【darknet源码解析-19】region_layer.h 和 region_layer.c 解析

本系列为darknet源码解析，本次解析src/region_layer.h 与 src/region_layer.c 两个。region_layer主要完成了yolo v2最后一层13*13*125，是yolo v2这篇论文的核心部分。

在阅读本节源码之前，请先了解一下 13*13*125 是什么样子的逻辑存储形式，在物体存储是一维数组；以及yolov2中bbox的[x, y, w, h]是如何进行表示的，本节只解析了yolov2的训练阶段的源码，inference阶段未进行解析；配对的cfg文件为cfg/yolov2-voc.cfg

 yolov2 对每个预测box的[x,y]，confidence进行逻辑回归，类别进行softmax回归；

region_layer.h 的定义如下：

#ifndef REORG_LAYER_H
#define REORG_LAYER_H

#include "image.h"
#include "cuda.h"
#include "layer.h"
#include "network.h"

// 构造yolov2 reorg层
layer make_reorg_layer(int batch, int w, int h, int c, int stride, int reverse, int flatten, int extra);
void resize_reorg_layer(layer *l, int w, int h);
// yolov2 reorg层前向传播函数
void forward_reorg_layer(const layer l, network net);
// yolov2 reorg层反向传播函数
void backward_reorg_layer(const layer l, network net);

#ifdef GPU
void forward_reorg_layer_gpu(layer l, network net);
void backward_reorg_layer_gpu(layer l, network net);
#endif

#endif

region_layer.c 详细解释如下：

//
// Created by wdong on 19-11-7.
//


#include "region_layer.h"
#include "activations.h"
#include "blas.h"
#include "box.h"
#include "cuda.h"
#include "utils.h"

#include <stdio.h>
#include <assert.h>
#include <string.h>
#include <stdlib.h>

/**
 * yolov2中的region层
 * @param batch 一个batch中包含图片的张数
 * @param w 输入特征图的宽度
 * @param h 输入特征图的高度
 * @param n 一个cell预测多少个bbox
 * @param classes 网络需要识别的物体类数
 * @param coords 一个bbox包含的[x, y, w, h]
 * @return
 */
layer make_region_layer(int batch, int w, int h, int n, int classes, int coords)
{
    layer l = {0};
    l.type = REGION; // 层类别

    l.n = n; // 一个cell预测多少个bbox
    l.batch = batch; // 一个batch中包含图片的张数
    l.h = h; // 输入图片的宽度
    l.w = w; // 输入图片的宽度
    l.c = n*(classes + coords + 1); // 输入图片的通道数 n*(20 + 4 + 1)
    l.out_w = l.w; // 输出图片的宽度
    l.out_h = l.h; // 输出图片的高度
    l.out_c = l.c; // 输出图片的通道数
    l.classes = classes; // 网络需要识别的物体类数
    l.coords = coords; // 一个bbox包含的[x, y, w, h]
    l.cost = calloc(1, sizeof(float)); // region层的总损失
    l.biases = calloc(n*2, sizeof(float)); // 存储bbox的Anchor box的[w,h]
    l.bias_updates = calloc(n*2, sizeof(float)); //存储bbox的Anchor box的[w,h]的更新值
    l.outputs = h*w*n*(classes + coords + 1); // reorg层对应输入图片的输出元素个数，reorg层输入输出元素个数不发生变化
    l.inputs = l.outputs; // reorg层一张输入图片的元素个数
    l.truths = 30*(l.coords + 1); // GT: 30*(4+1) 存储30个bbox的信息，这里是假设图片中GT bbox的数量是小于30的，这里是写死的；此处与yolov1 处是不同的；
    l.delta = calloc(batch*l.outputs, sizeof(float)); // region层误差项（包含整个batch的）
    l.output = calloc(batch*l.outputs, sizeof(float)); // region层所有输出（包含整个batch的）
    int i;
    // 存储bbox的Anchor box的[w,h]的初始化,在src/parse.c中parse_yolo函数会加载cfg中Anchor尺寸
    for(i = 0; i < n*2; ++i){
        l.biases[i] = .5;
    }

