更多darknet学习笔记,见darknet学习笔记
背景描述:
工业设备为了实时性和节省空间,保存图片往往时单通道二进制文件,目前darknet训练不支持单通道二进制格式,修改其代码使其可以直接运用公司设备。
代码详解:
1.main入口
darknet.c是darknet框架入口
mian函数中调用:
} else if (0 == strcmp(argv[1], "classifier")){
run_classifier(argc, argv);
之所以要看这段代码,因为darknet框架图像分类指令为:
cmd模式,在x64路径下——
训练:darknet classifier train data/METAL/metal.data data/METAL/darknet19_448.cfg pretrained_weights/darknet19_448.conv.23
验证:darknet classifier valid data/METAL/metal.data data/METAL/darknet19_448.cfg backup/_darknet19_448_final.weights
预测:darknet classifier test data/METAL/metal.data data/METAL/darknet19_448.cfg backup/_darknet19_448_final.weights
当第二个参数等于classifier时调用run_classifier()
2.图像分类函数run_classifier()
classifier.c文件下的run_classifier()函数:
if(0==strcmp(argv[2], "predict")) predict_classifier(data, cfg, weights, filename, top);
else if(0==strcmp(argv[2], "try")) try_classifier(data, cfg, weights, filename, atoi(layer_s));
else if(0==strcmp(argv[2], "train")) train_classifier(data, cfg, weights, gpus, ngpus, clear, dont_show, mjpeg_port, calc_topk);
else if(0==strcmp(argv[2], "demo")) demo_classifier(data, cfg, weights, cam_index, filename);
else if(0==strcmp(argv[2], "gun")) gun_classifier(data, cfg, weights, cam_index, filename);
else if(0==strcmp(argv[2], "threat")) threat_classifier(data, cfg, weights, cam_index, filename);
else if(0==strcmp(argv[2], "test")) test_classifier(data, cfg, weights, layer);
else if(0==strcmp(argv[2], "label")) label_classifier(data, cfg, weights);
else if(0==strcmp(argv[2], "valid")) validate_classifier_single(data, cfg, weights, NULL, -1);
else if(0==strcmp(argv[2], "validmulti")) validate_classifier_multi(data, cfg, weights);
else if(0==strcmp(argv[2], "valid10")) validate_classifier_10(data, cfg, weights);
else if(0==strcmp(argv[2], "validcrop")) validate_classifier_crop(data, cfg, weights);
else if(0==strcmp(argv[2], "validfull")) validate_classifier_full(data, cfg, weights);
图像分类的指令可以包括:predict(预测),try,train(训练),demo,gun,threat,test(测试),label,valid(验证),validmulti,valid10,validcrop,validfull。
这里我们只是想修改训练图像分类部分代码,所以只看train_classifier():
else if(0==strcmp(argv[2], "train")) train_classifier(data, cfg, weights, gpus, ngpus, clear, dont_show, mjpeg_port, calc_topk);
3.训练分类图片函数train_classifier()
classifier.c文件中train_classifier()函数源代码:
void train_classifier(char *datacfg, char *cfgfile, char *weightfile, int *gpus, int ngpus, int clear, int dont_show, int mjpeg_port, int calc_topk)
{
int i;
float avg_loss = -1;
char *base = basecfg(cfgfile);
printf("%s\n", base);
printf("%d\n", ngpus);
network* nets = (network*)calloc(ngpus, sizeof(network));
srand(time(0));
int seed = rand();
for(i = 0; i < ngpus; ++i){
srand(seed);
#ifdef GPU
cuda_set_device(gpus[i]);
#endif
nets[i] = parse_network_cfg(cfgfile);
if(weightfile){
load_weights(&nets[i], weightfile);
}
if(clear) *nets[i].seen = 0;
nets[i].learning_rate *= ngpus;
}
srand(time(0));
network net = nets[0];
int imgs = net.batch * net.subdivisions * ngpus;
printf("Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay);
list *options = read_data_cfg(datacfg);
char *backup_directory = option_find_str(options, "backup", "/backup/");
char *label_list = option_find_str(options, "labels", "data/labels.list");
char *train_list = option_find_str(options, "train", "data/train.list");
int classes = option_find_int(options, "classes", 2);
char **labels = get_labels(label_list);
list *plist = get_paths(train_list);
char **paths = (char **)list_to_array(plist);
printf("%d\n", plist->size);
int train_images_num = plist->size;
clock_t time;
load_args args = {0};
args.w = net.w;
