由于之前总结性的看了yolo系列的3篇，这两天好奇具体的yolo实现，于是看了看darknet有关的具体代码。darknet整体的CPU的代码直白清晰，挺好看的，虽然可能缺少一些小技巧的优化，但是都用CPU了，谁在意这个。
examples/darknet.c
官网上的训练命令是
./darknet detector train cfg/voc.data cfg/yolov3-voc.cfg darknet53.conv.74
顺路继续找下去的位置就是
examples/detector.c
里面的：
void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, int ngpus, int clear);
当然，其实上面也没什么卵用。反正直接看cfg里面的文件，可以知道用的是yolo层。可以在parse.c里面看到实际用的就是src/yolo_layer.c里面的实现。
直接加注释了，有些配置都是在cfg文件，还有parser里面可以看到。
layer make_yolo_layer(int batch, int w, int h, int n, int total, int *mask, int classes)
{
int i;
layer l = {0};
l.type = YOLO;
l.n = n; // 这一层用的anchor数量
l.total = total; // 所有的anchor数量
l.batch = batch;
l.h = h;
l.w = w;
l.c = n*(classes + 4 + 1); // classe, bbox, objectness
l.out_w = l.w;
l.out_h = l.h;
l.out_c = l.c;
l.classes = classes;
l.cost = calloc(1, sizeof(float));
l.biases = calloc(total*2, sizeof(float)); // anchor的具体值
// below 具体使用的那几个anchor
if(mask) l.mask = mask;
else{
l.mask = calloc(n, sizeof(int));
for(i = 0; i < n; ++i){
l.mask[i] = i;
}
}
l.bias_updates = calloc(n*2, sizeof(float));
l.outputs = h*w*n*(classes + 4 + 1);
l.inputs = l.outputs;
l.truths = 90*(4 + 1);
l.delta = calloc(batch*l.outputs, sizeof(float)); // MSE的差
l.output = calloc(batch*l.outputs, sizeof(float));
for(i = 0; i < total*2; ++i){
l.biases[i] = .5;
}
l.forward = forward_yolo_layer;
l.backward = backward_yolo_layer;
#ifdef GPU
l.forward_gpu = forward_yolo_layer_gpu;
l.backward_gpu = backward_yolo_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, "yolo\n");
srand(0);
return l;
}
两个循环。
首先，网络的每个输出的bbox都对比groudtruth，如果IOU > ignore则不参与训练，进一步的，大于truth则计算loss，参与训练，但是cfg文件中这个值设置的是1,所以应该就是忽略后面这个进一步的了。
第二个循环，对每个目标，查找最合适的anchor，如果本层负责这个尺寸的anchor，就计算对应的各loss。否则忽略。
box get_yolo_box(float *x, float *biases, int n, int index, int i, int j, int lw, int lh, int w, int h, int stride)
{
box b;
// lw/lh为网络输出大小， b.x, b.y 为全图相对尺寸
b.x = (i + x[index + 0*stride]) / lw;
b.y = (j + x[index + 1*stride]) / lh;
// w/h为网络输入大小， bias为anchor尺寸，b.w, b.h为全图相对尺寸
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;
}
// 计算boundbox的loss
float delta_yolo_box(box truth, float *x, float *biases, int n, int index,
int i, int j, int lw, int lh, int w, int h, float *delta,
float scale, int stride)
{
box pred = get_yolo_box(x, biases, n, index, i, j, lw, lh, w, h, stride);
float iou = box_iou(pred, truth);
float tx = (truth.x*lw - i);
float ty = (truth.y*lh - j);
float tw = log(truth.w*w / biases[2*n]);
float th = log(truth.h*h / biases[2*n + 1]);
// scale = 2 - truth.w * truth.h 干毛线的？
delta[index + 0*stride] = scale * (tx - x[index + 0*stride]);
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;
}
// 计算分类loss
void delta_yolo_class(float *output,
float *delta,
int index,
int class,
int classes,
int stride,
float *avg_cat)
{
int n;
// delta[index] is not 0. one class?
