yolov3_loss_op.h 14.8 KB
Newer Older
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15
/* Copyright (c) 2018 PaddlePaddle Authors. All Rights Reserve.
   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
#include <algorithm>
#include <vector>
#include "paddle/fluid/framework/op_registry.h"
16
#include "paddle/fluid/operators/math/math_function.h"
17 18 19 20 21 22 23 24 25 26 27 28 29

namespace paddle {
namespace operators {

using Tensor = framework::Tensor;
template <typename T, size_t D, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenTensor = framework::EigenTensor<T, D, MajorType, IndexType>;
template <typename T, int MajorType = Eigen::RowMajor,
          typename IndexType = Eigen::DenseIndex>
using EigenVector = framework::EigenVector<T, MajorType, IndexType>;

template <typename T>
D
dengkaipeng 已提交
30 31
static inline bool LessEqualZero(T x) {
  return x < 1e-6;
32 33
}

34
template <typename T>
35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53
static T SCE(T x, T label) {
  return (x > 0 ? x : 0.0) - x * label + std::log(1.0 + std::exp(-std::abs(x)));
}

template <typename T>
static T L1Loss(T x, T y) {
  return std::abs(y - x);
}

template <typename T>
static T SCEGrad(T x, T label) {
  return 1.0 / (1.0 + std::exp(-x)) - label;
}

template <typename T>
static T L1LossGrad(T x, T y) {
  return x > y ? 1.0 : -1.0;
}

D
dengkaipeng 已提交
54 55
static int GetMaskIndex(std::vector<int> mask, int val) {
  for (size_t i = 0; i < mask.size(); i++) {
56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73
    if (mask[i] == val) {
      return i;
    }
  }
  return -1;
}

template <typename T>
struct Box {
  float x, y, w, h;
};

template <typename T>
static inline T sigmoid(T x) {
  return 1.0 / (1.0 + std::exp(-x));
}

template <typename T>
D
dengkaipeng 已提交
74 75 76
static inline Box<T> GetYoloBox(const T* x, std::vector<int> anchors, int i,
                                int j, int an_idx, int grid_size,
                                int input_size, int index, int stride) {
77 78 79 80 81 82 83 84 85
  Box<T> b;
  b.x = (i + sigmoid<T>(x[index])) / grid_size;
  b.y = (j + sigmoid<T>(x[index + stride])) / grid_size;
  b.w = std::exp(x[index + 2 * stride]) * anchors[2 * an_idx] / input_size;
  b.h = std::exp(x[index + 3 * stride]) * anchors[2 * an_idx + 1] / input_size;
  return b;
}

template <typename T>
D
dengkaipeng 已提交
86
static inline Box<T> GetGtBox(const T* gt, int batch, int max_boxes, int idx) {
87 88 89 90 91 92 93 94 95
  Box<T> b;
  b.x = gt[(batch * max_boxes + idx) * 4];
  b.y = gt[(batch * max_boxes + idx) * 4 + 1];
  b.w = gt[(batch * max_boxes + idx) * 4 + 2];
  b.h = gt[(batch * max_boxes + idx) * 4 + 3];
  return b;
}

template <typename T>
D
dengkaipeng 已提交
96
static inline T BoxOverlap(T c1, T w1, T c2, T w2) {
97 98 99 100 101 102 103 104 105 106
  T l1 = c1 - w1 / 2.0;
  T l2 = c2 - w2 / 2.0;
  T left = l1 > l2 ? l1 : l2;
  T r1 = c1 + w1 / 2.0;
  T r2 = c2 + w2 / 2.0;
  T right = r1 < r2 ? r1 : r2;
  return right - left;
}

template <typename T>
D
dengkaipeng 已提交
107 108 109
static inline T CalcBoxIoU(Box<T> b1, Box<T> b2) {
  T w = BoxOverlap(b1.x, b1.w, b2.x, b2.w);
  T h = BoxOverlap(b1.y, b1.h, b2.y, b2.h);
110 111 112 113 114
  T inter_area = (w < 0 || h < 0) ? 0.0 : w * h;
  T union_area = b1.w * b1.h + b2.w * b2.h - inter_area;
  return inter_area / union_area;
}

