基于darknet框架的数据标注及训练流程
本文标注训练数据所使用的环境为Ubuntu20.04,CUDA 11.0, OpenCV 3.2, 训练则是在服务器上进行的。
主要使用的工具包括
开源标注工具Yolo_mark:
开源标注工具网址
红绿灯模型训练darknet框架:
darknet网址
具体实现如下:
标注
首先clone Yolo_mark工具到本地后,
标注过程:
- 将所需标注图片*.jpg文件添加到Yolo_mark相对路径:
x64/Release/data/img
目录下; - 在
x64/Release/data/obj.data
文件下更改classes数量; - 在
x64/Release/data/obj.names
文件中加入类别名称,每一行一个;
同时支持OPenCV2.x及OpenCV3.x
将上述所需信息更改后,便可在命令行中执行:
cmake.
make
./linux_mark.sh
Yolo_mark标注过程中,可以通过数字键选择类别;
键盘左右箭头选择前后照片;
h:开启help提示;
c:清除标记;
m:显示坐标;
p:拷贝上一张图片的标记结果;(一般标注连续帧图片时可采用,比较省事)
r:删除所选标注框;
ESC:退出标注。
训练过程
训练数据
在标注完数据后,生成训练数据后,可进行模型训练。
- 首先clone darknet: darknet
- config文件及脚本文件在
build\darknet\x64\data
文件中,具体可参考如何训练自己的数据 - 下载预训练权重,此处所给为darknet53.conv.74下载地址
- run:
./darknet detector train data/obj.data cfg/my_yolov3.cfg darknet53.conv.74 -dont_show
模型训练结束后,可在backup文件夹中找到训练所得的权重.weights文件。
验证模型
run:
./darknet detector test data/obj.data cfg/my_test.cfg backup/yourweights
另附批量检测图片并存储到指定目录方式:
- 用下面代码替换src文件夹下detector.c文件中的test_detector函数(可更改输出文件夹名,此处为result_img):
void test_detector(char *datacfg, char *cfgfile, char *weightfile, char *filename, float thresh,
float hier_thresh, int dont_show, int ext_output, int save_labels, char *outfile, int letter_box, int benchmark_layers)
{
list *options = read_data_cfg(datacfg);
char *name_list = option_find_str(options, "names", "data/names.list");
int names_size = 0;
char **names = get_labels_custom(name_list, &names_size); //get_labels(name_list);
image **alphabet = load_alphabet();
network net = parse_network_cfg_custom(cfgfile, 1, 1); // set batch=1
if (weightfile) {
load_weights(&net, weightfile);
}
net.benchmark_layers = benchmark_layers;
fuse_conv_batchnorm(net);
calculate_binary_weights(net);
if (net.layers[net.n - 1].classes != names_size) {
printf("\n Error: in the file %s number of names %d that isn't equal to classes=%d in the file %s \n",
name_list, names_size, net.layers[net.n - 1].classes, cfgfile);
if (net.layers[net.n - 1].classes > names_size) getchar();
}
srand(2222222);
char buff[256];
char *input = buff;
char *json_buf = NULL;
int json_image_id = 0;
FILE* json_file = NULL;
if (outfile) {
json_file = fopen(outfile, "wb");
if(!json_file) {
error("fopen failed");
}
char *tmp = "[\n";
fwrite(tmp, sizeof(char), strlen(tmp), json_file);
}
int j;
float nms = .45; // 0.4F
//开始循环//
int i;
while (1) {
if (filename) {
strncpy(input, filename, 256);
list *plist = get_paths(input);
char **paths = (char **)list_to_array(plist);
printf("Start Testing!\n");
int m = plist->size;
for (i = 0; i < m; ++i) {
char *path = paths[i];
image im = load_image(path, 0, 0, net.c);
int letterbox = 0;
image sized = resize_image(im, net.w, net.h);
//image sized = letterbox_image(im, net.w, net.h); letterbox = 1;
layer l = net.layers[net.n - 1];
float *X = sized.data;
double time = get_time_point();
network_predict(net, X);
printf("%s: Predicted in %lf milli-seconds.\n", input, ((double)get_time_point() - time) / 1000);
printf("Try Very Hard:");
printf("%s: Predicted in %lf milli-seconds.\n", path, ((double)get_time_point() - time) / 1000);
int nboxes = 0;
detection *dets = get_network_boxes(&net, im.w, im.h, thresh, hier_thresh, 0, 1, &nboxes, letterbox);
if (nms) do_nms_sort(dets, nboxes, l.classes, nms);
// draw_detections_v3(basecfg(input), im, dets, nboxes, thresh, names, alphabet, l.classes, ext_output);
draw_detections_v3(im, dets, nboxes, thresh, names, alphabet, l.classes, ext_output);
char b[2048];
sprintf(b, "result_img/%s", GetFilename(path)); //***改成自己的文件夹路径
save_image(im, b);
printf("save %s successfully!\n", GetFilename(path));//文件命名在这个地方,可改为i
if (save_labels)
{
char labelpath[4096];
replace_image_to_label(input, labelpath);
FILE* fw = fopen(labelpath, "wb");
int i;
for (i = 0; i < nboxes; ++i) {
char buff[1024];
int class_id = -1;
float prob = 0;
for (j = 0; j < l.classes; ++j) {
if (dets[i].prob[j] > thresh && dets[i].prob[j] > prob) {
prob = dets[i].prob[j];
class_id = j;
}
}
if (class_id >= 0) {