    l.forward = forward_region_layer; // reorg层前向传播
    l.backward = backward_region_layer; // reorg层反向传播
#ifdef GPU
    l.forward_gpu = forward_region_layer_gpu;
    l.backward_gpu = backward_region_layer_gpu;
    l.output_gpu = cuda_make_array(l.output, batch*l.outputs);
    l.delta_gpu = cuda_make_array(l.delta, batch*l.outputs);
#endif

    fprintf(stderr, "detection\n");
    srand(0);

    return l;
}

void resize_region_layer(layer *l, int w, int h)
{
    l->w = w;
    l->h = h;

    l->outputs = h*w*l->n*(l->classes + l->coords + 1);
    l->inputs = l->outputs;

    l->output = realloc(l->output, l->batch*l->outputs*sizeof(float));
    l->delta = realloc(l->delta, l->batch*l->outputs*sizeof(float));

#ifdef GPU
    cuda_free(l->delta_gpu);
    cuda_free(l->output_gpu);

    l->delta_gpu =     cuda_make_array(l->delta, l->batch*l->outputs);
    l->output_gpu =    cuda_make_array(l->output, l->batch*l->outputs);
#endif
}

// get_region_box(l.output, l.biases, n, box_index, i, j, l.w, l.h, l.w*l.h);
box get_region_box(float *x, float *biases, int n, int index, int i, int j, int w, int h, int stride)
{
    box b;
    b.x = (i + x[index + 0*stride]) / w;
    b.y = (j + x[index + 1*stride]) / h;
    b.w = exp(x[index + 2*stride]) * biases[2*n]   / w;
    b.h = exp(x[index + 3*stride]) * biases[2*n+1] / h;
    return b;
}


// float_to_box(net.truth + t*(l.coords + 1) + b*l.truths, 1)
box float_to_box(float *f, int stride)
{
    box b = {0};
    b.x = f[0];
    b.y = f[1*stride];
    b.w = f[2*stride];
    b.h = f[3*stride];
    return b;
}

// delta_region_box(truth, l.output, l.biases, n, box_index, i, j, l.w, l.h, l.delta, .01, l.w*l.h);
float delta_region_box(box truth, float *x, float *biases, int n, int index, int i, int j, int w, int h, float *delta, float scale, int stride)
{
    // 获得第j*w+i个cell第n个bbox在当前特征图上位置和宽高
    box pred = get_region_box(x, biases, n, index, i, j, w, h, stride);
    float iou = box_iou(pred, truth); // 计算pred bbox 与 GT bbox的IOU【前12800GT boox为当前cell第n个bbox的Anchor】

    float tx = (truth.x*w - i); // 计算GT bbox的tx,ty,tw,th
    float ty = (truth.y*h - j);
    float tw = log(truth.w*w / biases[2*n]);
    float th = log(truth.h*h / biases[2*n + 1]);

    delta[index + 0*stride] = scale * (tx - x[index + 0*stride]); // 计算tx,ty,tw,th梯度
    delta[index + 1*stride] = scale * (ty - x[index + 1*stride]);
    delta[index + 2*stride] = scale * (tw - x[index + 2*stride]);
    delta[index + 3*stride] = scale * (th - x[index + 3*stride]);
    return iou;
}

void delta_region_mask(float *truth, float *x, int n, int index, float *delta, int stride, int scale)
{
    int i;
    for(i = 0; i < n; ++i){
        delta[index + i*stride] = scale*(truth[i] - x[index + i*stride]);
    }
}

// delta_region_class(l.output, l.delta, class_index, class, l.classes, l.softmax_tree, l.class_scale, l.w*l.h, &avg_cat, !l.softmax);
void delta_region_class(float *output, float *delta, int index, int class, int classes, tree *hier, float scale, int stride, float *avg_cat, int tag)
{
    int i, n;
    if(hier){ // 在yolov2 中region层, 此部分不参与计算【这是在yolo9000才参与计算】
        float pred = 1;
        while(class >= 0){
            pred *= output[index + stride*class];
            int g = hier->group[class];
            int offset = hier->group_offset[g];
            for(i = 0; i < hier->group_size[g]; ++i){
                delta[index + stride*(offset + i)] = scale * (0 - output[index + stride*(offset + i)]);
            }
            delta[index + stride*class] = scale * (1 - output[index + stride*class]);