args.h = net.h;
args.threads = 32;
args.hierarchy = net.hierarchy;
args.min = net.min_crop;
args.max = net.max_crop;
args.flip = net.flip;
args.angle = net.angle;
args.aspect = net.aspect;
args.exposure = net.exposure;
args.saturation = net.saturation;
args.hue = net.hue;
args.size = net.w > net.h ? net.w : net.h;
args.paths = paths;
args.classes = classes;
args.n = imgs;
args.m = train_images_num;
args.labels = labels;
args.type = CLASSIFICATION_DATA;
#ifdef OPENCV
//args.threads = 3;
mat_cv* img = NULL;
float max_img_loss = 10;
int number_of_lines = 100;
int img_size = 1000;
img = draw_train_chart(max_img_loss, net.max_batches, number_of_lines, img_size, dont_show);
#endif //OPENCV
data train;
data buffer;
pthread_t load_thread;
args.d = &buffer;
load_thread = load_data(args);
int iter_save = get_current_batch(net);
int iter_save_last = get_current_batch(net);
int iter_topk = get_current_batch(net);
float topk = 0;
while(get_current_batch(net) < net.max_batches || net.max_batches == 0){
time=clock();
pthread_join(load_thread, 0);
train = buffer;
load_thread = load_data(args);
printf("Loaded: %lf seconds\n", sec(clock()-time));
time=clock();
float loss = 0;
#ifdef GPU
if(ngpus == 1){
loss = train_network(net, train);
} else {
loss = train_networks(nets, ngpus, train, 4);
}
#else
loss = train_network(net, train);
#endif
if(avg_loss == -1 || isnan(avg_loss) || isinf(avg_loss)) avg_loss = loss;
avg_loss = avg_loss*.9 + loss*.1;
i = get_current_batch(net);
int calc_topk_for_each = iter_topk + 2 * train_images_num / (net.batch * net.subdivisions); // calculate TOPk for each 2 Epochs
calc_topk_for_each = fmax(calc_topk_for_each, net.burn_in);
calc_topk_for_each = fmax(calc_topk_for_each, 100);
if (i % 10 == 0) {
if (calc_topk) {
fprintf(stderr, "\n (next TOP5 calculation at %d iterations) ", calc_topk_for_each);
if (topk > 0) fprintf(stderr, " Last accuracy TOP5 = %2.2f %% \n", topk * 100);
}
if (net.cudnn_half) {
if (i < net.burn_in * 3) fprintf(stderr, " Tensor Cores are disabled until the first %d iterations are reached.\n", 3 * net.burn_in);
else fprintf(stderr, " Tensor Cores are used.\n");
}
}
int draw_precision = 0;
if (calc_topk && (i >= calc_topk_for_each || i == net.max_batches)) {
iter_topk = i;
topk = validate_classifier_single(datacfg, cfgfile, weightfile, &net, 5); // calc TOP5
printf("\n accuracy TOP5 = %f \n", topk);
draw_precision = 1;
}
printf("%d, %.3f: %f, %f avg, %f rate, %lf seconds, %ld images\n", get_current_batch(net), (float)(*net.seen)/ train_images_num, loss, avg_loss, get_current_rate(net), sec(clock()-time), *net.seen);
#ifdef OPENCV
draw_train_loss(img, img_size, avg_loss, max_img_loss, i, net.max_batches, topk, draw_precision, "top5", dont_show, mjpeg_port);
#endif // OPENCV
if (i >= (iter_save + 1000)) {
iter_save = i;
#ifdef GPU
if (ngpus != 1) sync_nets(nets, ngpus, 0);
#endif
char buff[256];
sprintf(buff, "%s/%s_%d.weights", backup_directory, base, i);
save_weights(net, buff);
}
if (i >= (iter_save_last + 100)) {
iter_save_last = i;
#ifdef GPU
if (ngpus != 1) sync_nets(nets, ngpus, 0);
#endif
char buff[256];
sprintf(buff, "%s/%s_last.weights", backup_directory, base);
save_weights(net, buff);
}
free_data(train);
}
#ifdef GPU
if (ngpus != 1) sync_nets(nets, ngpus, 0);
#endif
char buff[256];
sprintf(buff, "%s/%s_final.weights", backup_directory, base);
save_weights(net, buff);
#ifdef OPENCV
release_mat(&img);
destroy_all_windows_cv();
#endif
free_network(net);
free_ptrs((void**)labels, classes);
free_ptrs((void**)paths, plist->size);
free_list(plist);
free(base);
}
3.1train_classifier()函数调用参数说明: