prior_box.cc 13.9 KB
Newer Older
Y
Yan Chunwei 已提交
1 2 3 4 5 6 7 8 9 10 11 12 13 14
// Copyright (c) 2019 PaddlePaddle Authors. All Rights Reserved.
//
// 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.

15
#include "lite/backends/arm/math/prior_box.h"
Y
Yan Chunwei 已提交
16
#include <algorithm>
17
#include "lite/backends/arm/math/funcs.h"
Y
Yan Chunwei 已提交
18 19 20 21 22 23

namespace paddle {
namespace lite {
namespace arm {
namespace math {

24
const int MALLOC_ALIGN = 16;
Y
Yan Chunwei 已提交
25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53

void* fast_malloc(size_t size) {
  size_t offset = sizeof(void*) + MALLOC_ALIGN - 1;
  char* p = static_cast<char*>(malloc(offset + size));

  if (!p) {
    return nullptr;
  }

  void* r = reinterpret_cast<void*>(reinterpret_cast<size_t>(p + offset) &
                                    (~(MALLOC_ALIGN - 1)));
  static_cast<void**>(r)[-1] = p;
  memset(r, 0, size);
  return r;
}

void fast_free(void* ptr) {
  if (ptr) {
    free(static_cast<void**>(ptr)[-1]);
  }
}

void density_prior_box(const lite::Tensor* input,
                       const lite::Tensor* image,
                       lite::Tensor** boxes,
                       lite::Tensor** variances,
                       const std::vector<float>& min_size_,
                       const std::vector<float>& fixed_size_,
                       const std::vector<float>& fixed_ratio_,
T
TianXiaogang 已提交
54
                       const std::vector<int>& density_size_,
Y
Yan Chunwei 已提交
55 56 57 58 59 60 61 62 63 64 65
                       const std::vector<float>& max_size_,
                       const std::vector<float>& aspect_ratio_,
                       const std::vector<float>& variance_,
                       int img_w_,
                       int img_h_,
                       float step_w_,
                       float step_h_,
                       float offset_,
                       int prior_num_,
                       bool is_flip_,
                       bool is_clip_,
66 67
                       const std::vector<std::string>& order_,
                       bool min_max_aspect_ratios_order) {
Y
Yan Chunwei 已提交
68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287
  // compute output shape
  int win1 = input->dims()[3];
  int hin1 = input->dims()[2];
  DDim shape_out({hin1, win1, prior_num_, 4});
  (*boxes)->Resize(shape_out);
  (*variances)->Resize(shape_out);

  float* _cpu_data = (*boxes)->mutable_data<float>();
  float* _variance_data = (*variances)->mutable_data<float>();

  const int width = win1;
  const int height = hin1;
  int img_width = img_w_;
  int img_height = img_h_;
  if (img_width == 0 || img_height == 0) {
    img_width = image->dims()[3];
    img_height = image->dims()[2];
  }
  float step_w = step_w_;
  float step_h = step_h_;
  if (step_w == 0 || step_h == 0) {
    step_w = static_cast<float>(img_width) / width;
    step_h = static_cast<float>(img_height) / height;
  }
  float offset = offset_;
  int step_average = static_cast<int>((step_w + step_h) * 0.5);  // add
  int channel_size = height * width * prior_num_ * 4;
  int idx = 0;
  for (int h = 0; h < height; ++h) {
    for (int w = 0; w < width; ++w) {
      float center_x = (w + offset) * step_w;
      float center_y = (h + offset) * step_h;
      float box_width;
      float box_height;
      if (fixed_size_.size() > 0) {
        // add
        for (int s = 0; s < fixed_size_.size(); ++s) {
          int fixed_size = fixed_size_[s];
          int com_idx = 0;
          box_width = fixed_size;
          box_height = fixed_size;

          if (fixed_ratio_.size() > 0) {
            for (int r = 0; r < fixed_ratio_.size(); ++r) {
              float ar = fixed_ratio_[r];
              int density = density_size_[s];
              int shift = step_average / density;
              float box_width_ratio = fixed_size_[s] * sqrt(ar);
              float box_height_ratio = fixed_size_[s] / sqrt(ar);