            class = hier->parent[class];
        }
        *avg_cat += pred;
    } else {
        if (delta[index] && tag){ //yolo v2中softmax = 1, 所以此处tag=0, 不参与计算
            delta[index + stride*class] = scale * (1 - output[index + stride*class]);
            return;
        }
        for(n = 0; n < classes; ++n){
            // 计算类别损失的梯度, 反向传递到误差项l.delta中, 在yolo v2中scale=1
            delta[index + stride*n] = scale * (((n == class)?1 : 0) - output[index + stride*n]);
            if(n == class) *avg_cat += output[index + stride*n]; // 统计正确的得分
        }
    }
}

float logit(float x)
{
    return log(x/(1.-x));
}

float tisnan(float x)
{
    return (x != x);
}

// int box_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0);
int entry_index(layer l, int batch, int location, int entry)
{
    int n =   location / (l.w*l.h);
    int loc = location % (l.w*l.h);
    return batch*l.outputs + n*l.w*l.h*(l.coords+l.classes+1) + entry*l.w*l.h + loc;
}

/**
 * region层前向传播函数
 * @param l 当前region层
 * @param net 整个网络
 */
void forward_region_layer(const layer l, network net)
{
    int i,j,b,t,n;
    // 内存拷贝, l.output = net.input
    memcpy(l.output, net.input, l.outputs*l.batch*sizeof(float));

#ifndef GPU
    for (b = 0; b < l.batch; ++b){
        for(n = 0; n < l.n; ++n){
            int index = entry_index(l, b, n*l.w*l.h, 0);
            activate_array(l.output + index, 2*l.w*l.h, LOGISTIC); // 对[x,y]进行逻辑回归
            index = entry_index(l, b, n*l.w*l.h, l.coords);
            // l.background = 0, 对confidence进行逻辑回归
            if(!l.background) activate_array(l.output + index,   l.w*l.h, LOGISTIC);
            index = entry_index(l, b, n*l.w*l.h, l.coords + 1);
            if(!l.softmax && !l.softmax_tree) activate_array(l.output + index, l.classes*l.w*l.h, LOGISTIC);
        }
    }
    if (l.softmax_tree){
        int i;
        int count = l.coords + 1;
        for (i = 0; i < l.softmax_tree->groups; ++i) {
            int group_size = l.softmax_tree->group_size[i];
            softmax_cpu(net.input + count, group_size, l.batch, l.inputs, l.n*l.w*l.h, 1, l.n*l.w*l.h, l.temperature, l.output + count);
            count += group_size;
        }
    } else if (l.softmax){ // l.softmax 对class进行softmax回归
        int index = entry_index(l, 0, 0, l.coords + !l.background);
        softmax_cpu(net.input + index, l.classes + l.background, l.batch*l.n, l.inputs/l.n, l.w*l.h, 1, l.w*l.h, 1, l.output + index);
    }
#endif