sprintf(buff, "%d %2.4f %2.4f %2.4f %2.4f\n", class_id, dets[i].bbox.x, dets[i].bbox.y, dets[i].bbox.w, dets[i].bbox.h);
fwrite(buff, sizeof(char), strlen(buff), fw);
}
}
fclose(fw);
}
free_detections(dets, nboxes);
free_image(im);
free_image(sized);
}
printf("All Done!\n");
exit(0);
}
else {
//这里这个if可以删掉,为了更好地看到哪里改动了,这块我就没有删。
if (filename) {
strncpy(input, filename, 256);
if (strlen(input) > 0)
if (input[strlen(input) - 1] == 0x0d) input[strlen(input) - 1] = 0;
}
else {
printf("Enter Image Path: ");
fflush(stdout);
input = fgets(input, 256, stdin);
if (!input) break;
strtok(input, "\n");
}
//image im;
//image sized = load_image_resize(input, net.w, net.h, net.c, &im);
image im = load_image(input, 0, 0, net.c);
image sized;
if (letter_box) sized = letterbox_image(im, net.w, net.h);
else sized = resize_image(im, net.w, net.h);
layer l = net.layers[net.n - 1];
//box *boxes = calloc(l.w*l.h*l.n, sizeof(box));
//float **probs = calloc(l.w*l.h*l.n, sizeof(float*));
//for(j = 0; j < l.w*l.h*l.n; ++j) probs[j] = (float*)xcalloc(l.classes, sizeof(float));
float *X = sized.data;
//time= what_time_is_it_now();
double time = get_time_point();
network_predict(net, X);
//network_predict_image(&net, im); letterbox = 1;
printf("%s: Predicted in %lf milli-seconds.\n", input, ((double)get_time_point() - time) / 1000);
//printf("%s: Predicted in %f seconds.\n", input, (what_time_is_it_now()-time));
int nboxes = 0;
detection *dets = get_network_boxes(&net, im.w, im.h, thresh, hier_thresh, 0, 1, &nboxes, letter_box);
if (nms) {
if (l.nms_kind == DEFAULT_NMS) do_nms_sort(dets, nboxes, l.classes, nms);
else diounms_sort(dets, nboxes, l.classes, nms, l.nms_kind, l.beta_nms);
}
draw_detections_v3(im, dets, nboxes, thresh, names, alphabet, l.classes, ext_output);
save_image(im, "predictions");
if (!dont_show) {
show_image(im, "predictions");
}
if (json_file) {
if (json_buf) {
char *tmp = ", \n";
fwrite(tmp, sizeof(char), strlen(tmp), json_file);
}
++json_image_id;
json_buf = detection_to_json(dets, nboxes, l.classes, names, json_image_id, input);
fwrite(json_buf, sizeof(char), strlen(json_buf), json_file);
free(json_buf);
}
// pseudo labeling concept - fast.ai
if (save_labels)
{
char labelpath[4096];
replace_image_to_label(input, labelpath);
FILE* fw = fopen(labelpath, "wb");
int i;
for (i = 0; i < nboxes; ++i) {
char buff[1024];
int class_id = -1;
float prob = 0;
for (j = 0; j < l.classes; ++j) {
if (dets[i].prob[j] > thresh && dets[i].prob[j] > prob) {
prob = dets[i].prob[j];
class_id = j;
}
}
if (class_id >= 0) {
sprintf(buff, "%d %2.4f %2.4f %2.4f %2.4f\n", class_id, dets[i].bbox.x, dets[i].bbox.y, dets[i].bbox.w, dets[i].bbox.h);
fwrite(buff, sizeof(char), strlen(buff), fw);
}
}
fclose(fw);
}
free_detections(dets, nboxes);
free_image(im);
free_image(sized);
if (!dont_show) {
wait_until_press_key_cv();
destroy_all_windows_cv();
}
if (filename) break;
}
}
//这里之前的是循环部分//
if (json_file) {
char *tmp = "\n]";
fwrite(tmp, sizeof(char), strlen(tmp), json_file);
fclose(json_file);
}
// free memory
free_ptrs((void**)names, net.layers[net.n - 1].classes);
free_list_contents_kvp(options);
free_list(options);
const int nsize = 8;
for (j = 0; j < nsize; ++j) {
for (i = 32; i < 127; ++i) {
free_image(alphabet[j][i]);
}
free(alphabet[j]);
}
free(alphabet);
free_network(net);
}
- 在前面添加GetFilename函数:
//change pic name
char *GetFilename(char *p)
{
static char name_[40] = { "" };
static char name[40] = { "" };
char *q = strrchr(p, '/') + 1;
strncpy(name_, q, 40);//后面的40是图片名长度(不包括后缀),根据自己的需要进行修改
int len=strlen(name_);
strncpy(name,name_,len-4);
return name;
}
- 在darknet下重新make(一定要记得!)
- 执行测试命令:
./darknet detector test data/obj.data cfg/my_test.cfg backup/yourweights data/test.txt -dont_show
注意上述指令中 test.txt文件存储的即是所需测试的图片。
获得方式可见另一篇博客:
好家伙,联动了
评估模型:
构造验证集 valid dataset,then run:
./darknet detector map data/obj.data cfg/yourcfg backup/yourweights -points 0 -thresh 0.5 -iou_thresh 0.5
可以得到模型的MAP,precision,recall和F1-score值。