yolov3_loss_op.h

/* 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"

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>;

using Array5 = Eigen::DSizes<int64_t, 5>;

template <typename T>
static inline bool isZero(T x) {
  return abs(x) < 1e-6;
}

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

template <typename T>
static inline T CalcMaskPointNum(const Tensor& mask) {
  auto mask_t = EigenVector<int>::Flatten(mask);
  T count = 0.0;
  for (int i = 0; i < mask_t.dimensions()[0]; i++) {
    if (mask_t(i)) {
      count += 1.0;
    }
  }
  return count;
}

template <typename T>
static inline T CalcMSEWithMask(const Tensor& x, const Tensor& y,
                                const Tensor& mask) {
  auto x_t = EigenVector<T>::Flatten(x);
  auto y_t = EigenVector<T>::Flatten(y);
  auto mask_t = EigenVector<int>::Flatten(mask);

  T error_sum = 0.0;
  T points = 0.0;
  for (int i = 0; i < x_t.dimensions()[0]; i++) {
    if (mask_t(i)) {
      error_sum += pow(x_t(i) - y_t(i), 2);
      points += 1;
    }
  }
  return (error_sum / points);
}

template <typename T>
static void CalcMSEGradWithMask(Tensor* grad, const Tensor& x, const Tensor& y,
                                const Tensor& mask, T mf) {
  auto grad_t = EigenVector<T>::Flatten(*grad).setConstant(0.0);
  auto x_t = EigenVector<T>::Flatten(x);
  auto y_t = EigenVector<T>::Flatten(y);
  auto mask_t = EigenVector<int>::Flatten(mask);

  for (int i = 0; i < x_t.dimensions()[0]; i++) {
    if (mask_t(i)) {
      grad_t(i) = 2.0 * (x_t(i) - y_t(i)) / mf;
    }
  }
}

template <typename T>
static inline T CalcBCEWithMask(const Tensor& x, const Tensor& y,
                                const Tensor& mask) {
  auto x_t = EigenVector<T>::Flatten(x);
  auto y_t = EigenVector<T>::Flatten(y);
  auto mask_t = EigenVector<int>::Flatten(mask);

  T error_sum = 0.0;
  T points = 0.0;
  for (int i = 0; i < x_t.dimensions()[0]; i++) {
    if (mask_t(i)) {
      error_sum +=
          -1.0 * (y_t(i) * log(x_t(i)) + (1.0 - y_t(i)) * log(1.0 - x_t(i)));
      points += 1;
    }
  }
  return (error_sum / points);
}

template <typename T>
static inline void CalcBCEGradWithMask(Tensor* grad, const Tensor& x,
                                       const Tensor& y, const Tensor& mask,
                                       T mf) {
  auto grad_t = EigenVector<T>::Flatten(*grad).setConstant(0.0);
  auto x_t = EigenVector<T>::Flatten(x);
  auto y_t = EigenVector<T>::Flatten(y);
  auto mask_t = EigenVector<int>::Flatten(mask);

  for (int i = 0; i < x_t.dimensions()[0]; i++) {
    if (mask_t(i)) {
      grad_t(i) = ((1.0 - y_t(i)) / (1.0 - x_t(i)) - y_t(i) / x_t(i)) / mf;
    }
  }
}

template <typename T>
static void CalcPredResult(const Tensor& input, Tensor* pred_conf,
                           Tensor* pred_class, Tensor* pred_x, Tensor* pred_y,
                           Tensor* pred_w, Tensor* pred_h, const int anchor_num,
                           const int class_num) {
  const int n = input.dims()[0];
  const int h = input.dims()[2];
  const int w = input.dims()[3];
  const int box_attr_num = 5 + class_num;

  auto input_t = EigenTensor<T, 4>::From(input);
  auto pred_conf_t = EigenTensor<T, 4>::From(*pred_conf);
  auto pred_class_t = EigenTensor<T, 5>::From(*pred_class);
  auto pred_x_t = EigenTensor<T, 4>::From(*pred_x);
  auto pred_y_t = EigenTensor<T, 4>::From(*pred_y);
  auto pred_w_t = EigenTensor<T, 4>::From(*pred_w);
  auto pred_h_t = EigenTensor<T, 4>::From(*pred_h);