    memset(l.delta, 0, l.outputs * l.batch * sizeof(float)); // 将reorg层的误差项进行初始化（包含整个batch的）
    if(!net.train) return; // inference阶段，则到此结束
    float avg_iou = 0;
    float recall = 0;
    float avg_cat = 0;
    float avg_obj = 0;
    float avg_anyobj = 0;
    int count = 0;
    int class_count = 0;
    *(l.cost) = 0; // // region层的总损失初始化为0
    for (b = 0; b < l.batch; ++b) { // 遍历batch中每一张图片
        if(l.softmax_tree){ //【这是在yolo9000才参与计算】
            int onlyclass = 0;
            for(t = 0; t < 30; ++t){
                box truth = float_to_box(net.truth + t*(l.coords + 1) + b*l.truths, 1);
                if(!truth.x) break;
                int class = net.truth[t*(l.coords + 1) + b*l.truths + l.coords];
                float maxp = 0;
                int maxi = 0;
                if(truth.x > 100000 && truth.y > 100000){
                    for(n = 0; n < l.n*l.w*l.h; ++n){
                        int class_index = entry_index(l, b, n, l.coords + 1);
                        int obj_index = entry_index(l, b, n, l.coords);
                        float scale =  l.output[obj_index];
                        l.delta[obj_index] = l.noobject_scale * (0 - l.output[obj_index]);
                        float p = scale*get_hierarchy_probability(l.output + class_index, l.softmax_tree, class, l.w*l.h);
                        if(p > maxp){
                            maxp = p;
                            maxi = n;
                        }
                    }
                    int class_index = entry_index(l, b, maxi, l.coords + 1);
                    int obj_index = entry_index(l, b, maxi, l.coords);
                    delta_region_class(l.output, l.delta, class_index, class, l.classes, l.softmax_tree, l.class_scale, l.w*l.h, &avg_cat, !l.softmax);
                    if(l.output[obj_index] < .3) l.delta[obj_index] = l.object_scale * (.3 - l.output[obj_index]);
                    else  l.delta[obj_index] = 0;
                    l.delta[obj_index] = 0;
                    ++class_count;
                    onlyclass = 1;
                    break;
                }
            }
            if(onlyclass) continue;
        } // if(l.softmax_tree)
        for (j = 0; j < l.h; ++j) {
            for (i = 0; i < l.w; ++i) { // 遍历每个cell, 当前cell编号为[j, i]
                for (n = 0; n < l.n; ++n) { // 遍历每个bbox，当前bbox编号为[n]
                    int box_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0); // 获得第j*w+i个cell第n个bbox的index
                    // 获得第j*w+i个cell第n个bbox在当前特征图上位置和宽高
                    box pred = get_region_box(l.output, l.biases, n, box_index, i, j, l.w, l.h, l.w*l.h);
                    float best_iou = 0; // 保存最大IOU
                    for(t = 0; t < 30; ++t){ // 遍历每一个GT bbox
                        // 将第t个bbox由float数组转bbox结构体,方便计算IOU
                        box truth = float_to_box(net.truth + t*(l.coords + 1) + b*l.truths, 1);
                        if(!truth.x) break; // 如果x坐标为0则取消, 因为yolov2这里定义了30 bbox, 可能实际上没有bbox
                        float iou = box_iou(pred, truth); // 计算pred与第t个GT之间的IOU
                        if (iou > best_iou) {
                            best_iou = iou; // 最大IOU更新
                        }
                    }
                    int obj_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, l.coords); // 获得第j*w+i个cell第n个bbox的confidence的index
                    avg_anyobj += l.output[obj_index]; // 统计有目标的概率
                    // 与yolov1相似, 先将所有pred bbox都当做noobject，计算其confidence损失梯度
                    l.delta[obj_index] = l.noobject_scale * (0 - l.output[obj_index]);
                    if(l.background) l.delta[obj_index] = l.noobject_scale * (1 - l.output[obj_index]); // 在yolov2中并没有执行
                    if (best_iou > l.thresh) { // best_iou大于阈值则说明有object, 在yolo v2中阈值为0.6
                        l.delta[obj_index] = 0;
                    }

                    if(*(net.seen) < 12800){ // net.seen 保存当前是训练第多少张图片

                        // 对于训练阶段的前12800张图片,GT bbox 直接用了anchor box
                        box truth = {0}; // 计算第[j, i]cell, 第n个bbox的anchor bbox
                        truth.x = (i + .5)/l.w; // +0.5是因为x位于几何中心, 然后计算x相对整张特征图的位置
                        truth.y = (j + .5)/l.h;
                        truth.w = l.biases[2*n]/l.w;
                        truth.h = l.biases[2*n+1]/l.h;
                        // 将pred bbox的tx,ty,tw,th和上面的truth box的差值反向传递到l.detla
                        delta_region_box(truth, l.output, l.biases, n, box_index, i, j, l.w, l.h, l.delta, .01, l.w*l.h);
                    }
                }
            }
        }
        for(t = 0; t < 30; ++t){ // 遍历每一个GT bbox
            // 将第t个bbox由float数组转bbox结构体,方便计算IOU
            box truth = float_to_box(net.truth + t*(l.coords + 1) + b*l.truths, 1);

            if(!truth.x) break; // 如果x坐标为0则取消, 因为yolov2这里定义了30 bbox, 可能实际上没有bbox
            float best_iou = 0; // 保存最大IOU
            int best_n = 0; // 保存最大IOU的bbox index
            i = (truth.x * l.w); // 获得当前第t个GT bbox所在cell
            j = (truth.y * l.h);
            box truth_shift = truth;
            truth_shift.x = 0; // 将truth_shift的box移动到0,0
            truth_shift.y = 0;
            for(n = 0; n < l.n; ++n){ // 遍历cell[j,i]所在的n个预测bbox
                int box_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0); // 获得第j*w+i个cell第n个bbox的index
                // 获得第j*w+i个cell第n个bbox在当前特征图上位置和宽高
                box pred = get_region_box(l.output, l.biases, n, box_index, i, j, l.w, l.h, l.w*l.h);
                if(l.bias_match){ // yolov2 reorg层 bias_match = 1
                    pred.w = l.biases[2*n]/l.w; // 然后计算pred box的w相对整张特征图的位置
                    pred.h = l.biases[2*n+1]/l.h; // 然后计算pred box的h相对整张特征图的位置
                }
                pred.x = 0; // 将预测的bbox移动到0,0
                pred.y = 0;
                float iou = box_iou(pred, truth_shift); // 计算GT box truth_shift 与 预测bbox pred 二者之间的IOU
                if (iou > best_iou){
                    best_iou = iou; // 记录IOU最大的IOU
                    best_n = n; // 以及记录该bbox的编号n
                }
            }

            int box_index = entry_index(l, b, best_n*l.w*l.h + j*l.w + i, 0); // 获得best_iou对应bbox的index
            // 计算获得best_iou的pred bbox 与 GT bbox之间的真实iou, 之前best_iou是方便计算,以及加速,
            // 同时完成坐标损失的反向传递
            float iou = delta_region_box(truth, l.output, l.biases, best_n, box_index, i, j, l.w, l.h, l.delta, l.coord_scale *  (2 - truth.w*truth.h), l.w*l.h);
            if(l.coords > 4){
                int mask_index = entry_index(l, b, best_n*l.w*l.h + j*l.w + i, 4);
                delta_region_mask(net.truth + t*(l.coords + 1) + b*l.truths + 5, l.output, l.coords - 4, mask_index, l.delta, l.w*l.h, l.mask_scale);
            }
            // 如果iou大于0.5, recall ++;
            if(iou > .5) recall += 1;
            avg_iou += iou;

            // 获得best_iou对应bbox的confidence的index
            int obj_index = entry_index(l, b, best_n*l.w*l.h + j*l.w + i, l.coords);
            avg_obj += l.output[obj_index]; // 统计有目标的概率

            // 与yolov1相似, 该pred bbox中是有object，计算其confidence损失梯度; object_scale = 5
            l.delta[obj_index] = l.object_scale * (1 - l.output[obj_index]);  // 实际上是没有执行, 因为被下面一句覆盖了
            if (l.rescore) { // yolov2 reorg层中rescore = 1, 参于计算
                // 该pred bbox中是有object，计算其confidence损失梯度的方法发生变化; object_scale = 5,
                l.delta[obj_index] = l.object_scale * (iou - l.output[obj_index]);
            }
            if(l.background){ // yolov2 reorg层中background = 0, 不参与计算
                l.delta[obj_index] = l.object_scale * (0 - l.output[obj_index]);
            }

            // 获取GT bbox的类别信息
            int class = net.truth[t*(l.coords + 1) + b*l.truths + l.coords];
            if (l.map) class = l.map[class]; // yolov2 reorg层中map = 0, 不参与计算 【这是在yolo9000才参与计算】
            // 获取best_iou对应bbox的class的index
            int class_index = entry_index(l, b, best_n*l.w*l.h + j*l.w + i, l.coords + 1);
            //  // 与yolov1相似, 该pred bbox中是有object，计算其class损失梯度; class_scale = 1
            delta_region_class(l.output, l.delta, class_index, class, l.classes, l.softmax_tree, l.class_scale, l.w*l.h, &avg_cat, !l.softmax);
            ++count; // 正样本个数+1
            ++class_count; // emmmm, 跟count的含义貌似一样啊
        }
    }
    // 计算均方误差
    *(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);
    printf("Region Avg IOU: %f, Class: %f, Obj: %f, No Obj: %f, Avg Recall: %f,  count: %d\n", avg_iou/count, avg_cat/class_count, avg_obj/count, avg_anyobj/(l.w*l.h*l.n*l.batch), recall/count, count);
}

