| | |
| | | l.outputs = h*w*n*(classes + 4 + 1);
|
| | | l.inputs = l.outputs;
|
| | | l.max_boxes = max_boxes;
|
| | | l.truths = l.max_boxes*(4 + 1); // 90*(4 + 1);
|
| | | l.truth_size = 4 + 2; |
| | | l.truths = l.max_boxes*l.truth_size; // 90*(4 + 1); |
| | | l.labels = (int*)xcalloc(batch * l.w*l.h*l.n, sizeof(int)); |
| | | for (i = 0; i < batch * l.w*l.h*l.n; ++i) l.labels[i] = -1; |
| | | l.class_ids = (int*)xcalloc(batch * l.w*l.h*l.n, sizeof(int)); |
| | | for (i = 0; i < batch * l.w*l.h*l.n; ++i) l.class_ids[i] = -1; |
| | | |
| | | l.delta = (float*)xcalloc(batch * l.outputs, sizeof(float));
|
| | | l.output = (float*)xcalloc(batch * l.outputs, sizeof(float));
|
| | | for(i = 0; i < total*2; ++i){
|
| | |
| | | l->outputs = h*w*l->n*(l->classes + 4 + 1);
|
| | | l->inputs = l->outputs;
|
| | |
|
| | | if (l->embedding_output) l->embedding_output = (float*)xrealloc(l->output, l->batch * l->embedding_size * l->n * l->h * l->w * sizeof(float)); |
| | | if (l->labels) l->labels = (int*)xrealloc(l->labels, l->batch * l->n * l->h * l->w * sizeof(int)); |
| | | if (l->class_ids) l->class_ids = (int*)xrealloc(l->class_ids, l->batch * l->n * l->h * l->w * sizeof(int)); |
| | | |
| | | if (!l->output_pinned) l->output = (float*)xrealloc(l->output, l->batch*l->outputs * sizeof(float));
|
| | | if (!l->delta_pinned) l->delta = (float*)xrealloc(l->delta, l->batch*l->outputs*sizeof(float));
|
| | |
|
| | |
| | | #endif
|
| | | }
|
| | |
|
| | | 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 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, int new_coords) |
| | | {
|
| | | box b;
|
| | | // ln - natural logarithm (base = e)
|
| | |
| | | // y` = t.y * lh - i; // y = ln(y`/(1-y`)) // y - output of previous conv-layer
|
| | | // w = ln(t.w * net.w / anchors_w); // w - output of previous conv-layer
|
| | | // h = ln(t.h * net.h / anchors_h); // h - output of previous conv-layer
|
| | | if (new_coords) { |
| | | b.x = (i + x[index + 0 * stride]) / lw; |
| | | b.y = (j + x[index + 1 * stride]) / lh; |
| | | b.w = x[index + 2 * stride] * x[index + 2 * stride] * 4 * biases[2 * n] / w; |
| | | b.h = x[index + 3 * stride] * x[index + 3 * stride] * 4 * biases[2 * n + 1] / h; |
| | | } |
| | | else { |
| | | b.x = (i + x[index + 0*stride]) / lw;
|
| | | b.y = (j + x[index + 1*stride]) / lh;
|
| | | 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;
|
| | | }
|
| | |
|
| | |
| | | return val;
|
| | | }
|
| | |
|
| | | ious 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, float iou_normalizer, IOU_LOSS iou_loss, int accumulate, float max_delta)
|
| | | ious 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, float iou_normalizer, IOU_LOSS iou_loss, int accumulate, float max_delta, int *rewritten_bbox, int new_coords) |
| | | {
|
| | | if (delta[index + 0 * stride] || delta[index + 1 * stride] || delta[index + 2 * stride] || delta[index + 3 * stride]) { |
| | | (*rewritten_bbox)++; |
| | | } |
| | | |
| | | ious all_ious = { 0 };
|
| | | // i - step in layer width
|
| | | // j - step in layer height
|
| | | // Returns a box in absolute coordinates
|
| | | box pred = get_yolo_box(x, biases, n, index, i, j, lw, lh, w, h, stride);
|