  for (int i = 0; i < n; i++) {
    for (int an_idx = 0; an_idx < anchor_num; an_idx++) {
      for (int j = 0; j < h; j++) {
        for (int k = 0; k < w; k++) {
          pred_x_t(i, an_idx, j, k) =
              sigmoid(input_t(i, box_attr_num * an_idx, j, k));
          pred_y_t(i, an_idx, j, k) =
              sigmoid(input_t(i, box_attr_num * an_idx + 1, j, k));
          pred_w_t(i, an_idx, j, k) =
              input_t(i, box_attr_num * an_idx + 2, j, k);
          pred_h_t(i, an_idx, j, k) =
              input_t(i, box_attr_num * an_idx + 3, j, k);

          pred_conf_t(i, an_idx, j, k) =
              sigmoid(input_t(i, box_attr_num * an_idx + 4, j, k));

          for (int c = 0; c < class_num; c++) {
            pred_class_t(i, an_idx, j, k, c) =
                sigmoid(input_t(i, box_attr_num * an_idx + 5 + c, j, k));
          }
        }
      }
    }
  }
}

template <typename T>
static T CalcBoxIoU(std::vector<T> box1, std::vector<T> box2) {
  T b1_x1 = box1[0] - box1[2] / 2;
  T b1_x2 = box1[0] + box1[2] / 2;
  T b1_y1 = box1[1] - box1[3] / 2;
  T b1_y2 = box1[1] + box1[3] / 2;
  T b2_x1 = box2[0] - box2[2] / 2;
  T b2_x2 = box2[0] + box2[2] / 2;
  T b2_y1 = box2[1] - box2[3] / 2;
  T b2_y2 = box2[1] + box2[3] / 2;

  T b1_area = (b1_x2 - b1_x1) * (b1_y2 - b1_y1);
  T b2_area = (b2_x2 - b2_x1) * (b2_y2 - b2_y1);

  T inter_rect_x1 = std::max(b1_x1, b2_x1);
  T inter_rect_y1 = std::max(b1_y1, b2_y1);
  T inter_rect_x2 = std::min(b1_x2, b2_x2);
  T inter_rect_y2 = std::min(b1_y2, b2_y2);
  T inter_area = std::max(inter_rect_x2 - inter_rect_x1, static_cast<T>(0.0)) *
                 std::max(inter_rect_y2 - inter_rect_y1, static_cast<T>(0.0));

  return inter_area / (b1_area + b2_area - inter_area);
}

template <typename T>
static void PrePorcessGTBox(const Tensor& gt_boxes, const float ignore_thresh,
                            std::vector<int> anchors, const int grid_size,
                            Tensor* obj_mask, Tensor* noobj_mask, Tensor* tx,
                            Tensor* ty, Tensor* tw, Tensor* th, Tensor* tconf,
                            Tensor* tclass) {
  const int n = gt_boxes.dims()[0];
  const int b = gt_boxes.dims()[1];
  const int anchor_num = anchors.size() / 2;
  auto gt_boxes_t = EigenTensor<T, 3>::From(gt_boxes);
  auto obj_mask_t = EigenTensor<int, 4>::From(*obj_mask).setConstant(0);
  auto noobj_mask_t = EigenTensor<int, 4>::From(*noobj_mask).setConstant(1);
  auto tx_t = EigenTensor<T, 4>::From(*tx).setConstant(0.0);
  auto ty_t = EigenTensor<T, 4>::From(*ty).setConstant(0.0);
  auto tw_t = EigenTensor<T, 4>::From(*tw).setConstant(0.0);
  auto th_t = EigenTensor<T, 4>::From(*th).setConstant(0.0);
  auto tconf_t = EigenTensor<T, 4>::From(*tconf).setConstant(0.0);
  auto tclass_t = EigenTensor<T, 5>::From(*tclass).setConstant(0.0);

  for (int i = 0; i < n; i++) {
    for (int j = 0; j < b; j++) {
      if (isZero(gt_boxes_t(i, j, 0)) && isZero(gt_boxes_t(i, j, 1)) &&
          isZero(gt_boxes_t(i, j, 2)) && isZero(gt_boxes_t(i, j, 3))) {
        continue;
      }

      int gt_label = static_cast<int>(gt_boxes_t(i, j, 0));
      T gx = gt_boxes_t(i, j, 1) * grid_size;
      T gy = gt_boxes_t(i, j, 2) * grid_size;
      T gw = gt_boxes_t(i, j, 3) * grid_size;
      T gh = gt_boxes_t(i, j, 4) * grid_size;
      int gi = static_cast<int>(gx);
      int gj = static_cast<int>(gy);

