Skip to content

Commit

Permalink
Merge pull request #17 from fenrir-z/dev
Browse files Browse the repository at this point in the history 
mAP and class APs
  • Loading branch information
@phoenix-xhuang
phoenix-xhuang authored Jan 28, 2022
2 parents 9f4d308 + f441a7d commit cabc76a
Show file tree
Hide file tree
Showing 3 changed files with 63 additions and 24 deletions.
4 changes: 3 additions & 1 deletion det-yolov4-training/include/darknet.h
View file
Original file line number Diff line number Diff line change
Expand Up @@ -1045,7 +1045,9 @@ LIB_API detection *make_network_boxes(network *net, float thresh, int *num);
LIB_API void reset_rnn(network *net);
LIB_API float *network_predict_image(network *net, image im);
LIB_API float *network_predict_image_letterbox(network *net, image im);
LIB_API float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, float thresh_calc_avg_iou, const float iou_thresh, const int map_points, int letter_box, network *existing_net);
LIB_API float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, float thresh_calc_avg_iou,
const float iou_thresh, const int map_points, int letter_box, network *existing_net,
char **names, int names_size, double *aps, float best_map);
LIB_API void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, int ngpus, int clear, int dont_show, int calc_map, float thresh, float iou_thresh, int mjpeg_port, int show_imgs, int benchmark_layers, char* chart_path, int max_batches, char* task_id);
LIB_API 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);
Expand Down
71 changes: 53 additions & 18 deletions det-yolov4-training/src/detector.c
View file
Original file line number Diff line number Diff line change
Expand Up @@ -327,14 +327,6 @@ void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, i
printf("\n %d: %f, %f avg loss, %f rate, %lf seconds, %d images, %f hours left\n", iteration, loss, avg_loss, current_rate, (what_time_is_it_now() - time), iteration*imgs, avg_time);
fflush(stdout);

// write iteration, loss, avg loss, rate to train log yaml
FILE *file_ptr = fopen("/out/models/train-log.yaml", "w");
if (file_ptr != NULL) {
fprintf(file_ptr, "iteration: %d\nloss: %f\navg_loss: %f\nrate: %f\n", iteration, loss, avg_loss, current_rate);
fclose(file_ptr);
} // don't write if file open failed
// write train log yaml ends

int draw_precision = 0;
if (calc_map && (iteration >= next_map_calc || iteration == net.max_batches)) {
if (l.random) {
Expand Down Expand Up @@ -372,7 +364,14 @@ void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, i
//network net_combined = combine_train_valid_networks(net, net_map);

iter_map = iteration;
mean_average_precision = validate_detector_map(datacfg, cfgfile, weightfile, thresh, iou_thresh, 0, net.letter_box, &net_map);// &net_combined);
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);
double *aps = (double *)xcalloc(names_size, sizeof(double));
mean_average_precision = validate_detector_map(datacfg, cfgfile, weightfile, thresh, iou_thresh, 0,
net.letter_box, &net_map, names, names_size, aps, best_map);

printf("\n mean_average_precision (mAP@%0.2f) = %f \n", iou_thresh, mean_average_precision);
if (mean_average_precision > best_map) {
best_map = mean_average_precision;
Expand All @@ -383,6 +382,31 @@ void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, i
}

draw_precision = 1;

// write iteration, loss, avg loss, rate, average map to train log yaml
FILE *file_ptr = fopen("/out/models/train-log.yaml", "w");
if (file_ptr != NULL) {
fprintf(file_ptr, "iteration: %d\nloss: %f\navg_loss: %f\nrate: %f\nmap: %f\n", iteration, loss, avg_loss, current_rate, mean_average_precision);
if (names_size > 0) {
fprintf(file_ptr, "aps:\n");
for (int idx = 0; idx < names_size; ++idx) {
fprintf(file_ptr, " %s: %f\n", names[idx], aps[idx]);
}
}
fclose(file_ptr);
} // don't write if file open failed
// write train log yaml ends

free_ptrs((void **)names, names_size);
free(aps);
} else {
// write iteration, loss, avg loss, rate to train log yaml
FILE *file_ptr = fopen("/out/models/train-log.yaml", "w");
if (file_ptr != NULL) {
fprintf(file_ptr, "iteration: %d\nloss: %f\navg_loss: %f\nrate: %f\n", iteration, loss, avg_loss, current_rate);
fclose(file_ptr);
} // don't write if file open failed
// write train log yaml ends
}
time_remaining = ((net.max_batches - iteration) / ngpus)*(what_time_is_it_now() - time + load_time) / 60 / 60;
// set initial value, even if resume training from 10000 iteration
Expand Down Expand Up @@ -954,15 +978,14 @@ int detections_comparator(const void *pa, const void *pb)
return 0;
}