/**
 * yolov2 reorg层反向传播函数
 * @param l 当前reorg层
 * @param net 整个网络
 * 说明：其实这部分其实作者已经写入到前向传播函数中了
 */
void backward_region_layer(const layer l, network net)
{
    /*
       int b;
       int size = l.coords + l.classes + 1;
       for (b = 0; b < l.batch*l.n; ++b){
       int index = (b*size + 4)*l.w*l.h;
       gradient_array(l.output + index, l.w*l.h, LOGISTIC, l.delta + index);
       }
       axpy_cpu(l.batch*l.inputs, 1, l.delta, 1, net.delta, 1);
     */
}

void correct_region_boxes(detection *dets, int n, int w, int h, int netw, int neth, int relative)
{
    int i;
    int new_w=0;
    int new_h=0;
    if (((float)netw/w) < ((float)neth/h)) {
        new_w = netw;
        new_h = (h * netw)/w;
    } else {
        new_h = neth;
        new_w = (w * neth)/h;
    }
    for (i = 0; i < n; ++i){
        box b = dets[i].bbox;
        b.x =  (b.x - (netw - new_w)/2./netw) / ((float)new_w/netw);
        b.y =  (b.y - (neth - new_h)/2./neth) / ((float)new_h/neth);
        b.w *= (float)netw/new_w;
        b.h *= (float)neth/new_h;
        if(!relative){
            b.x *= w;
            b.w *= w;
            b.y *= h;
            b.h *= h;
        }
        dets[i].bbox = b;
    }
}

void get_region_detections(layer l, int w, int h, int netw, int neth, float thresh, int *map, float tree_thresh, int relative, detection *dets)
{
    int i,j,n,z;
    float *predictions = l.output;
    if (l.batch == 2) {
        float *flip = l.output + l.outputs;
        for (j = 0; j < l.h; ++j) {
            for (i = 0; i < l.w/2; ++i) {
                for (n = 0; n < l.n; ++n) {
                    for(z = 0; z < l.classes + l.coords + 1; ++z){
                        int i1 = z*l.w*l.h*l.n + n*l.w*l.h + j*l.w + i;
                        int i2 = z*l.w*l.h*l.n + n*l.w*l.h + j*l.w + (l.w - i - 1);
                        float swap = flip[i1];
                        flip[i1] = flip[i2];
                        flip[i2] = swap;
                        if(z == 0){
                            flip[i1] = -flip[i1];
                            flip[i2] = -flip[i2];
                        }
                    }
                }
            }
        }
        for(i = 0; i < l.outputs; ++i){
            l.output[i] = (l.output[i] + flip[i])/2.;
        }
    }
    for (i = 0; i < l.w*l.h; ++i){
        int row = i / l.w;
        int col = i % l.w;
        for(n = 0; n < l.n; ++n){
            int index = n*l.w*l.h + i;
            for(j = 0; j < l.classes; ++j){
                dets[index].prob[j] = 0;
            }
            int obj_index  = entry_index(l, 0, n*l.w*l.h + i, l.coords);
            int box_index  = entry_index(l, 0, n*l.w*l.h + i, 0);
            int mask_index = entry_index(l, 0, n*l.w*l.h + i, 4);
            float scale = l.background ? 1 : predictions[obj_index];
            dets[index].bbox = get_region_box(predictions, l.biases, n, box_index, col, row, l.w, l.h, l.w*l.h);
            dets[index].objectness = scale > thresh ? scale : 0;
            if(dets[index].mask){
                for(j = 0; j < l.coords - 4; ++j){
                    dets[index].mask[j] = l.output[mask_index + j*l.w*l.h];
                }
            }

            int class_index = entry_index(l, 0, n*l.w*l.h + i, l.coords + !l.background);
            if(l.softmax_tree){

