26 27
    PADDLE_ENFORCE(ctx->HasInput("ImgSize"),
                   "Input(ImgSize) of YoloBoxOp should not be null.");
D
dengkaipeng 已提交
28 29 30 31 32 33
    PADDLE_ENFORCE(ctx->HasOutput("Boxes"),
                   "Output(Boxes) of YoloBoxOp should not be null.");
    PADDLE_ENFORCE(ctx->HasOutput("Scores"),
                   "Output(Scores) of YoloBoxOp should not be null.");

    auto dim_x = ctx->GetInputDim("X");
34
    auto dim_imgsize = ctx->GetInputDim("ImgSize");
D
dengkaipeng 已提交
35 36 37 38 39 40 41 42 43
    auto anchors = ctx->Attrs().Get<std::vector<int>>("anchors");
    int anchor_num = anchors.size() / 2;
    auto class_num = ctx->Attrs().Get<int>("class_num");

    PADDLE_ENFORCE_EQ(dim_x.size(), 4, "Input(X) should be a 4-D tensor.");
    PADDLE_ENFORCE_EQ(
        dim_x[1], anchor_num * (5 + class_num),
        "Input(X) dim[1] should be equal to (anchor_mask_number * (5 "
        "+ class_num)).");
44 45 46 47 48 49
    PADDLE_ENFORCE_EQ(dim_imgsize.size(), 2,
                      "Input(ImgSize) should be a 2-D tensor.");
    PADDLE_ENFORCE_EQ(
        dim_imgsize[0], dim_x[0],
        "Input(ImgSize) dim[0] and Input(X) dim[0] should be same.");
    PADDLE_ENFORCE_EQ(dim_imgsize[1], 2, "Input(ImgSize) dim[1] should be 2.");
D
dengkaipeng 已提交
50
    PADDLE_ENFORCE_GT(anchors.size(), 0,
D
dengkaipeng 已提交
51
                      "Attr(anchors) length should be greater than 0.");
D
dengkaipeng 已提交
52 53 54
    PADDLE_ENFORCE_EQ(anchors.size() % 2, 0,
                      "Attr(anchors) length should be even integer.");
    PADDLE_ENFORCE_GT(class_num, 0,
D
dengkaipeng 已提交
55
                      "Attr(class_num) should be an integer greater than 0.");
D
dengkaipeng 已提交
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77

    int box_num = dim_x[2] * dim_x[3] * anchor_num;
    std::vector<int64_t> dim_boxes({dim_x[0], box_num, 4});
    ctx->SetOutputDim("Boxes", framework::make_ddim(dim_boxes));

    std::vector<int64_t> dim_scores({dim_x[0], box_num, class_num});
    ctx->SetOutputDim("Scores", framework::make_ddim(dim_scores));
  }

 protected:
  framework::OpKernelType GetExpectedKernelType(
      const framework::ExecutionContext& ctx) const override {
    return framework::OpKernelType(ctx.Input<Tensor>("X")->type(),
                                   ctx.GetPlace());
  }
};

class YoloBoxOpMaker : public framework::OpProtoAndCheckerMaker {
 public:
  void Make() override {
    AddInput("X",
             "The input tensor of YoloBox operator, "
D
dengkaipeng 已提交
78 79 80 81
             "This is a 4-D tensor with shape of [N, C, H, W]. "
             "H and W should be same, and the second dimension(C) stores "
             "box locations, confidence score and classification one-hot "
             "keys of each anchor box. Generally, X should be the output "
D
dengkaipeng 已提交
82
             "of YOLOv3 network.");
83 84
    AddInput("ImgSize",
             "The image size tensor of YoloBox operator, "
D
dengkaipeng 已提交
85 86
             "This is a 2-D tensor with shape of [N, 2]. This tensor holds "
             "height and width of each input image using for resize output "
87
             "box in input image scale.");
D
dengkaipeng 已提交
88 89
    AddOutput("Boxes",
              "The output tensor of detection boxes of YoloBox operator, "
D
dengkaipeng 已提交
90 91
              "This is a 3-D tensor with shape of [N, M, 4], N is the "
              "batch num, M is output box number, and the 3rd dimension "
D
dengkaipeng 已提交
92 93
              "stores [xmin, ymin, xmax, ymax] coordinates of boxes.");
    AddOutput("Scores",
D
dengkaipeng 已提交
94 95 96
              "The output tensor ofdetection boxes scores of YoloBox "
              "operator, This is a 3-D tensor with shape of [N, M, C], "
              "N is the batch num, M is output box number, C is the "
D
dengkaipeng 已提交
97 98 99 100 101 102 103 104 105 106 107 108 109
              "class number.");