if (delta[index]){
delta[index + stride*class] = 1 - output[index + stride*class];
if(avg_cat) *avg_cat += output[index + stride*class];
return;
}
// multi-class
for(n = 0; n < classes; ++n){
delta[index + stride*n] = ((n == class)?1 : 0) - output[index + stride*n];
if(n == class && avg_cat) *avg_cat += output[index + stride*n];
}
}
void forward_yolo_layer(const layer l, network net)
{
int i,j,b,t,n;
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); // obj
index = entry_index(l, b, n*l.w*l.h, 4);
activate_array(l.output + index, (1+l.classes)*l.w*l.h, LOGISTIC); // classes
}
}
#endif
memset(l.delta, 0, l.outputs * l.batch * sizeof(float));
if(!net.train) return;
// 以上是测试
float avg_iou = 0;
float recall = 0;
float recall75 = 0;
float avg_cat = 0;
float avg_obj = 0;
float avg_anyobj = 0;
int count = 0;
int class_count = 0;
*(l.cost) = 0;
for (b = 0; b < l.batch; ++b) {
for (j = 0; j < l.h; ++j) {
for (i = 0; i < l.w; ++i) {
for (n = 0; n < l.n; ++n) {
int box_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0);
box pred = get_yolo_box(l.output, l.biases, l.mask[n],
box_index, i, j,
l.w, l.h,
net.w, net.h,
l.w*l.h);
float best_iou = 0;
int best_t = 0;
for(t = 0; t < l.max_boxes; ++t){
box truth = float_to_box(net.truth + t*(4 + 1) + b*l.truths, 1);
if(!truth.x) break;
float iou = box_iou(pred, truth);
if (iou > best_iou) {
best_iou = iou;
best_t = t;
}
}
// obj
int obj_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4);
avg_anyobj += l.output[obj_index];
// Negative
l.delta[obj_index] = 0 - l.output[obj_index];
// 如果大于igonre_thresh就不参与训练
if (best_iou > l.ignore_thresh) {
l.delta[obj_index] = 0;
}
// 进一步大于truth_thresh，参与训练
if (best_iou > l.truth_thresh) {
l.delta[obj_index] = 1 - l.output[obj_index];
// groudtruth的类型
int class = net.truth[best_t*(4 + 1) + b*l.truths + 4];
if (l.map) class = l.map[class];
// 网络预测类型的index
int class_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4 + 1);
delta_yolo_class(l.output, l.delta, class_index, class, l.classes, l.w*l.h, 0);
box truth = float_to_box(net.truth + best_t*(4 + 1) + b*l.truths, 1);
delta_yolo_box(truth, l.output, l.biases, l.mask[n], box_index,
i, j, l.w, l.h, net.w, net.h, l.delta,
(2-truth.w*truth.h), l.w*l.h);
}
}
}
}
// 保证所有的目标都应该出现在预测中。
for(t = 0; t < l.max_boxes; ++t){
box truth = float_to_box(net.truth + t*(4 + 1) + b*l.truths, 1);
if(!truth.x) break;
float best_iou = 0;
int best_n = 0;
i = (truth.x * l.w);
j = (truth.y * l.h);
box truth_shift = truth;
truth_shift.x = truth_shift.y = 0;
// 搜索最合适的anchor
for(n = 0; n < l.total; ++n){
box pred = {0};
pred.w = l.biases[2*n]/net.w;
pred.h = l.biases[2*n+1]/net.h;
float iou = box_iou(pred, truth_shift);
if (iou > best_iou){
best_iou = iou;
best_n = n;
}
}
int mask_n = int_index(l.mask, best_n, l.n);
// 如果最合适的anchor由本层负责预测（由mask来决定）
if(mask_n >= 0){
// 类似上面的工作。
int box_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 0);
float iou = delta_yolo_box(truth, l.output, l.biases,
best_n, box_index,
i, j, l.w, l.h, net.w, net.h,
l.delta,
(2-truth.w*truth.h),
l.w*l.h);
int obj_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4);
avg_obj += l.output[obj_index];
l.delta[obj_index] = 1 - l.output[obj_index];
int class = net.truth[t*(4 + 1) + b*l.truths + 4];
if (l.map) class = l.map[class];
int class_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4 + 1);
delta_yolo_class(l.output, l.delta, class_index, class, l.classes, l.w*l.h, &avg_cat);
++count;
++class_count;
if(iou > .5) recall += 1;
if(iou > .75) recall75 += 1;
avg_iou += iou;
}
}
}
// MSE loss
*(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);
printf("Region %d Avg IOU: %f, Class: %f, Obj: %f, No Obj: %f, .5R: %f, .75R: %f, count: %d\n", net.index, avg_iou/count, avg_cat/class_count, avg_obj/count, avg_anyobj/(l.w*l.h*l.n*l.batch), recall/count, recall75/count, count);
}
就一句， 相当于直接拷贝之前的delta了。
axpy_cpu(l.batch*l.inputs, 1, l.delta, 1, net.delta, 1);