| | | box pred = get_yolo_box(x, biases, n, index, i, j, lw, lh, w, h, stride, new_coords); |
| | | all_ious.iou = box_iou(pred, truth);
|
| | | all_ious.giou = box_giou(pred, truth);
|
| | | all_ious.diou = box_diou(pred, truth);
|
| | |
| | | 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]);
|
| | | |
| | | if (new_coords) { |
| | | //tx = (truth.x*lw - i + 0.5) / 2; |
| | | //ty = (truth.y*lh - j + 0.5) / 2; |
| | | tw = sqrt(truth.w*w / (4 * biases[2 * n])); |
| | | th = sqrt(truth.h*h / (4 * biases[2 * n + 1])); |
| | | } |
| | |
|
| | | //printf(" tx = %f, ty = %f, tw = %f, th = %f \n", tx, ty, tw, th);
|
| | | //printf(" x = %f, y = %f, w = %f, h = %f \n", x[index + 0 * stride], x[index + 1 * stride], x[index + 2 * stride], x[index + 3 * stride]);
|
| | |
| | | float dw = all_ious.dx_iou.dl;
|
| | | float dh = all_ious.dx_iou.dr;
|
| | |
|
| | | |
| | | // predict exponential, apply gradient of e^delta_t ONLY for w,h
|
| | | if (new_coords) { |
| | | //dw *= 8 * x[index + 2 * stride]; |
| | | //dh *= 8 * x[index + 3 * stride]; |
| | | //dw *= 8 * x[index + 2 * stride] * biases[2 * n] / w; |
| | | //dh *= 8 * x[index + 3 * stride] * biases[2 * n + 1] / h; |
| | | |
| | | //float grad_w = 8 * exp(-x[index + 2 * stride]) / pow(exp(-x[index + 2 * stride]) + 1, 3); |
| | | //float grad_h = 8 * exp(-x[index + 3 * stride]) / pow(exp(-x[index + 3 * stride]) + 1, 3); |
| | | //dw *= grad_w; |
| | | //dh *= grad_h; |
| | | } |
| | | else { |
| | | dw *= exp(x[index + 2 * stride]);
|
| | | dh *= exp(x[index + 3 * stride]);
|
| | | } |
| | | |
| | | |
| | | //dw *= exp(x[index + 2 * stride]); |
| | | //dh *= exp(x[index + 3 * stride]); |
| | |
|
| | | // normalize iou weight
|
| | | dx *= iou_normalizer;
|
| | |
| | | }
|
| | | }
|
| | |
|
| | | 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, float *classes_multipliers)
|
| | | 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, float *classes_multipliers, float cls_normalizer) |
| | | {
|
| | | int n;
|
| | | if (delta[index + stride*class_id]){
|
| | |
| | | float result_delta = y_true - output[index + stride*n];
|
| | | if (!isnan(result_delta) && !isinf(result_delta)) delta[index + stride*n] = result_delta;
|
| | |
|
| | | if (classes_multipliers && n == class_id) delta[index + stride*class_id] *= classes_multipliers[class_id];
|
| | | if (classes_multipliers && n == class_id) delta[index + stride*class_id] *= classes_multipliers[class_id] * cls_normalizer; |
| | | if (n == class_id && avg_cat) *avg_cat += output[index + stride*n];
|
| | | }
|
| | | }
|
| | |
| | | return batch*l.outputs + n*l.w*l.h*(4+l.classes+1) + entry*l.w*l.h + loc;
|
| | | }
|
| | |
|
| | | void forward_yolo_layer(const layer l, network_state state)
|
| | | typedef struct train_yolo_args { |
| | | layer l; |
| | | network_state state; |
| | | int b; |
| | | |
| | | float tot_iou; |
| | | float tot_giou_loss; |
| | | float tot_iou_loss; |
| | | int count; |
| | | int class_count; |
| | | } train_yolo_args; |
| | | |
| | | void *process_batch(void* ptr) |
| | | {
|
| | | int i, j, b, t, n;
|
| | | memcpy(l.output, state.input, l.outputs*l.batch * sizeof(float));
|
| | | { |
| | | train_yolo_args *args = (train_yolo_args*)ptr; |
