detection_layer对应Yolov1的实现,理解detection_layer的实现主要需要理解Yolov1中输出层的存储方式,以及论文中损失函数的构成。
论文深入理解,请参考YOLO v1深入理解。本文只对detection_layer的实现进行注释分析。
/*
** 构建detection层
** 输入: batch 一个batch中含有的图片张数(等于net.batch)
** inputs 该层输入数据维度大小
** n 一个grid cell预测bound box的数量
** side grid cell的尺寸大小
** classes 目标类别数
** coords 学习的位置参数个数
** max_boxes 图像真实目标的最多个数
** 返回: detection_layer
*/
detection_layer make_detection_layer(int batch, int inputs, int n, int side, int classes, int coords, int rescore)
{
detection_layer l = { (LAYER_TYPE)0 };
l.type = DETECTION;
l.n = n;
l.batch = batch;
l.inputs = inputs;
l.classes = classes;
l.coords = coords;
l.rescore = rescore;
l.side = side;
l.w = side;
l.h = side;
assert(side*side*((1 + l.coords)*l.n + l.classes) == inputs);
l.cost = (float*)calloc(1, sizeof(float));
l.outputs = l.inputs;
//一个grid cell 只负责预测一个ground truth,因此此处不需要乘以l.n
l.truths = l.side*l.side*(1+l.coords+l.classes);
l.output = (float*)calloc(batch * l.outputs, sizeof(float));
l.delta = (float*)calloc(batch * l.outputs, sizeof(float));
l.forward = forward_detection_layer;
l.backward = backward_detection_layer;
#ifdef GPU
l.forward_gpu = forward_detection_layer_gpu;
l.backward_gpu = backward_detection_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 Layer\n");
srand(time(0));
return l;
}
Yolov1中的损失函数如下:
void forward_detection_layer(const detection_layer l, network_state state)
{
//7*7=49个grid cell
int locations = l.side*l.side;
int i,j;
memcpy(l.output, state.input, l.outputs*l.batch*sizeof(float));//copy state.input to l.output
//if(l.reorg) reorg(l.output, l.w*l.h, size*l.n, l.batch, 1);
int b;
if (l.softmax){
for(b = 0; b < l.batch; ++b){
int index = b*l.inputs;
for (i = 0; i < locations; ++i) {
int offset = i*l.classes;
//注意输出数据的存储方式(跟yolov2,v3的区别)
softmax(l.output + index + offset, l.classes, 1,
l.output + index + offset, 1);
}
}
}
if(state.train){
float avg_iou = 0;
float avg_cat = 0;
float avg_allcat = 0;
float avg_obj = 0;
float avg_anyobj = 0;
int count = 0;
*(l.cost) = 0;
int size = l.inputs * l.batch;
//初始化当前层的sensitivity
memset(l.delta, 0, size * sizeof(float));
for (b = 0; b < l.batch; ++b){
int index = b*l.inputs;
for (i = 0; i < locations; ++i) {
int truth_index = (b*locations + i)*(1+l.coords+l.classes);//ground truth索引
int is_obj = state.truth[truth_index];
for (j = 0; j < l.n; ++j) {
//预测边框的索引(第i个grid cell的第j个bounding box)
int p_index = index + locations*l.classes + i*l.n + j;
//预测边框置信度损失(相当于对第四行对$\hat_{C_i}$求导)
l.delta[p_index] = l.noobject_scale*(0 - l.output[p_index]);
*(l.cost) += l.noobject_scale*pow(l.output[p_index], 2);//相当于第四行,当该box没有对象的时候,Ci为0
avg_anyobj += l.output[p_index];
}
int best_index = -1;
float best_iou = 0;
float best_rmse = 20;
//该grid cell没有object,则只计算第四行的损失
if (!is_obj){
continue;
}
//20个类别概率索引
int class_index = index + i*l.classes;
for(j = 0; j < l.classes; ++j) {
l.delta[class_index+j] = l.class_scale * (state.truth[truth_index+1+j] - l.output[class_index+j]);
