github
/
darknet
mirror of https://github.com/AlexeyAB/darknet.git
1
0
Fork 0
Code Issues Releases Wiki Activity

Added label_smooth_eps=0.1 for [yolo] and [Gaussian_yolo] layer for Label Smoothing

Browse Source
pull/7577/head
AlexeyAB 6 years ago
parent 67e94dc1de
commit 318919e1cb
5 changed files with 16 additions and 13 deletions
Whitespace
Show all changes Ignore whitespace when comparing lines Ignore changes in amount of whitespace Ignore changes in whitespace at EOL
Unified View
Diff Options
Show Stats Download Patch File Download Diff File
  1. 1
      include/darknet.h
  2. 2
      src/detector.c
  3. 12
      src/gaussian_yolo_layer.c
  4. 2
      src/parser.c
  5. 12
      src/yolo_layer.c

1
include/darknet.h
Unescape View File

@ -257,6 +257,7 @@ struct layer {
float learning_rate_scale; float learning_rate_scale;
float clip; float clip;
int focal_loss; int focal_loss;
float label_smooth_eps;
int noloss; int noloss;
int softmax; int softmax;
int classes; int classes;

2
src/detector.c
Unescape View File

@ -141,7 +141,7 @@ void train_detector(char *datacfg, char *cfgfile, char *weightfile, int *gpus, i
args.show_imgs = show_imgs; args.show_imgs = show_imgs;
#ifdef OPENCV #ifdef OPENCV
args.threads = 3 * ngpus; // Amazon EC2 Tesla V100: p3.2xlarge (8 logical cores) - p3.16xlarge args.threads = 6 * ngpus; // 3 for - Amazon EC2 Tesla V100: p3.2xlarge (8 logical cores) - p3.16xlarge
//args.threads = 12 * ngpus; // Ryzen 7 2700X (16 logical cores) //args.threads = 12 * ngpus; // Ryzen 7 2700X (16 logical cores)
mat_cv* img = NULL; mat_cv* img = NULL;
float max_img_loss = 5; float max_img_loss = 5;