      T max_iou = static_cast<T>(0);
      T iou;
      int best_an_index = -1;
      std::vector<T> gt_box({0, 0, gw, gh});
      for (int an_idx = 0; an_idx < anchor_num; an_idx++) {
        std::vector<T> anchor_shape({0, 0, static_cast<T>(anchors[2 * an_idx]),
                                     static_cast<T>(anchors[2 * an_idx + 1])});
        iou = CalcBoxIoU<T>(gt_box, anchor_shape);
        if (iou > max_iou) {
          max_iou = iou;
          best_an_index = an_idx;
        }
        if (iou > ignore_thresh) {
          noobj_mask_t(i, an_idx, gj, gi) = 0;
        }
      }
      obj_mask_t(i, best_an_index, gj, gi) = 1;
      noobj_mask_t(i, best_an_index, gj, gi) = 0;
      tx_t(i, best_an_index, gj, gi) = gx - gi;
      ty_t(i, best_an_index, gj, gi) = gy - gj;
      tw_t(i, best_an_index, gj, gi) = log(gw / anchors[2 * best_an_index]);
      th_t(i, best_an_index, gj, gi) = log(gh / anchors[2 * best_an_index + 1]);
      tclass_t(i, best_an_index, gj, gi, gt_label) = 1;
      tconf_t(i, best_an_index, gj, gi) = 1;
    }
  }
}

static void ExpandObjMaskByClassNum(Tensor* obj_mask_expand,
                                    const Tensor& obj_mask) {
  const int n = obj_mask_expand->dims()[0];
  const int an_num = obj_mask_expand->dims()[1];
  const int h = obj_mask_expand->dims()[2];
  const int w = obj_mask_expand->dims()[3];
  const int class_num = obj_mask_expand->dims()[4];
  auto obj_mask_expand_t = EigenTensor<int, 5>::From(*obj_mask_expand);
  auto obj_mask_t = EigenTensor<int, 4>::From(obj_mask);

  obj_mask_expand_t = obj_mask_t.reshape(Array5(n, an_num, h, w, 1))
                          .broadcast(Array5(1, 1, 1, 1, class_num));
}

template <typename T>
static void AddAllGradToInputGrad(
    Tensor* grad, T loss, const Tensor& pred_x, const Tensor& pred_y,
    const Tensor& pred_conf, const Tensor& pred_class, const Tensor& grad_x,
    const Tensor& grad_y, const Tensor& grad_w, const Tensor& grad_h,
    const Tensor& grad_conf_obj, const Tensor& grad_conf_noobj,
    const Tensor& grad_class, const int class_num) {
  const int n = pred_x.dims()[0];
  const int an_num = pred_x.dims()[1];
  const int h = pred_x.dims()[2];
  const int w = pred_x.dims()[3];
  const int attr_num = class_num + 5;
  auto grad_t = EigenTensor<T, 4>::From(*grad).setConstant(0.0);
  auto pred_x_t = EigenTensor<T, 4>::From(pred_x);
  auto pred_y_t = EigenTensor<T, 4>::From(pred_y);
  auto pred_conf_t = EigenTensor<T, 4>::From(pred_conf);
  auto pred_class_t = EigenTensor<T, 5>::From(pred_class);
  auto grad_x_t = EigenTensor<T, 4>::From(grad_x);
  auto grad_y_t = EigenTensor<T, 4>::From(grad_y);
  auto grad_w_t = EigenTensor<T, 4>::From(grad_w);
  auto grad_h_t = EigenTensor<T, 4>::From(grad_h);
  auto grad_conf_obj_t = EigenTensor<T, 4>::From(grad_conf_obj);
  auto grad_conf_noobj_t = EigenTensor<T, 4>::From(grad_conf_noobj);
  auto grad_class_t = EigenTensor<T, 5>::From(grad_class);