//计算第5行
*(l.cost) += l.class_scale * pow(state.truth[truth_index+1+j] - l.output[class_index+j], 2);
if(state.truth[truth_index + 1 + j]) avg_cat += l.output[class_index+j];
avg_allcat += l.output[class_index+j];
}
//注意state.truth的存储方式
box truth = float_to_box(state.truth + truth_index + 1 + l.classes);
truth.x /= l.side;
truth.y /= l.side;
for(j = 0; j < l.n; ++j){
//边框坐标位置索引
int box_index = index + locations*(l.classes + l.n) + (i*l.n + j) * l.coords;
box out = float_to_box(l.output + box_index);
out.x /= l.side;
out.y /= l.side;
if (l.sqrt){
out.w = out.w*out.w;
out.h = out.h*out.h;
}
float iou = box_iou(out, truth);
//iou = 0;
float rmse = box_rmse(out, truth);
if(best_iou > 0 || iou > 0){
if(iou > best_iou){
best_iou = iou;
best_index = j;
}
}else{
if(rmse < best_rmse){
best_rmse = rmse;
best_index = j;
}
}
}
if(l.forced){
if(truth.w*truth.h < .1){
best_index = 1;
}else{
best_index = 0;
}
}
if(l.random && *(state.net.seen) < 64000){
best_index = rand()%l.n;
}
int box_index = index + locations*(l.classes + l.n) + (i*l.n + best_index) * l.coords;
int tbox_index = truth_index + 1 + l.classes;
box out = float_to_box(l.output + box_index);
out.x /= l.side;
out.y /= l.side;
if (l.sqrt) {
out.w = out.w*out.w;
out.h = out.h*out.h;
}
float iou = box_iou(out, truth);
//printf("%d,", best_index);
int p_index = index + locations*l.classes + i*l.n + best_index;//负责预测ground truth的bounding box
*(l.cost) -= l.noobject_scale * pow(l.output[p_index], 2);
*(l.cost) += l.object_scale * pow(1-l.output[p_index], 2);
avg_obj += l.output[p_index];
l.delta[p_index] = l.object_scale * (1.-l.output[p_index]);
if(l.rescore){
l.delta[p_index] = l.object_scale * (iou - l.output[p_index]);
}
//第一行求导
l.delta[box_index+0] = l.coord_scale*(state.truth[tbox_index + 0] - l.output[box_index + 0]);
l.delta[box_index+1] = l.coord_scale*(state.truth[tbox_index + 1] - l.output[box_index + 1]);
//第二行求导
l.delta[box_index+2] = l.coord_scale*(state.truth[tbox_index + 2] - l.output[box_index + 2]);
l.delta[box_index+3] = l.coord_scale*(state.truth[tbox_index + 3] - l.output[box_index + 3]);
if(l.sqrt){
l.delta[box_index+2] = l.coord_scale*(sqrt(state.truth[tbox_index + 2]) - l.output[box_index + 2]);
l.delta[box_index+3] = l.coord_scale*(sqrt(state.truth[tbox_index + 3]) - l.output[box_index + 3]);
}
*(l.cost) += pow(1-iou, 2);
avg_iou += iou;
++count;
}
}
if(0){
float* costs = (float*)calloc(l.batch * locations * l.n, sizeof(float));
for (b = 0; b < l.batch; ++b) {
int index = b*l.inputs;
for (i = 0; i < locations; ++i) {
for (j = 0; j < l.n; ++j) {
int p_index = index + locations*l.classes + i*l.n + j;
costs[b*locations*l.n + i*l.n + j] = l.delta[p_index]*l.delta[p_index];
}
}
}
int indexes[100];
top_k(costs, l.batch*locations*l.n, 100, indexes);
float cutoff = costs[indexes[99]];
for (b = 0; b < l.batch; ++b) {