| | | const layer l = args->l; |
| | | network_state state = args->state; |
| | | int b = args->b; |
| | |
|
| | | #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); // x,y,
|
| | | scal_add_cpu(2 * l.w*l.h, l.scale_x_y, -0.5*(l.scale_x_y - 1), l.output + index, 1); // scale x,y
|
| | | index = entry_index(l, b, n*l.w*l.h, 4);
|
| | | activate_array(l.output + index, (1 + l.classes)*l.w*l.h, LOGISTIC);
|
| | | }
|
| | | }
|
| | | #endif
|
| | | int i, j, t, n; |
| | |
|
| | | // delta is zeroed
|
| | | memset(l.delta, 0, l.outputs * l.batch * sizeof(float));
|
| | | if (!state.train) return;
|
| | | //float avg_iou = 0;
|
| | | float tot_iou = 0;
|
| | | //printf(" b = %d \n", b, b); |
| | | |
| | | //float tot_iou = 0; |
| | | float tot_giou = 0;
|
| | | float tot_diou = 0;
|
| | | float tot_ciou = 0;
|
| | | float tot_iou_loss = 0;
|
| | | float tot_giou_loss = 0;
|
| | | //float tot_iou_loss = 0; |
| | | //float tot_giou_loss = 0; |
| | | float tot_diou_loss = 0;
|
| | | float tot_ciou_loss = 0;
|
| | | float recall = 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) {
|
| | | //int count = 0; |
| | | //int class_count = 0; |
| | | |
| | | for (j = 0; j < l.h; ++j) {
|
| | | for (i = 0; i < l.w; ++i) {
|
| | | for (n = 0; n < l.n; ++n) {
|
| | |
| | | const int obj_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4);
|
| | | const int box_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0);
|
| | | const int stride = l.w*l.h;
|
| | | box pred = get_yolo_box(l.output, l.biases, l.mask[n], box_index, i, j, l.w, l.h, state.net.w, state.net.h, l.w*l.h);
|
| | | box pred = get_yolo_box(l.output, l.biases, l.mask[n], box_index, i, j, l.w, l.h, state.net.w, state.net.h, l.w * l.h, l.new_coords); |
| | | float best_match_iou = 0;
|
| | | int best_match_t = 0;
|
| | | float best_iou = 0;
|
| | | int best_t = 0;
|
| | | for (t = 0; t < l.max_boxes; ++t) {
|
| | | box truth = float_to_box_stride(state.truth + t*(4 + 1) + b*l.truths, 1);
|
| | | int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
|
| | | box truth = float_to_box_stride(state.truth + t * l.truth_size + b * l.truths, 1); |
| | | if (!truth.x) break; // continue; |
| | | int class_id = state.truth[t * l.truth_size + b * l.truths + 4]; |
| | | if (class_id >= l.classes || class_id < 0) {
|
| | | printf("\n Warning: in txt-labels class_id=%d >= classes=%d in cfg-file. In txt-labels class_id should be [from 0 to %d] \n", class_id, l.classes, l.classes - 1);
|
| | | printf("\n truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f, class_id = %d \n", truth.x, truth.y, truth.w, truth.h, class_id);
|
| | | if (check_mistakes) getchar();
|
| | | continue; // if label contains class_id more than number of classes in the cfg-file and class_id check garbage value
|
| | | }
|
| | | if (!truth.x) break; // continue;
|
| | |
|
| | | float objectness = l.output[obj_index];
|
| | | if (isnan(objectness) || isinf(objectness)) l.output[obj_index] = 0;
|
| | |
| | | }
|
| | |
|
| | | avg_anyobj += l.output[obj_index];
|
| | | l.delta[obj_index] = l.cls_normalizer * (0 - l.output[obj_index]);
|
| | | l.delta[obj_index] = l.obj_normalizer * (0 - l.output[obj_index]); |
| | | if (best_match_iou > l.ignore_thresh) {