    AddAttr<int>("class_num", "The number of classes to predict.");
    AddAttr<std::vector<int>>("anchors",
                              "The anchor width and height, "
                              "it will be parsed pair by pair.")
        .SetDefault(std::vector<int>{});
    AddAttr<int>("downsample_ratio",
                 "The downsample ratio from network input to YoloBox operator "
                 "input, so 32, 16, 8 should be set for the first, second, "
                 "and thrid YoloBox operators.")
        .SetDefault(32);
    AddAttr<float>("conf_thresh",
D
dengkaipeng 已提交
110 111
                   "The confidence scores threshold of detection boxes. "
                   "Boxes with confidence scores under threshold should "
D
dengkaipeng 已提交
112 113 114
                   "be ignored.")
        .SetDefault(0.01);
    AddComment(R"DOC(
D
dengkaipeng 已提交
115
         This operator generate YOLO detection boxes from output of YOLOv3 network.
D
dengkaipeng 已提交
116 117
         
         The output of previous network is in shape [N, C, H, W], while H and W
D
dengkaipeng 已提交
118 119 120 121 122 123 124 125 126 127 128
         should be the same, H and W specify the grid size, each grid point predict 
         given number boxes, this given number, which following will be represented as S,
         is specified by the number of anchors, In the second dimension(the channel
         dimension), C should be equal to S * (class_num + 5), class_num is the object 
         category number of source dataset(such as 80 in coco dataset), so in the 
         second(channel) dimension, apart from 4 box location coordinates x, y, w, h, 
         also includes confidence score of the box and class one-hot key of each anchor 
         box.

         Assume the 4 location coordinates are :math:`t_x, t_y, t_w, t_h`, the box 
         predictions should be as follows:
D
dengkaipeng 已提交
129 130

         $$
D
dengkaipeng 已提交
131
         b_x = \\sigma(t_x) + c_x
D
dengkaipeng 已提交
132 133
         $$
         $$
D
dengkaipeng 已提交
134
         b_y = \\sigma(t_y) + c_y
D
dengkaipeng 已提交
135 136
         $$
         $$
D
dengkaipeng 已提交
137
         b_w = p_w e^{t_w}
D
dengkaipeng 已提交
138 139
         $$
         $$
D
dengkaipeng 已提交
140 141 142
         b_h = p_h e^{t_h}
         $$

D
dengkaipeng 已提交
143 144
         in the equation above, :math:`c_x, c_y` is the left top corner of current grid
         and :math:`p_w, p_h` is specified by anchors.
D
dengkaipeng 已提交
145

D
dengkaipeng 已提交
146 147 148 149 150
         The logistic regression value of the 5rd channel of each anchor prediction boxes
         represent the confidence score of each prediction box, and the logistic
         regression value of the last :attr:`class_num` channels of each anchor prediction 
         boxes represent the classifcation scores. Boxes with confidence scores less than
         :attr:`conf_thresh` should be ignored, and box final scores is the product of 
D
dengkaipeng 已提交
151
         confidence scores and classification scores.
D
dengkaipeng 已提交
152 153 154 155 156 157 158 159 160 161 162 163 164

         )DOC");
  }
};

}  // namespace operators
}  // namespace paddle

namespace ops = paddle::operators;
REGISTER_OPERATOR(yolo_box, ops::YoloBoxOp, ops::YoloBoxOpMaker,
                  paddle::framework::EmptyGradOpMaker);
REGISTER_OP_CPU_KERNEL(yolo_box, ops::YoloBoxKernel<float>,
                       ops::YoloBoxKernel<double>);