D
dengkaipeng 已提交
115 116
static inline int GetEntryIndex(int batch, int an_idx, int hw_idx, int an_num,
                                int an_stride, int stride, int entry) {
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
  return (batch * an_num + an_idx) * an_stride + entry * stride + hw_idx;
}

template <typename T>
static void CalcBoxLocationLoss(T* loss, const T* input, Box<T> gt,
                                std::vector<int> anchors, int an_idx,
                                int box_idx, int gi, int gj, int grid_size,
                                int input_size, int stride) {
  T tx = gt.x * grid_size - gi;
  T ty = gt.y * grid_size - gj;
  T tw = std::log(gt.w * input_size / anchors[2 * an_idx]);
  T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]);

  T scale = 2.0 - gt.w * gt.h;
  loss[0] += SCE<T>(input[box_idx], tx) * scale;
  loss[0] += SCE<T>(input[box_idx + stride], ty) * scale;
  loss[0] += L1Loss<T>(input[box_idx + 2 * stride], tw) * scale;
  loss[0] += L1Loss<T>(input[box_idx + 3 * stride], th) * scale;
}

template <typename T>
static void CalcBoxLocationLossGrad(T* input_grad, const T loss, const T* input,
                                    Box<T> gt, std::vector<int> anchors,
                                    int an_idx, int box_idx, int gi, int gj,
                                    int grid_size, int input_size, int stride) {
  T tx = gt.x * grid_size - gi;
  T ty = gt.y * grid_size - gj;
  T tw = std::log(gt.w * input_size / anchors[2 * an_idx]);
  T th = std::log(gt.h * input_size / anchors[2 * an_idx + 1]);

  T scale = 2.0 - gt.w * gt.h;
  input_grad[box_idx] = SCEGrad<T>(input[box_idx], tx) * scale * loss;
  input_grad[box_idx + stride] =
      SCEGrad<T>(input[box_idx + stride], ty) * scale * loss;
  input_grad[box_idx + 2 * stride] =
      L1LossGrad<T>(input[box_idx + 2 * stride], tw) * scale * loss;
  input_grad[box_idx + 3 * stride] =
      L1LossGrad<T>(input[box_idx + 3 * stride], th) * scale * loss;
}

template <typename T>
static inline void CalcLabelLoss(T* loss, const T* input, const int index,
                                 const int label, const int class_num,
                                 const int stride) {
  for (int i = 0; i < class_num; i++) {
    loss[0] += SCE<T>(input[index + i * stride], (i == label) ? 1.0 : 0.0);
  }
}

template <typename T>
static inline void CalcLabelLossGrad(T* input_grad, const T loss,
                                     const T* input, const int index,
                                     const int label, const int class_num,
                                     const int stride) {
  for (int i = 0; i < class_num; i++) {
    input_grad[index + i * stride] =
        SCEGrad<T>(input[index + i * stride], (i == label) ? 1.0 : 0.0) * loss;
  }
}

template <typename T>
static inline void CalcObjnessLoss(T* loss, const T* input, const int* objness,
                                   const int n, const int an_num, const int h,
                                   const int w, const int stride,
                                   const int an_stride) {
  for (int i = 0; i < n; i++) {
    for (int j = 0; j < an_num; j++) {
      for (int k = 0; k < h; k++) {
        for (int l = 0; l < w; l++) {
          int obj = objness[k * w + l];
          if (obj >= 0) {
            loss[i] += SCE<T>(input[k * w + l], static_cast<T>(obj));
          }
        }
      }
      objness += stride;
      input += an_stride;
    }
  }
}

template <typename T>
static inline void CalcObjnessLossGrad(T* input_grad, const T* loss,
                                       const T* input, const int* objness,
                                       const int n, const int an_num,
                                       const int h, const int w,
                                       const int stride, const int an_stride) {
  for (int i = 0; i < n; i++) {
    for (int j = 0; j < an_num; j++) {
      for (int k = 0; k < h; k++) {
        for (int l = 0; l < w; l++) {
          int obj = objness[k * w + l];
          if (obj >= 0) {
            input_grad[k * w + l] =
                SCEGrad<T>(input[k * w + l], static_cast<T>(obj)) * loss[i];
          }
        }
      }
      objness += stride;
      input += an_stride;
      input_grad += an_stride;
    }
  }
}