              for (int p = 0; p < density; ++p) {
                for (int c = 0; c < density; ++c) {
                  float center_x_temp =
                      center_x - step_average / 2.0f + shift / 2.f + c * shift;
                  float center_y_temp =
                      center_y - step_average / 2.0f + shift / 2.f + p * shift;
                  // xmin
                  _cpu_data[idx++] =
                      (center_x_temp - box_width_ratio / 2.f) / img_width >= 0
                          ? (center_x_temp - box_width_ratio / 2.f) / img_width
                          : 0;
                  // ymin
                  _cpu_data[idx++] =
                      (center_y_temp - box_height_ratio / 2.f) / img_height >= 0
                          ? (center_y_temp - box_height_ratio / 2.f) /
                                img_height
                          : 0;
                  // xmax
                  _cpu_data[idx++] =
                      (center_x_temp + box_width_ratio / 2.f) / img_width <= 1
                          ? (center_x_temp + box_width_ratio / 2.f) / img_width
                          : 1;
                  // ymax
                  _cpu_data[idx++] =
                      (center_y_temp + box_height_ratio / 2.f) / img_height <= 1
                          ? (center_y_temp + box_height_ratio / 2.f) /
                                img_height
                          : 1;
                }
              }
            }
          } else {
            // this code for density anchor box
            if (density_size_.size() > 0) {
              CHECK_EQ(fixed_size_.size(), density_size_.size())
                  << "fixed_size_ should be same with density_size_";
              int density = density_size_[s];
              int shift = fixed_size_[s] / density;

              for (int r = 0; r < density; ++r) {
                for (int c = 0; c < density; ++c) {
                  float center_x_temp =
                      center_x - fixed_size / 2.f + shift / 2.f + c * shift;
                  float center_y_temp =
                      center_y - fixed_size / 2.f + shift / 2.f + r * shift;
                  // xmin
                  _cpu_data[idx++] =
                      (center_x_temp - box_width / 2.f) / img_width >= 0
                          ? (center_x_temp - box_width / 2.f) / img_width
                          : 0;
                  // ymin
                  _cpu_data[idx++] =
                      (center_y_temp - box_height / 2.f) / img_height >= 0
                          ? (center_y_temp - box_height / 2.f) / img_height
                          : 0;
                  // xmax
                  _cpu_data[idx++] =
                      (center_x_temp + box_width / 2.f) / img_width <= 1
                          ? (center_x_temp + box_width / 2.f) / img_width
                          : 1;
                  // ymax
                  _cpu_data[idx++] =
                      (center_y_temp + box_height / 2.f) / img_height <= 1
                          ? (center_y_temp + box_height / 2.f) / img_height
                          : 1;
                }
              }
            }

            // rest of priors: will never come here!!!
            for (int r = 0; r < aspect_ratio_.size(); ++r) {
              float ar = aspect_ratio_[r];

              if (fabs(ar - 1.) < 1e-6) {
                continue;
              }

              int density = density_size_[s];
              int shift = fixed_size_[s] / density;
              float box_width_ratio = fixed_size_[s] * sqrt(ar);
              float box_height_ratio = fixed_size_[s] / sqrt(ar);

              for (int p = 0; p < density; ++p) {
                for (int c = 0; c < density; ++c) {
                  float center_x_temp =
                      center_x - fixed_size / 2.f + shift / 2.f + c * shift;
                  float center_y_temp =
                      center_y - fixed_size / 2.f + shift / 2.f + p * shift;
                  // xmin
                  _cpu_data[idx++] =
                      (center_x_temp - box_width_ratio / 2.f) / img_width >= 0
                          ? (center_x_temp - box_width_ratio / 2.f) / img_width
                          : 0;
                  // ymin
                  _cpu_data[idx++] =
                      (center_y_temp - box_height_ratio / 2.f) / img_height >= 0
                          ? (center_y_temp - box_height_ratio / 2.f) /
                                img_height
                          : 0;
                  // xmax
                  _cpu_data[idx++] =
                      (center_x_temp + box_width_ratio / 2.f) / img_width <= 1
                          ? (center_x_temp + box_width_ratio / 2.f) / img_width
                          : 1;
                  // ymax
                  _cpu_data[idx++] =
                      (center_y_temp + box_height_ratio / 2.f) / img_height <= 1
                          ? (center_y_temp + box_height_ratio / 2.f) /
                                img_height
                          : 1;
                }
              }
            }
          }
        }
      } else {
        float* min_buf =
            reinterpret_cast<float*>(fast_malloc(sizeof(float) * 4));
        float* max_buf =
            reinterpret_cast<float*>(fast_malloc(sizeof(float) * 4));
        float* com_buf = reinterpret_cast<float*>(
            fast_malloc(sizeof(float) * aspect_ratio_.size() * 4));

        for (int s = 0; s < min_size_.size(); ++s) {
          int min_idx = 0;
          int max_idx = 0;
          int com_idx = 0;
          int min_size = min_size_[s];
          // first prior: aspect_ratio = 1, size = min_size
          box_width = box_height = min_size;
          //! xmin
          min_buf[min_idx++] = (center_x - box_width / 2.f) / img_width;
          //! ymin
          min_buf[min_idx++] = (center_y - box_height / 2.f) / img_height;
          //! xmax
          min_buf[min_idx++] = (center_x + box_width / 2.f) / img_width;
          //! ymax
          min_buf[min_idx++] = (center_y + box_height / 2.f) / img_height;