  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++) {
          grad_t(i, j * attr_num, k, l) = grad_x_t(i, j, k, l) *
                                          pred_x_t(i, j, k, l) *
                                          (1.0 - pred_x_t(i, j, k, l)) * loss;
          grad_t(i, j * attr_num + 1, k, l) =
              grad_y_t(i, j, k, l) * pred_y_t(i, j, k, l) *
              (1.0 - pred_y_t(i, j, k, l)) * loss;
          grad_t(i, j * attr_num + 2, k, l) = grad_w_t(i, j, k, l) * loss;
          grad_t(i, j * attr_num + 3, k, l) = grad_h_t(i, j, k, l) * loss;
          grad_t(i, j * attr_num + 4, k, l) =
              grad_conf_obj_t(i, j, k, l) * pred_conf_t(i, j, k, l) *
              (1.0 - pred_conf_t(i, j, k, l)) * loss;
          grad_t(i, j * attr_num + 4, k, l) +=
              grad_conf_noobj_t(i, j, k, l) * pred_conf_t(i, j, k, l) *
              (1.0 - pred_conf_t(i, j, k, l)) * loss;

          for (int c = 0; c < class_num; c++) {
            grad_t(i, j * attr_num + 5 + c, k, l) =
                grad_class_t(i, j, k, l, c) * pred_class_t(i, j, k, l, c) *
                (1.0 - pred_class_t(i, j, k, l, c)) * loss;
          }
        }
      }
    }
  }
}

template <typename DeviceContext, typename T>
class Yolov3LossKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* input = ctx.Input<Tensor>("X");
    auto* gt_boxes = ctx.Input<Tensor>("GTBox");
    auto* loss = ctx.Output<Tensor>("Loss");
    auto anchors = ctx.Attr<std::vector<int>>("anchors");
    int class_num = ctx.Attr<int>("class_num");
    float ignore_thresh = ctx.Attr<float>("ignore_thresh");

    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;

    Tensor pred_x, pred_y, pred_w, pred_h;
    Tensor pred_conf, pred_class;
    pred_x.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_y.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_w.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_h.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_conf.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_class.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
    CalcPredResult<T>(*input, &pred_conf, &pred_class, &pred_x, &pred_y,
                      &pred_w, &pred_h, an_num, class_num);

    Tensor obj_mask, noobj_mask;
    Tensor tx, ty, tw, th, tconf, tclass;
    obj_mask.mutable_data<int>({n, an_num, h, w}, ctx.GetPlace());
    noobj_mask.mutable_data<int>({n, an_num, h, w}, ctx.GetPlace());
    tx.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    ty.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    tw.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    th.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    tconf.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    tclass.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
    PrePorcessGTBox<T>(*gt_boxes, ignore_thresh, anchors, h, &obj_mask,
                       &noobj_mask, &tx, &ty, &tw, &th, &tconf, &tclass);

    Tensor obj_mask_expand;
    obj_mask_expand.mutable_data<int>({n, an_num, h, w, class_num},
                                      ctx.GetPlace());
    ExpandObjMaskByClassNum(&obj_mask_expand, obj_mask);

    T loss_x = CalcMSEWithMask<T>(pred_x, tx, obj_mask);
    T loss_y = CalcMSEWithMask<T>(pred_y, ty, obj_mask);
    T loss_w = CalcMSEWithMask<T>(pred_w, tw, obj_mask);
    T loss_h = CalcMSEWithMask<T>(pred_h, th, obj_mask);
    T loss_conf_obj = CalcBCEWithMask<T>(pred_conf, tconf, obj_mask);
    T loss_conf_noobj = CalcBCEWithMask<T>(pred_conf, tconf, noobj_mask);
    T loss_class = CalcBCEWithMask<T>(pred_class, tclass, obj_mask_expand);

    auto* loss_data = loss->mutable_data<T>({1}, ctx.GetPlace());
    loss_data[0] = loss_x + loss_y + loss_w + loss_h + loss_conf_obj +
                   loss_conf_noobj + loss_class;
  }
};