222
template <typename T>
223 224 225 226
class Yolov3LossKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* input = ctx.Input<Tensor>("X");
D
dengkaipeng 已提交
227 228
    auto* gt_box = ctx.Input<Tensor>("GTBox");
    auto* gt_label = ctx.Input<Tensor>("GTLabel");
D
dengkaipeng 已提交
229
    auto* loss = ctx.Output<Tensor>("Loss");
230 231
    auto* objness_mask = ctx.Output<Tensor>("ObjectnessMask");
    auto* gt_match_mask = ctx.Output<Tensor>("GTMatchMask");
232
    auto anchors = ctx.Attr<std::vector<int>>("anchors");
233
    auto anchor_mask = ctx.Attr<std::vector<int>>("anchor_mask");
234 235
    int class_num = ctx.Attr<int>("class_num");
    float ignore_thresh = ctx.Attr<float>("ignore_thresh");
236
    int downsample = ctx.Attr<int>("downsample");
237 238 239 240 241

    const int n = input->dims()[0];
    const int h = input->dims()[2];
    const int w = input->dims()[3];
    const int an_num = anchors.size() / 2;
242 243 244
    const int mask_num = anchor_mask.size();
    const int b = gt_box->dims()[1];
    int input_size = downsample * h;
245

246 247 248
    const int stride = h * w;
    const int an_stride = (class_num + 5) * stride;

249 250 251
    const T* input_data = input->data<T>();
    const T* gt_box_data = gt_box->data<T>();
    const int* gt_label_data = gt_label->data<int>();
252
    T* loss_data = loss->mutable_data<T>({n}, ctx.GetPlace());
D
dengkaipeng 已提交
253
    memset(loss_data, 0, loss->numel() * sizeof(T));
254 255 256 257 258
    int* obj_mask_data =
        objness_mask->mutable_data<int>({n, mask_num, h, w}, ctx.GetPlace());
    memset(obj_mask_data, 0, objness_mask->numel() * sizeof(int));
    int* gt_match_mask_data =
        gt_match_mask->mutable_data<int>({n, b}, ctx.GetPlace());
259 260 261 262 263 264

    for (int i = 0; i < n; i++) {
      for (int j = 0; j < mask_num; j++) {
        for (int k = 0; k < h; k++) {
          for (int l = 0; l < w; l++) {
            int box_idx =
D
dengkaipeng 已提交
265 266 267
                GetEntryIndex(i, j, k * w + l, mask_num, an_stride, stride, 0);
            Box<T> pred = GetYoloBox(input_data, anchors, l, k, anchor_mask[j],
                                     h, input_size, box_idx, stride);
268 269
            T best_iou = 0;
            for (int t = 0; t < b; t++) {
D
dengkaipeng 已提交
270 271
              Box<T> gt = GetGtBox(gt_box_data, i, b, t);
              if (LessEqualZero<T>(gt.w) || LessEqualZero<T>(gt.h)) {
272 273
                continue;
              }
D
dengkaipeng 已提交
274
              T iou = CalcBoxIoU(pred, gt);
275 276 277 278 279 280 281
              if (iou > best_iou) {
                best_iou = iou;
              }
            }

            if (best_iou > ignore_thresh) {
              int obj_idx = (i * mask_num + j) * stride + k * w + l;
282
              obj_mask_data[obj_idx] = -1;
283 284 285 286 287
            }
          }
        }
      }
      for (int t = 0; t < b; t++) {
D
dengkaipeng 已提交
288 289
        Box<T> gt = GetGtBox(gt_box_data, i, b, t);
        if (LessEqualZero<T>(gt.w) || LessEqualZero<T>(gt.h)) {
290
          gt_match_mask_data[i * b + t] = -1;
291 292 293 294 295 296 297 298 299 300 301 302 303 304 305
          continue;
        }
        int gi = static_cast<int>(gt.x * w);
        int gj = static_cast<int>(gt.y * h);
        Box<T> gt_shift = gt;
        gt_shift.x = 0.0;
        gt_shift.y = 0.0;
        T best_iou = 0.0;
        int best_n = 0;
        for (int an_idx = 0; an_idx < an_num; an_idx++) {
          Box<T> an_box;
          an_box.x = 0.0;
          an_box.y = 0.0;
          an_box.w = anchors[2 * an_idx] / static_cast<T>(input_size);
          an_box.h = anchors[2 * an_idx + 1] / static_cast<T>(input_size);
D
dengkaipeng 已提交
306
          float iou = CalcBoxIoU<T>(an_box, gt_shift);
307 308 309 310 311 312 313
          // TO DO: iou > 0.5 ?
          if (iou > best_iou) {
            best_iou = iou;
            best_n = an_idx;
          }
        }