|
| | | const float iou_multiplier = best_match_iou*best_match_iou;// (best_match_iou - l.ignore_thresh) / (1.0 - l.ignore_thresh);
|
| | | if (l.objectness_smooth) {
|
| | | l.delta[obj_index] = l.cls_normalizer * (iou_multiplier - l.output[obj_index]);
|
| | | const float delta_obj = l.obj_normalizer * (best_match_iou - l.output[obj_index]); |
| | | if (delta_obj > l.delta[obj_index]) l.delta[obj_index] = delta_obj; |
| | |
|
| | | int class_id = state.truth[best_match_t*(4 + 1) + b*l.truths + 4];
|
| | | if (l.map) class_id = l.map[class_id];
|
| | | const float class_multiplier = (l.classes_multipliers) ? l.classes_multipliers[class_id] : 1.0f;
|
| | | l.delta[class_index + stride*class_id] = class_multiplier * (iou_multiplier - l.output[class_index + stride*class_id]);
|
| | | }
|
| | | else l.delta[obj_index] = 0;
|
| | | }
|
| | |
| | | int stride = l.w*l.h;
|
| | | float scale = pred.w * pred.h;
|
| | | if (scale > 0) scale = sqrt(scale);
|
| | | l.delta[obj_index] = scale * l.cls_normalizer * (0 - l.output[obj_index]);
|
| | | l.delta[obj_index] = scale * l.obj_normalizer * (0 - l.output[obj_index]); |
| | | int cl_id;
|
| | | int found_object = 0; |
| | | for (cl_id = 0; cl_id < l.classes; ++cl_id) {
|
| | | if(l.output[class_index + stride*cl_id] * l.output[obj_index] > 0.25)
|
| | | if (l.output[class_index + stride * cl_id] * l.output[obj_index] > 0.25) { |
| | | l.delta[class_index + stride*cl_id] = scale * (0 - l.output[class_index + stride*cl_id]);
|
| | | found_object = 1; |
| | | } |
| | | } |
| | | if (found_object) { |
| | | // don't use this loop for adversarial attack drawing |
| | | for (cl_id = 0; cl_id < l.classes; ++cl_id) |
| | | if (l.output[class_index + stride * cl_id] * l.output[obj_index] < 0.25) |
| | | l.delta[class_index + stride * cl_id] = scale * (1 - l.output[class_index + stride * cl_id]); |
| | | |
| | | l.delta[box_index + 0 * stride] += scale * (0 - l.output[box_index + 0 * stride]); |
| | | l.delta[box_index + 1 * stride] += scale * (0 - l.output[box_index + 1 * stride]); |
| | | l.delta[box_index + 2 * stride] += scale * (0 - l.output[box_index + 2 * stride]); |
| | | l.delta[box_index + 3 * stride] += scale * (0 - l.output[box_index + 3 * stride]); |
| | | }
|
| | | }
|
| | | if (best_iou > l.truth_thresh) {
|
| | | const float iou_multiplier = best_iou*best_iou;// (best_iou - l.truth_thresh) / (1.0 - l.truth_thresh);
|
| | | if (l.objectness_smooth) l.delta[obj_index] = l.cls_normalizer * (iou_multiplier - l.output[obj_index]);
|
| | | else l.delta[obj_index] = l.cls_normalizer * (1 - l.output[obj_index]);
|
| | | //l.delta[obj_index] = l.cls_normalizer * (1 - l.output[obj_index]);
|
| | | if (l.objectness_smooth) l.delta[obj_index] = l.obj_normalizer * (iou_multiplier - l.output[obj_index]); |
| | | else l.delta[obj_index] = l.obj_normalizer * (1 - l.output[obj_index]); |
| | | //l.delta[obj_index] = l.obj_normalizer * (1 - l.output[obj_index]); |
| | |
|
| | | int class_id = state.truth[best_t*(4 + 1) + b*l.truths + 4];
|
| | | int class_id = state.truth[best_t * l.truth_size + b * l.truths + 4]; |
| | | if (l.map) class_id = l.map[class_id];
|
| | | 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, l.classes_multipliers);