int index = b*l.inputs;
for (i = 0; i < locations; ++i) {
for (j = 0; j < l.n; ++j) {
int p_index = index + locations*l.classes + i*l.n + j;
if (l.delta[p_index]*l.delta[p_index] < cutoff) l.delta[p_index] = 0;
}
}
}
free(costs);
}
*(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);
printf("Detection Avg IOU: %f, Pos Cat: %f, All Cat: %f, Pos Obj: %f, Any Obj: %f, count: %d\n", avg_iou/count, avg_cat/count, avg_allcat/(count*l.classes), avg_obj/count, avg_anyobj/(l.batch*locations*l.n), count);
//if(l.reorg) reorg(l.delta, l.w*l.h, size*l.n, l.batch, 0);
}
}
/*
** 获取detection层的检测边框(跟get_detection_detections一样只是返回结果不一样)
** 输入: l 网络检测层
** w, h 网络输入图像的宽高
** thresh 阈值
** probs 类别概率
** boxes 检测到的边框
** only_objectness 用于决定probs返回的是置信度还是对象类别概率
** 返回: detection_layer
*/
void get_detection_boxes(layer l, int w, int h, float thresh, float **probs, box *boxes, int only_objectness)
{
int i,j,n;
float *predictions = l.output;
//int per_cell = 5*num+classes;
for (i = 0; i < l.side*l.side; ++i){
int row = i / l.side;
int col = i % l.side;
for(n = 0; n < l.n; ++n){
int index = i*l.n + n;//预测候选框索引(最多能预测7*7*2=98个候选框)
int p_index = l.side*l.side*l.classes + i*l.n + n;
float scale = predictions[p_index];
int box_index = l.side*l.side*(l.classes + l.n) + (i*l.n + n)*4;
boxes[index].x = (predictions[box_index + 0] + col) / l.side * w;
boxes[index].y = (predictions[box_index + 1] + row) / l.side * h;
boxes[index].w = pow(predictions[box_index + 2], (l.sqrt?2:1)) * w;
boxes[index].h = pow(predictions[box_index + 3], (l.sqrt?2:1)) * h;
for(j = 0; j < l.classes; ++j){
int class_index = i*l.classes;
float prob = scale*predictions[class_index+j];
probs[index][j] = (prob > thresh) ? prob : 0;
}
if(only_objectness){
probs[index][0] = scale;
}
}
}
}
/*
** 获取detection层的检测结果
** 输入: batch 一个batch中含有的图片张数(等于net.batch)
** w, h 网络输入图像的宽高
** thresh 阈值
** dets 检测结果
** 返回: detection_layer
*/
void get_detection_detections(layer l, int w, int h, float thresh, detection *dets)
{
int i, j, n;
float *predictions = l.output;
//int per_cell = 5*num+classes;
for (i = 0; i < l.side*l.side; ++i) {
int row = i / l.side;
int col = i % l.side;
for (n = 0; n < l.n; ++n) {
int index = i*l.n + n;//预测候选框索引(最多能预测7*7*2=98个候选框)
int p_index = l.side*l.side*l.classes + i*l.n + n;////预测候选框目标置信度存储索引
float scale = predictions[p_index];
int box_index = l.side*l.side*(l.classes + l.n) + (i*l.n + n) * 4;////预测候选框坐标位置存储索引
box b;
b.x = (predictions[box_index + 0] + col) / l.side * w;
b.y = (predictions[box_index + 1] + row) / l.side * h;
b.w = pow(predictions[box_index + 2], (l.sqrt ? 2 : 1)) * w;
b.h = pow(predictions[box_index + 3], (l.sqrt ? 2 : 1)) * h;
dets[index].bbox = b;
dets[index].objectness = scale;
for (j = 0; j < l.classes; ++j) {
int class_index = i*l.classes;
float prob = scale*predictions[class_index + j];//第i个grid cell中的第n个预测框是第j类的概率
dets[index].prob[j] = (prob > thresh) ? prob : 0;
}
}
}
}