原 YOLO源码详解（五）-追本溯源7*7个grid
原 YOLO源码详解（五）- YOLO中的7*7个grid和RPN中的9个anchors
原 YOLO源码详解（四）- 反向传播（back propagation）
原 YOLOv2如何fine-tuning？
原 用YOLOv2模型训练VOC数据集
原 YOLO源码详解（三）- 前向传播（forward）
原 YOLO源码详解（二）- 函数剖析
原 YOLO源码详解（一）-训练
detection_layer make_detection_layer(int batch, int inputs, int n, int side, int classes, int coords, int rescore)
{
detection_layer l = {0};
l.type = DETECTION;
l.n = n;//对应论文里面的B,表示每个cell生成的bounding box数。
l.batch = batch;
l.inputs = inputs;//输入detection的维度，计算方式根据论文是S*S*(B*5 + C)，论文里面对应7*7*（2*5 + 20）
l.classes = classes;//类别数，论文里面对应是20
l.coords = coords;//坐标，（x,y,w,h）, (x,y)是相对于网格的数值[0,1],(w,h)，是相对于整个图片的大小。
l.rescore = rescore;
l.side = side;//网格side*side
assert(side*side*((1 + l.coords)*l.n + l.classes) == inputs);
l.cost = calloc(1, sizeof(float));
l.outputs = l.inputs;
l.truths = l.side*l.side*(1+l.coords+l.classes);
l.output = calloc(batch*l.outputs, sizeof(float));
l.delta = calloc(batch*l.outputs, sizeof(float));
#ifdef 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(0);
return l;
}void forward_detection_layer(const detection_layer l, network_state state)
{
int locations = l.side*l.side;
int i, j;
memcpy(l.output, state.input, l.outputs*l.batch*sizeof(float));
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;
softmax_array(l.output + index + offset, l.classes, 1,
l.output + index + offset);
}
int offset = locations*l.classes;
activate_array(l.output + index + offset, locations*l.n*(1 + l.coords), LOGISTIC);
}
}
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;
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) {
//truth 的存储方式是 [l.side*l.side*(1+l.classes+l.coords)]
int truth_index = (b*locations + i)*(1 + l.coords + l.classes);//这个地方表示的是输入图片的gound truth的index
int is_obj = state.truth[truth_index];//表示这个cell是否是object，数据存放在state里面
for (j = 0; j < l.n; ++j) {//计算noobj这一项的损失
int p_index = index + locations*l.classes + i*l.n + j;
//b*l.inputs + locations*l.classes + i*l.n + j;
//inputs 的存储方式是 [l.side*l.side*l.classes][l.side*l.side*l.n][l.side*l.side*l.n*l.coords]
l.delta[p_index] = l.noobject_scale*(0 - l.output[p_index]);//对应论文损失的第四项。这个计算很是奇怪，论文里面还是要判断一下是否是noobj？？？？他怎么没有判断就直接计算了
*(l.cost) += l.noobject_scale*pow(l.output[p_index], 2);//同上
avg_anyobj += l.output[p_index];
}
int best_index = -1;
float best_iou = 0;
float best_rmse = 20;
if (!is_obj){
continue;
}
//后面都是对应于object的情况。
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]);//对应论文五项损失的第五项
*(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];//float类型做判断这样写不太好。
avg_allcat += l.output[class_index + j];
}//l.classes
box truth = float_to_box(state.truth + truth_index + 1 + l.classes);//进行了loaction 对应 truth的获取。
truth.x /= l.side;//这个地方的计算不清楚。
truth.y /= l.side;
for (j = 0; j < l.n; ++j){
int box_index = index + locations*(l.classes + l.n) + (i*l.n + j) * l.coords;
box out = float_to_box(l.output + box_index);//cell里面预测的boundingbox
out.x /= l.side;//同上面的不清楚
out.y /= l.side;
if (l.sqrt){
out.w = out.w*out.w;
out.h = out.h*out.h;
}//if
float iou = box_iou(out, truth);//计算iou值，即out与truth的交除以他们的并。
//iou = 0;
float rmse = box_rmse(out, truth);//这个是计算boundingbox各项之间的差的平方和的开根号。
if (best_iou > 0 || iou > 0){
if (iou > best_iou){
best_iou = iou;
best_index = j;//存放cell里面最好的预测的boundingbox的idx
}
}
else{
if (rmse < best_rmse){
best_rmse = rmse;
best_index = j;
}
}//if
}//for
if (l.forced){//？？
if (truth.w*truth.h < .1){
best_index = 1;
}
else{
best_index = 0;
}
}
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;
//对应于论文里面的第三项。
*(l.cost) -= l.noobject_scale * pow(l.output[p_index], 2);//这里揭开前面的问题，这里相当于已经判定是objcet，就把之前算在里面noobjcet的减掉。
*(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]);//这里就覆盖到之前被noobject赋值的数。
if (l.rescore){
l.delta[p_index] = l.object_scale * (iou - l.output[p_index]);
}
//下面的四个对应于论文里面损失的第一项和第二项。
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]);
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 (l.softmax){
gradient_array(l.output + index + locations*l.classes, locations*l.n*(1 + l.coords),
LOGISTIC, l.delta + index + locations*l.classes);
}
}//loaction
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);
}//batchsize
}