float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, float thresh_calc_avg_iou, const float iou_thresh, const int map_points, int letter_box, network *existing_net)
float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, float thresh_calc_avg_iou,
const float iou_thresh, const int map_points, int letter_box, network *existing_net,
char **names, int names_size, double *aps, float best_map)
{
int j;
list *options = read_data_cfg(datacfg);
char *valid_images = option_find_str(options, "valid", "data/train.txt");
char *difficult_valid_images = option_find_str(options, "difficult", NULL);
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);
//char *mapf = option_find_str(options, "map", 0);
//int *map = 0;
//if (mapf) map = read_map(mapf);
Expand All @@ -987,8 +1010,8 @@ float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, floa
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);
printf("\n Error: number of names %d that isn't equal to classes=%d in the file %s \n",
names_size, net.layers[net.n - 1].classes, cfgfile);
getchar();
}
srand(time(0));
Expand Down Expand Up @@ -1368,6 +1391,10 @@ float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, floa
//printf("Precision = %1.2f, Recall = %1.2f, avg IOU = %2.2f%% \n\n", class_precision, class_recall, avg_iou_per_class[i]);

mean_average_precision += avg_precision;

if (aps) {
aps[i] = avg_precision;
}
}

const float cur_precision = (float)tp_for_thresh / ((float)tp_for_thresh + (float)fp_for_thresh);
Expand All @@ -1385,7 +1412,7 @@ float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, floa
else printf("used Area-Under-Curve for each unique Recall \n");

printf(" mean average precision (mAP@%0.6f) = %f, or %2.2f %% \n", iou_thresh, mean_average_precision, mean_average_precision * 100);
fprintf(file_handle, "map: '%0.6f'\n", mean_average_precision);
fprintf(file_handle, "map: '%0.6f'\n", (best_map > mean_average_precision? best_map: mean_average_precision));

fprintf(file_handle, "model:\n");
fprintf(file_handle, "- model-symbol.json\n");
Expand Down Expand Up @@ -1418,7 +1445,7 @@ float validate_detector_map(char *datacfg, char *cfgfile, char *weightfile, floa
if (reinforcement_fd != NULL) fclose(reinforcement_fd);

// free memory
free_ptrs((void**)names, net.layers[net.n - 1].classes);
// free_ptrs((void **) names, net.layers[net.n - 1].classes);
free_list_contents_kvp(options);
free_list(options);

Expand Down Expand Up @@ -2065,7 +2092,15 @@ void run_detector(int argc, char **argv)
else if (0 == strcmp(argv[2], "train")) train_detector(datacfg, cfg, weights, gpus, ngpus, clear, dont_show, calc_map, thresh, iou_thresh, mjpeg_port, show_imgs, benchmark_layers, chart_path, max_batches, task_id);
else if (0 == strcmp(argv[2], "valid")) validate_detector(datacfg, cfg, weights, outfile);
else if (0 == strcmp(argv[2], "recall")) validate_detector_recall(datacfg, cfg, weights);
else if (0 == strcmp(argv[2], "map")) validate_detector_map(datacfg, cfg, weights, thresh, iou_thresh, map_points, letter_box, NULL);
else if (0 == strcmp(argv[2], "map")) {
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);
validate_detector_map(datacfg, cfg, weights, thresh, iou_thresh, map_points, letter_box, NULL, names,
names_size, NULL, -1);
free_ptrs((void **)names, names_size);
}
else if (0 == strcmp(argv[2], "calc_anchors")) calc_anchors(datacfg, num_of_clusters, width, height, show);
else if (0 == strcmp(argv[2], "draw")) {
int it_num = 100;
Expand Down
12 changes: 7 additions & 5 deletions det-yolov4-training/train_watcher.py
View file
Original file line number Diff line number Diff line change
Expand Up @@ -66,12 +66,16 @@ def _on_train_log_yaml_modified(self, src_path: str) -> None:
loss = float(train_log_dict['loss'])
avg_loss = float(train_log_dict['avg_loss'])
rate = float(train_log_dict['rate'])

logging.debug(f"writing tensorboard: i: {iteration}, l: {loss}, a: {avg_loss}, r: {rate}")
mean_ap = train_log_dict.get('map', None)
class_aps = train_log_dict.get('aps', None)

self._tensorboard_writer.add_scalar(tag="train/loss", scalar_value=loss, global_step=iteration)
self._tensorboard_writer.add_scalar(tag="train/avg_loss", scalar_value=avg_loss, global_step=iteration)
self._tensorboard_writer.add_scalar(tag="train/rate", scalar_value=rate, global_step=iteration)
if isinstance(mean_ap, float):
self._tensorboard_writer.add_scalar(tag="train/mAP", scalar_value=mean_ap, global_step=iteration)
if class_aps and isinstance(class_aps, dict):
self._tensorboard_writer.add_scalars(main_tag='train/aps', tag_scalar_dict=class_aps, global_step=iteration)
self._tensorboard_writer.flush()

# protected: general
Expand All @@ -93,9 +97,7 @@ def __init__(self, model_dir: str, width: int, height: int, class_num: int) -> N
super().__init__()
self._model_dir = model_dir
self._observer = None
self._event_handler = _DarknetTrainingHandler(width=width,
height=height,
class_num=class_num)
self._event_handler = _DarknetTrainingHandler(width=width, height=height, class_num=class_num)

def start(self) -> None:
if self._observer:
Expand Down

0 comments on commit cabc76a

Please sign in to comment.