12
src/gaussian_yolo_layer.c
Unescape View File

@ -330,16 +330,16 @@ void averages_gaussian_yolo_deltas(int class_index, int box_index, int stride, i
} }
} }
void delta_gaussian_yolo_class(float *output, float *delta, int index, int class_id, int classes, int stride, float *avg_cat) void delta_gaussian_yolo_class(float *output, float *delta, int index, int class_id, int classes, int stride, float *avg_cat, float label_smooth_eps)
{ {
int n; int n;
if (delta[index]){ if (delta[index]){
delta[index + stride*class_id] = 1 - output[index + stride*class_id]; delta[index + stride*class_id] = (1 - label_smooth_eps) - output[index + stride*class_id];
if(avg_cat) *avg_cat += output[index + stride*class_id]; if(avg_cat) *avg_cat += output[index + stride*class_id];
return; return;
} }
for(n = 0; n < classes; ++n){ for(n = 0; n < classes; ++n){
delta[index + stride*n] = ((n == class_id)?1 : 0) - output[index + stride*n]; delta[index + stride*n] = ((n == class_id) ? (1 - label_smooth_eps) : (0 + label_smooth_eps/classes)) - output[index + stride*n];
if(n == class_id && avg_cat) *avg_cat += output[index + stride*n]; if(n == class_id && avg_cat) *avg_cat += output[index + stride*n];
} }
} }
@ -452,7 +452,7 @@ void forward_gaussian_yolo_layer(const layer l, network_state state)
int class_id = state.truth[best_t*(4 + 1) + b*l.truths + 4]; int class_id = state.truth[best_t*(4 + 1) + b*l.truths + 4];
if (l.map) class_id = l.map[class_id]; if (l.map) class_id = l.map[class_id];
int class_index = entry_gaussian_index(l, b, n*l.w*l.h + j*l.w + i, 9); int class_index = entry_gaussian_index(l, b, n*l.w*l.h + j*l.w + i, 9);
delta_gaussian_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, 0); delta_gaussian_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, 0, l.label_smooth_eps);
box truth = float_to_box_stride(state.truth + best_t*(4 + 1) + b*l.truths, 1); box truth = float_to_box_stride(state.truth + best_t*(4 + 1) + b*l.truths, 1);
delta_gaussian_yolo_box(truth, l.output, l.biases, l.mask[n], box_index, i, j, l.w, l.h, state.net.w, state.net.h, l.delta, (2-truth.w*truth.h), l.w*l.h, l.iou_normalizer, l.iou_loss, l.uc_normalizer, 1, l.yolo_point); delta_gaussian_yolo_box(truth, l.output, l.biases, l.mask[n], box_index, i, j, l.w, l.h, state.net.w, state.net.h, l.delta, (2-truth.w*truth.h), l.w*l.h, l.iou_normalizer, l.iou_loss, l.uc_normalizer, 1, l.yolo_point);
} }
@ -506,7 +506,7 @@ void forward_gaussian_yolo_layer(const layer l, network_state state)
int class_id = state.truth[t*(4 + 1) + b*l.truths + 4]; int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
if (l.map) class_id = l.map[class_id]; if (l.map) class_id = l.map[class_id];
int class_index = entry_gaussian_index(l, b, mask_n*l.w*l.h + j*l.w + i, 9); int class_index = entry_gaussian_index(l, b, mask_n*l.w*l.h + j*l.w + i, 9);
delta_gaussian_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat); delta_gaussian_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat, l.label_smooth_eps);
++count; ++count;
++class_count; ++class_count;
@ -537,7 +537,7 @@ void forward_gaussian_yolo_layer(const layer l, network_state state)
int class_id = state.truth[t*(4 + 1) + b*l.truths + 4]; int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
if (l.map) class_id = l.map[class_id]; if (l.map) class_id = l.map[class_id];
int class_index = entry_gaussian_index(l, b, mask_n*l.w*l.h + j*l.w + i, 9); int class_index = entry_gaussian_index(l, b, mask_n*l.w*l.h + j*l.w + i, 9);
delta_gaussian_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat); delta_gaussian_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat, l.label_smooth_eps);
++count; ++count;
++class_count; ++class_count;

2
src/parser.c
Unescape View File

@ -365,6 +365,7 @@ layer parse_yolo(list *options, size_params params)
} }
//assert(l.outputs == params.inputs); //assert(l.outputs == params.inputs);
l.label_smooth_eps = option_find_float_quiet(options, "label_smooth_eps", 0.0f);
l.scale_x_y = option_find_float_quiet(options, "scale_x_y", 1); l.scale_x_y = option_find_float_quiet(options, "scale_x_y", 1);
l.iou_normalizer = option_find_float_quiet(options, "iou_normalizer", 0.75); l.iou_normalizer = option_find_float_quiet(options, "iou_normalizer", 0.75);
l.cls_normalizer = option_find_float_quiet(options, "cls_normalizer", 1); l.cls_normalizer = option_find_float_quiet(options, "cls_normalizer", 1);
@ -456,6 +457,7 @@ layer parse_gaussian_yolo(list *options, size_params params) // Gaussian_YOLOv3
} }
//assert(l.outputs == params.inputs); //assert(l.outputs == params.inputs);
l.label_smooth_eps = option_find_float_quiet(options, "label_smooth_eps", 0.0f);
l.scale_x_y = option_find_float_quiet(options, "scale_x_y", 1); l.scale_x_y = option_find_float_quiet(options, "scale_x_y", 1);
l.uc_normalizer = option_find_float_quiet(options, "uc_normalizer", 1.0); l.uc_normalizer = option_find_float_quiet(options, "uc_normalizer", 1.0);
l.iou_normalizer = option_find_float_quiet(options, "iou_normalizer", 0.75); l.iou_normalizer = option_find_float_quiet(options, "iou_normalizer", 0.75);