                hierarchy_predictions(predictions + class_index, l.classes, l.softmax_tree, 0, l.w*l.h);
                if(map){
                    for(j = 0; j < 200; ++j){
                        int class_index = entry_index(l, 0, n*l.w*l.h + i, l.coords + 1 + map[j]);
                        float prob = scale*predictions[class_index];
                        dets[index].prob[j] = (prob > thresh) ? prob : 0;
                    }
                } else {
                    int j =  hierarchy_top_prediction(predictions + class_index, l.softmax_tree, tree_thresh, l.w*l.h);
                    dets[index].prob[j] = (scale > thresh) ? scale : 0;
                }
            } else {
                if(dets[index].objectness){
                    for(j = 0; j < l.classes; ++j){
                        int class_index = entry_index(l, 0, n*l.w*l.h + i, l.coords + 1 + j);
                        float prob = scale*predictions[class_index];
                        dets[index].prob[j] = (prob > thresh) ? prob : 0;
                    }
                }
            }
        }
    }
    correct_region_boxes(dets, l.w*l.h*l.n, w, h, netw, neth, relative);
}

#ifdef GPU

void forward_region_layer_gpu(const layer l, network net)
{
    copy_gpu(l.batch*l.inputs, net.input_gpu, 1, l.output_gpu, 1);
    int b, n;
    for (b = 0; b < l.batch; ++b){
        for(n = 0; n < l.n; ++n){
            int index = entry_index(l, b, n*l.w*l.h, 0);
            activate_array_gpu(l.output_gpu + index, 2*l.w*l.h, LOGISTIC);
            if(l.coords > 4){
                index = entry_index(l, b, n*l.w*l.h, 4);
                activate_array_gpu(l.output_gpu + index, (l.coords - 4)*l.w*l.h, LOGISTIC);
            }
            index = entry_index(l, b, n*l.w*l.h, l.coords);
            if(!l.background) activate_array_gpu(l.output_gpu + index,   l.w*l.h, LOGISTIC);
            index = entry_index(l, b, n*l.w*l.h, l.coords + 1);
            if(!l.softmax && !l.softmax_tree) activate_array_gpu(l.output_gpu + index, l.classes*l.w*l.h, LOGISTIC);
        }
    }
    if (l.softmax_tree){
        int index = entry_index(l, 0, 0, l.coords + 1);
        softmax_tree(net.input_gpu + index, l.w*l.h, l.batch*l.n, l.inputs/l.n, 1, l.output_gpu + index, *l.softmax_tree);
    } else if (l.softmax) {
        int index = entry_index(l, 0, 0, l.coords + !l.background);
        softmax_gpu(net.input_gpu + index, l.classes + l.background, l.batch*l.n, l.inputs/l.n, l.w*l.h, 1, l.w*l.h, 1, l.output_gpu + index);
    }
    if(!net.train || l.onlyforward){
        cuda_pull_array(l.output_gpu, l.output, l.batch*l.outputs);
        return;
    }

    cuda_pull_array(l.output_gpu, net.input, l.batch*l.inputs);
    forward_region_layer(l, net);
    //cuda_push_array(l.output_gpu, l.output, l.batch*l.outputs);
    if(!net.train) return;
    cuda_push_array(l.delta_gpu, l.delta, l.batch*l.outputs);
}

void backward_region_layer_gpu(const layer l, network net)
{
    int b, n;
    for (b = 0; b < l.batch; ++b){
        for(n = 0; n < l.n; ++n){
            int index = entry_index(l, b, n*l.w*l.h, 0);
            gradient_array_gpu(l.output_gpu + index, 2*l.w*l.h, LOGISTIC, l.delta_gpu + index);
            if(l.coords > 4){
                index = entry_index(l, b, n*l.w*l.h, 4);
                gradient_array_gpu(l.output_gpu + index, (l.coords - 4)*l.w*l.h, LOGISTIC, l.delta_gpu + index);
            }
            index = entry_index(l, b, n*l.w*l.h, l.coords);
            if(!l.background) gradient_array_gpu(l.output_gpu + index,   l.w*l.h, LOGISTIC, l.delta_gpu + index);
        }
    }
    axpy_gpu(l.batch*l.inputs, 1, l.delta_gpu, 1, net.delta_gpu, 1);
}
#endif

void zero_objectness(layer l)
{
    int i, n;
    for (i = 0; i < l.w*l.h; ++i){
        for(n = 0; n < l.n; ++n){
            int obj_index = entry_index(l, 0, n*l.w*l.h + i, l.coords);
            l.output[obj_index] = 0;
        }
    }
}

完，