D
dengkaipeng 已提交
314
        int mask_idx = GetMaskIndex(anchor_mask, best_n);
315
        gt_match_mask_data[i * b + t] = mask_idx;
316
        if (mask_idx >= 0) {
D
dengkaipeng 已提交
317 318
          int box_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
                                      an_stride, stride, 0);
319 320 321 322
          CalcBoxLocationLoss<T>(loss_data + i, input_data, gt, anchors, best_n,
                                 box_idx, gi, gj, h, input_size, stride);

          int obj_idx = (i * mask_num + mask_idx) * stride + gj * w + gi;
323
          obj_mask_data[obj_idx] = 1;
324 325

          int label = gt_label_data[i * b + t];
D
dengkaipeng 已提交
326 327
          int label_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
                                        an_stride, stride, 5);
328 329 330 331 332 333
          CalcLabelLoss<T>(loss_data + i, input_data, label_idx, label,
                           class_num, stride);
        }
      }
    }

334
    CalcObjnessLoss<T>(loss_data, input_data + 4 * stride, obj_mask_data, n,
335
                       mask_num, h, w, stride, an_stride);
336 337 338
  }
};

339
template <typename T>
340 341 342
class Yolov3LossGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
343
    auto* input = ctx.Input<Tensor>("X");
D
dengkaipeng 已提交
344 345
    auto* gt_box = ctx.Input<Tensor>("GTBox");
    auto* gt_label = ctx.Input<Tensor>("GTLabel");
346 347
    auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
    auto* loss_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));
348 349
    auto* objness_mask = ctx.Input<Tensor>("ObjectnessMask");
    auto* gt_match_mask = ctx.Input<Tensor>("GTMatchMask");
350
    auto anchors = ctx.Attr<std::vector<int>>("anchors");
351
    auto anchor_mask = ctx.Attr<std::vector<int>>("anchor_mask");
352
    int class_num = ctx.Attr<int>("class_num");
353
    int downsample = ctx.Attr<int>("downsample");
354

355 356 357 358
    const int n = input_grad->dims()[0];
    const int c = input_grad->dims()[1];
    const int h = input_grad->dims()[2];
    const int w = input_grad->dims()[3];
359
    const int mask_num = anchor_mask.size();
360
    const int b = gt_match_mask->dims()[1];
361 362
    int input_size = downsample * h;

363 364 365
    const int stride = h * w;
    const int an_stride = (class_num + 5) * stride;

366 367 368 369
    const T* input_data = input->data<T>();
    const T* gt_box_data = gt_box->data<T>();
    const int* gt_label_data = gt_label->data<int>();
    const T* loss_grad_data = loss_grad->data<T>();
370 371
    const int* obj_mask_data = objness_mask->data<int>();
    const int* gt_match_mask_data = gt_match_mask->data<int>();
372 373
    T* input_grad_data =
        input_grad->mutable_data<T>({n, c, h, w}, ctx.GetPlace());
374 375 376 377
    memset(input_grad_data, 0, input_grad->numel() * sizeof(T));

    for (int i = 0; i < n; i++) {
      for (int t = 0; t < b; t++) {
D
dengkaipeng 已提交
378 379
        Box<T> gt = GetGtBox(gt_box_data, i, b, t);
        if (LessEqualZero<T>(gt.w) || LessEqualZero<T>(gt.h)) {
380 381 382 383 384
          continue;
        }
        int gi = static_cast<int>(gt.x * w);
        int gj = static_cast<int>(gt.y * h);

385
        int mask_idx = gt_match_mask_data[i * b + t];
386
        if (mask_idx >= 0) {
D
dengkaipeng 已提交
387 388
          int box_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
                                      an_stride, stride, 0);
389 390 391
          CalcBoxLocationLossGrad<T>(
              input_grad_data, loss_grad_data[i], input_data, gt, anchors,
              anchor_mask[mask_idx], box_idx, gi, gj, h, input_size, stride);
392 393

          int label = gt_label_data[i * b + t];
D
dengkaipeng 已提交
394 395
          int label_idx = GetEntryIndex(i, mask_idx, gj * w + gi, mask_num,
                                        an_stride, stride, 5);
396 397 398 399 400 401 402
          CalcLabelLossGrad<T>(input_grad_data, loss_grad_data[i], input_data,
                               label_idx, label, class_num, stride);
        }
      }
    }

    CalcObjnessLossGrad<T>(input_grad_data + 4 * stride, loss_grad_data,
403
                           input_data + 4 * stride, obj_mask_data, n, mask_num,
404
                           h, w, stride, an_stride);
405 406 407 408 409
  }
};

}  // namespace operators
}  // namespace paddle