prior_box_op.h 6.9 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 15

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. */

#pragma once
Y
Yi Wang 已提交
16 17 18
#include "paddle/fluid/framework/op_registry.h"
#include "paddle/fluid/operators/math/math_function.h"
#include "paddle/fluid/platform/transform.h"
W
wanghaox 已提交
19 20 21 22

namespace paddle {
namespace operators {

W
wanghaox 已提交
23 24 25
inline void ExpandAspectRatios(const std::vector<float>& input_aspect_ratior,
                               bool flip,
                               std::vector<float>& output_aspect_ratior) {
26
  constexpr float epsilon = 1e-6;
W
wanghaox 已提交
27
  output_aspect_ratior.clear();
C
chengduoZH 已提交
28
  output_aspect_ratior.push_back(1.0f);
W
wanghaox 已提交
29 30 31 32
  for (size_t i = 0; i < input_aspect_ratior.size(); ++i) {
    float ar = input_aspect_ratior[i];
    bool already_exist = false;
    for (size_t j = 0; j < output_aspect_ratior.size(); ++j) {
33
      if (fabs(ar - output_aspect_ratior[j]) < epsilon) {
W
wanghaox 已提交
34 35 36 37 38 39 40
        already_exist = true;
        break;
      }
    }
    if (!already_exist) {
      output_aspect_ratior.push_back(ar);
      if (flip) {
C
chengduoZH 已提交
41
        output_aspect_ratior.push_back(1.0f / ar);
W
wanghaox 已提交
42 43 44 45 46
      }
    }
  }
}

W
wanghaox 已提交
47 48
template <typename T>
struct ClipFunctor {
49
  HOSTDEVICE inline T operator()(T in) const {
W
wanghaox 已提交
50 51 52 53
    return std::min<T>(std::max<T>(in, 0.), 1.);
  }
};

W
wanghaox 已提交
54 55 56 57 58 59
template <typename Place, typename T>
class PriorBoxOpKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* input = ctx.Input<paddle::framework::Tensor>("Input");
    auto* image = ctx.Input<paddle::framework::Tensor>("Image");
W
wanghaox 已提交
60 61
    auto* boxes = ctx.Output<paddle::framework::Tensor>("Boxes");
    auto* vars = ctx.Output<paddle::framework::Tensor>("Variances");
W
wanghaox 已提交
62

C
chengduoZH 已提交
63 64
    auto min_sizes = ctx.Attr<std::vector<float>>("min_sizes");
    auto max_sizes = ctx.Attr<std::vector<float>>("max_sizes");
W
wanghaox 已提交
65 66 67 68 69 70
    auto input_aspect_ratio = ctx.Attr<std::vector<float>>("aspect_ratios");
    auto variances = ctx.Attr<std::vector<float>>("variances");
    auto flip = ctx.Attr<bool>("flip");
    auto clip = ctx.Attr<bool>("clip");

    std::vector<float> aspect_ratios;
W
wanghaox 已提交
71
    ExpandAspectRatios(input_aspect_ratio, flip, aspect_ratios);
W
wanghaox 已提交
72

W
wanghaox 已提交
73 74 75
    T step_w = static_cast<T>(ctx.Attr<float>("step_w"));
    T step_h = static_cast<T>(ctx.Attr<float>("step_h"));
    T offset = static_cast<T>(ctx.Attr<float>("offset"));
W
wanghaox 已提交
76

W
wanghaox 已提交
77 78
    auto img_width = image->dims()[3];
    auto img_height = image->dims()[2];
W
wanghaox 已提交
79

W
wanghaox 已提交
80 81
    auto feature_width = input->dims()[3];
    auto feature_height = input->dims()[2];
W
wanghaox 已提交
82

W
wanghaox 已提交
83
    T step_width, step_height;
W
wanghaox 已提交
84
    if (step_w == 0 || step_h == 0) {
W
wanghaox 已提交
85 86
      step_width = static_cast<T>(img_width) / feature_width;
      step_height = static_cast<T>(img_height) / feature_height;
W
wanghaox 已提交
87 88 89 90 91 92 93 94 95 96
    } else {
      step_width = step_w;
      step_height = step_h;
    }

    int num_priors = aspect_ratios.size() * min_sizes.size();
    if (max_sizes.size() > 0) {
      num_priors += max_sizes.size();
    }

W
wanghaox 已提交
97 98
    boxes->mutable_data<T>(ctx.GetPlace());
    vars->mutable_data<T>(ctx.GetPlace());
W
wanghaox 已提交
99

100 101 102
    T inv_img_width = 1.0 / img_width;
    T inv_img_height = 1.0 / img_height;