          if (max_size_.size() > 0) {
            int max_size = max_size_[s];
            //! second prior: aspect_ratio = 1, size = sqrt(min_size * max_size)
            box_width = box_height = sqrtf(min_size * max_size);
            //! xmin
            max_buf[max_idx++] = (center_x - box_width / 2.f) / img_width;
            //! ymin
            max_buf[max_idx++] = (center_y - box_height / 2.f) / img_height;
            //! xmax
            max_buf[max_idx++] = (center_x + box_width / 2.f) / img_width;
            //! ymax
            max_buf[max_idx++] = (center_y + box_height / 2.f) / img_height;
          }

          //! rest of priors
          for (int r = 0; r < aspect_ratio_.size(); ++r) {
            float ar = aspect_ratio_[r];
            if (fabs(ar - 1.) < 1e-6) {
              continue;
            }
            box_width = min_size * sqrt(ar);
            box_height = min_size / sqrt(ar);
            //! xmin
            com_buf[com_idx++] = (center_x - box_width / 2.f) / img_width;
            //! ymin
            com_buf[com_idx++] = (center_y - box_height / 2.f) / img_height;
            //! xmax
            com_buf[com_idx++] = (center_x + box_width / 2.f) / img_width;
            //! ymax
            com_buf[com_idx++] = (center_y + box_height / 2.f) / img_height;
          }
288 289 290 291 292 293 294 295 296 297 298 299 300 301 302
          if (min_max_aspect_ratios_order) {
            memcpy(_cpu_data + idx, min_buf, sizeof(float) * min_idx);
            idx += min_idx;
            memcpy(_cpu_data + idx, max_buf, sizeof(float) * max_idx);
            idx += max_idx;
            memcpy(_cpu_data + idx, com_buf, sizeof(float) * com_idx);
            idx += com_idx;
          } else {
            memcpy(_cpu_data + idx, min_buf, sizeof(float) * min_idx);
            idx += min_idx;
            memcpy(_cpu_data + idx, com_buf, sizeof(float) * com_idx);
            idx += com_idx;
            memcpy(_cpu_data + idx, max_buf, sizeof(float) * max_idx);
            idx += max_idx;
          }
Y
Yan Chunwei 已提交
303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345
        }
        fast_free(min_buf);
        fast_free(max_buf);
        fast_free(com_buf);
      }
    }
  }
  //! clip the prior's coordinate such that it is within [0, 1]
  if (is_clip_) {
    for (int d = 0; d < channel_size; ++d) {
      _cpu_data[d] = std::min(std::max(_cpu_data[d], 0.f), 1.f);
    }
  }
  //! set the variance.
  int count = 0;
  for (int h = 0; h < height; ++h) {
    for (int w = 0; w < width; ++w) {
      for (int i = 0; i < prior_num_; ++i) {
        for (int j = 0; j < 4; ++j) {
          _variance_data[count] = variance_[j];
          ++count;
        }
      }
    }
  }
}

void prior_box(const lite::Tensor* input,
               const lite::Tensor* image,
               lite::Tensor** boxes,
               lite::Tensor** variances,
               const std::vector<float>& min_size,
               const std::vector<float>& max_size,
               const std::vector<float>& aspect_ratio,
               const std::vector<float>& variance,
               int img_w,
               int img_h,
               float step_w,
               float step_h,
               float offset,
               int prior_num,
               bool is_flip,
               bool is_clip,
346 347
               const std::vector<std::string>& order,
               bool min_max_aspect_ratios_order) {
Y
Yan Chunwei 已提交
348 349 350 351 352 353 354
  density_prior_box(input,
                    image,
                    boxes,
                    variances,
                    min_size,
                    std::vector<float>(),
                    std::vector<float>(),
T
TianXiaogang 已提交
355
                    std::vector<int>(),
Y
Yan Chunwei 已提交
356 357 358 359 360 361 362 363 364 365 366
                    max_size,
                    aspect_ratio,
                    variance,
                    img_w,
                    img_h,
                    step_w,
                    step_h,
                    offset,
                    prior_num,
                    is_flip,
                    is_clip,
367 368
                    order,
                    min_max_aspect_ratios_order);
Y
Yan Chunwei 已提交
369 370 371 372 373 374
}

}  // namespace math
}  // namespace arm
}  // namespace lite
}  // namespace paddle