12
src/yolo_layer.c
Unescape View File

@ -217,11 +217,11 @@ void averages_yolo_deltas(int class_index, int box_index, int stride, int classe
} }
} }
void delta_yolo_class(float *output, float *delta, int index, int class_id, int classes, int stride, float *avg_cat, int focal_loss) void delta_yolo_class(float *output, float *delta, int index, int class_id, int classes, int stride, float *avg_cat, int focal_loss, float label_smooth_eps)
{ {
int n; int n;
if (delta[index + stride*class_id]){ if (delta[index + stride*class_id]){
delta[index + stride*class_id] = 1 - output[index + stride*class_id]; delta[index + stride*class_id] = (1 - label_smooth_eps) - output[index + stride*class_id];
if(avg_cat) *avg_cat += output[index + stride*class_id]; if(avg_cat) *avg_cat += output[index + stride*class_id];
return; return;
} }
@ -248,7 +248,7 @@ void delta_yolo_class(float *output, float *delta, int index, int class_id, int
else { else {
// default // default
for (n = 0; n < classes; ++n) { for (n = 0; n < classes; ++n) {
delta[index + stride*n] = ((n == class_id) ? 1 : 0) - output[index + stride*n]; delta[index + stride*n] = ((n == class_id) ? (1 - label_smooth_eps) : (0 + label_smooth_eps/classes)) - output[index + stride*n];
if (n == class_id && avg_cat) *avg_cat += output[index + stride*n]; if (n == class_id && avg_cat) *avg_cat += output[index + stride*n];
} }
} }
@ -359,7 +359,7 @@ void forward_yolo_layer(const layer l, network_state state)
int class_id = state.truth[best_t*(4 + 1) + b*l.truths + 4]; int class_id = state.truth[best_t*(4 + 1) + b*l.truths + 4];
if (l.map) class_id = l.map[class_id]; if (l.map) class_id = l.map[class_id];
int class_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4 + 1); 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_id, l.classes, l.w*l.h, 0, l.focal_loss); delta_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, 0, l.focal_loss, l.label_smooth_eps);
box truth = float_to_box_stride(state.truth + best_t*(4 + 1) + b*l.truths, 1); box truth = float_to_box_stride(state.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, state.net.w, state.net.h, l.delta, (2 - truth.w*truth.h), l.w*l.h, l.iou_normalizer, l.iou_loss, 1); delta_yolo_box(truth, l.output, l.biases, l.mask[n], box_index, i, j, l.w, l.h, state.net.w, state.net.h, l.delta, (2 - truth.w*truth.h), l.w*l.h, l.iou_normalizer, l.iou_loss, 1);
} }
@ -421,7 +421,7 @@ void forward_yolo_layer(const layer l, network_state state)
int class_id = state.truth[t*(4 + 1) + b*l.truths + 4]; int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
if (l.map) class_id = l.map[class_id]; if (l.map) class_id = l.map[class_id];
int class_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4 + 1); 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_id, l.classes, l.w*l.h, &avg_cat, l.focal_loss); delta_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat, l.focal_loss, l.label_smooth_eps);
++count; ++count;
++class_count; ++class_count;
@ -463,7 +463,7 @@ void forward_yolo_layer(const layer l, network_state state)
int class_id = state.truth[t*(4 + 1) + b*l.truths + 4]; int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
if (l.map) class_id = l.map[class_id]; if (l.map) class_id = l.map[class_id];
int class_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 4 + 1); 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_id, l.classes, l.w*l.h, &avg_cat, l.focal_loss); delta_yolo_class(l.output, l.delta, class_index, class_id, l.classes, l.w*l.h, &avg_cat, l.focal_loss, l.label_smooth_eps);
++count; ++count;
++class_count; ++class_count;

Write Preview
Loading…
Cancel
Save