W
wanghaox 已提交
103
    auto e_boxes = framework::EigenTensor<T, 4>::From(*boxes);
W
wanghaox 已提交
104 105
    for (int h = 0; h < feature_height; ++h) {
      for (int w = 0; w < feature_width; ++w) {
W
wanghaox 已提交
106 107 108
        T center_x = (w + offset) * step_width;
        T center_y = (h + offset) * step_height;
        T box_width, box_height;
109
        int idx = 0;
W
wanghaox 已提交
110
        for (size_t s = 0; s < min_sizes.size(); ++s) {
C
chengduoZH 已提交
111
          auto min_size = min_sizes[s];
W
wanghaox 已提交
112 113 114
          // first prior: aspect_ratio = 1, size = min_size
          box_width = box_height = min_size;
          // xmin
115
          e_boxes(h, w, idx, 0) = (center_x - box_width * 0.5) * inv_img_width;
W
wanghaox 已提交
116
          // ymin
117 118
          e_boxes(h, w, idx, 1) =
              (center_y - box_height * 0.5) * inv_img_height;
W
wanghaox 已提交
119
          // xmax
120
          e_boxes(h, w, idx, 2) = (center_x + box_width * 0.5) * inv_img_width;
W
wanghaox 已提交
121
          // ymax
122 123
          e_boxes(h, w, idx, 3) =
              (center_y + box_height * 0.5) * inv_img_height;
W
wanghaox 已提交
124

125
          idx++;
W
wanghaox 已提交
126
          if (max_sizes.size() > 0) {
C
chengduoZH 已提交
127
            auto max_size = max_sizes[s];
W
wanghaox 已提交
128 129 130 131
            // second prior: aspect_ratio = 1,
            // size = sqrt(min_size * max_size)
            box_width = box_height = sqrt(min_size * max_size);
            // xmin
132 133
            e_boxes(h, w, idx, 0) =
                (center_x - box_width * 0.5) * inv_img_width;
W
wanghaox 已提交
134
            // ymin
135 136
            e_boxes(h, w, idx, 1) =
                (center_y - box_height * 0.5) * inv_img_height;
W
wanghaox 已提交
137
            // xmax
138 139
            e_boxes(h, w, idx, 2) =
                (center_x + box_width * 0.5) * inv_img_width;
W
wanghaox 已提交
140
            // ymax
141 142
            e_boxes(h, w, idx, 3) =
                (center_y + box_height * 0.5) * inv_img_height;
143
            idx++;
W
wanghaox 已提交
144 145 146 147 148 149 150 151 152 153 154
          }

          // rest of priors
          for (size_t r = 0; r < aspect_ratios.size(); ++r) {
            float ar = aspect_ratios[r];
            if (fabs(ar - 1.) < 1e-6) {
              continue;
            }
            box_width = min_size * sqrt(ar);
            box_height = min_size / sqrt(ar);
            // xmin
155 156
            e_boxes(h, w, idx, 0) =
                (center_x - box_width * 0.5) * inv_img_width;
W
wanghaox 已提交
157
            // ymin
158 159
            e_boxes(h, w, idx, 1) =
                (center_y - box_height * 0.5) * inv_img_height;
W
wanghaox 已提交
160
            // xmax
161 162
            e_boxes(h, w, idx, 2) =
                (center_x + box_width * 0.5) * inv_img_width;
W
wanghaox 已提交
163
            // ymax
164 165
            e_boxes(h, w, idx, 3) =
                (center_y + box_height * 0.5) * inv_img_height;
166
            idx++;
W
wanghaox 已提交
167 168 169 170 171 172
          }
        }
      }
    }

    if (clip) {
W
wanghaox 已提交
173 174 175 176 177
      platform::Transform<platform::CPUDeviceContext> trans;
      ClipFunctor<T> clip_func;
      trans(ctx.template device_context<platform::CPUDeviceContext>(),
            boxes->data<T>(), boxes->data<T>() + boxes->numel(),
            boxes->data<T>(), clip_func);
W
wanghaox 已提交
178
    }
W
wanghaox 已提交
179

W
wanghaox 已提交
180 181 182 183 184
    framework::Tensor var_t;
    var_t.mutable_data<T>(
        framework::make_ddim({1, static_cast<int>(variances.size())}),
        ctx.GetPlace());
    auto var_et = framework::EigenTensor<T, 2>::From(var_t);
W
wanghaox 已提交
185
    for (size_t i = 0; i < variances.size(); ++i) {
W
wanghaox 已提交
186
      var_et(0, i) = variances[i];
W
wanghaox 已提交
187
    }
W
wanghaox 已提交
188

W
wanghaox 已提交
189
    int box_num = feature_height * feature_width * num_priors;
W
wanghaox 已提交
190 191 192 193 194 195 196
    auto var_dim = vars->dims();
    vars->Resize({box_num, static_cast<int>(variances.size())});

    auto e_vars = framework::EigenMatrix<T, Eigen::RowMajor>::From(*vars);
    e_vars = var_et.broadcast(Eigen::DSizes<int, 2>(box_num, 1));

    vars->Resize(var_dim);
W
wanghaox 已提交
197
  }
W
wanghaox 已提交
198
};  // namespace operators
W
wanghaox 已提交
199 200 201

}  // namespace operators
}  // namespace paddle