template <typename DeviceContext, typename T>
class Yolov3LossGradKernel : public framework::OpKernel<T> {
 public:
  void Compute(const framework::ExecutionContext& ctx) const override {
    auto* input = ctx.Input<Tensor>("X");
    auto* gt_boxes = ctx.Input<Tensor>("GTBox");
    auto anchors = ctx.Attr<std::vector<int>>("anchors");
    int class_num = ctx.Attr<int>("class_num");
    float ignore_thresh = ctx.Attr<float>("ignore_thresh");
    auto* input_grad = ctx.Output<Tensor>(framework::GradVarName("X"));
    auto* output_grad = ctx.Input<Tensor>(framework::GradVarName("Loss"));
    const T loss = output_grad->data<T>()[0];

    const int n = input->dims()[0];
    const int c = input->dims()[1];
    const int h = input->dims()[2];
    const int w = input->dims()[3];
    const int an_num = anchors.size() / 2;

    Tensor pred_x, pred_y, pred_w, pred_h;
    Tensor pred_conf, pred_class;
    pred_x.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_y.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_w.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_h.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_conf.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    pred_class.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
    CalcPredResult<T>(*input, &pred_conf, &pred_class, &pred_x, &pred_y,
                      &pred_w, &pred_h, an_num, class_num);

    Tensor obj_mask, noobj_mask;
    Tensor tx, ty, tw, th, tconf, tclass;
    obj_mask.mutable_data<int>({n, an_num, h, w}, ctx.GetPlace());
    noobj_mask.mutable_data<int>({n, an_num, h, w}, ctx.GetPlace());
    tx.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    ty.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    tw.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    th.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    tconf.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    tclass.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
    PrePorcessGTBox<T>(*gt_boxes, ignore_thresh, anchors, h, &obj_mask,
                       &noobj_mask, &tx, &ty, &tw, &th, &tconf, &tclass);

    Tensor obj_mask_expand;
    obj_mask_expand.mutable_data<int>({n, an_num, h, w, class_num},
                                      ctx.GetPlace());
    ExpandObjMaskByClassNum(&obj_mask_expand, obj_mask);

    Tensor grad_x, grad_y, grad_w, grad_h;
    Tensor grad_conf_obj, grad_conf_noobj, grad_class;
    grad_x.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    grad_y.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    grad_w.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    grad_h.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    grad_conf_obj.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    grad_conf_noobj.mutable_data<T>({n, an_num, h, w}, ctx.GetPlace());
    grad_class.mutable_data<T>({n, an_num, h, w, class_num}, ctx.GetPlace());
    T obj_mf = CalcMaskPointNum<int>(obj_mask);
    T noobj_mf = CalcMaskPointNum<int>(noobj_mask);
    T obj_expand_mf = CalcMaskPointNum<int>(obj_mask_expand);
    CalcMSEGradWithMask<T>(&grad_x, pred_x, tx, obj_mask, obj_mf);
    CalcMSEGradWithMask<T>(&grad_y, pred_y, ty, obj_mask, obj_mf);
    CalcMSEGradWithMask<T>(&grad_w, pred_w, tw, obj_mask, obj_mf);
    CalcMSEGradWithMask<T>(&grad_h, pred_h, th, obj_mask, obj_mf);
    CalcBCEGradWithMask<T>(&grad_conf_obj, pred_conf, tconf, obj_mask, obj_mf);
    CalcBCEGradWithMask<T>(&grad_conf_noobj, pred_conf, tconf, noobj_mask,
                           noobj_mf);
    CalcBCEGradWithMask<T>(&grad_class, pred_class, tclass, obj_mask_expand,
                           obj_expand_mf);

    input_grad->mutable_data<T>({n, c, h, w}, ctx.GetPlace());
    AddAllGradToInputGrad<T>(
        input_grad, loss, pred_x, pred_y, pred_conf, pred_class, grad_x, grad_y,
        grad_w, grad_h, grad_conf_obj, grad_conf_noobj, grad_class, class_num);
  }
};

}  // namespace operators
}  // namespace paddle