|
| | | 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, l.classes_multipliers, l.cls_normalizer); |
| | | const float class_multiplier = (l.classes_multipliers) ? l.classes_multipliers[class_id] : 1.0f;
|
| | | if (l.objectness_smooth) l.delta[class_index + stride*class_id] = class_multiplier * (iou_multiplier - l.output[class_index + stride*class_id]);
|
| | | 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 * class_multiplier, l.iou_loss, 1, l.max_delta);
|
| | | box truth = float_to_box_stride(state.truth + best_t * l.truth_size + 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 * class_multiplier, l.iou_loss, 1, l.max_delta, state.net.rewritten_bbox, l.new_coords); |
| | | (*state.net.total_bbox)++; |
| | | }
|
| | | }
|
| | | }
|
| | | }
|
| | | for (t = 0; t < l.max_boxes; ++t) {
|
| | | box truth = float_to_box_stride(state.truth + t*(4 + 1) + b*l.truths, 1);
|
| | | box truth = float_to_box_stride(state.truth + t * l.truth_size + b * l.truths, 1); |
| | | if (!truth.x) break; // continue; |
| | | if (truth.x < 0 || truth.y < 0 || truth.x > 1 || truth.y > 1 || truth.w < 0 || truth.h < 0) {
|
| | | char buff[256];
|
| | | printf(" Wrong label: truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f \n", truth.x, truth.y, truth.w, truth.h);
|
| | |
| | | truth.x, truth.y, truth.w, truth.h);
|
| | | system(buff);
|
| | | }
|
| | | int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
|
| | | int class_id = state.truth[t * l.truth_size + b * l.truths + 4]; |
| | | if (class_id >= l.classes || class_id < 0) continue; // if label contains class_id more than number of classes in the cfg-file and class_id check garbage value
|
| | |
|
| | | if (!truth.x) break; // continue;
|
| | | float best_iou = 0;
|
| | | int best_n = 0;
|
| | | i = (truth.x * l.w);
|
| | |
| | |
|
| | | int mask_n = int_index(l.mask, best_n, l.n);
|
| | | if (mask_n >= 0) {
|
| | | int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
|
| | | int class_id = state.truth[t * l.truth_size + b * l.truths + 4]; |
| | | if (l.map) class_id = l.map[class_id];
|
| | |
|
| | | int box_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 0);
|
| | | const float class_multiplier = (l.classes_multipliers) ? l.classes_multipliers[class_id] : 1.0f;
|
| | | ious all_ious = delta_yolo_box(truth, l.output, l.biases, best_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 * class_multiplier, l.iou_loss, 1, l.max_delta);
|
| | | ious all_ious = delta_yolo_box(truth, l.output, l.biases, best_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 * class_multiplier, l.iou_loss, 1, l.max_delta, state.net.rewritten_bbox, l.new_coords); |
| | | (*state.net.total_bbox)++; |
| | | |
| | | const int truth_in_index = t * l.truth_size + b * l.truths + 5; |
| | | const int track_id = state.truth[truth_in_index]; |
| | | const int truth_out_index = b * l.n * l.w * l.h + mask_n * l.w * l.h + j * l.w + i; |
| | | l.labels[truth_out_index] = track_id; |
| | | l.class_ids[truth_out_index] = class_id; |
| | | //printf(" track_id = %d, t = %d, b = %d, truth_in_index = %d, truth_out_index = %d \n", track_id, t, b, truth_in_index, truth_out_index); |
| | |
|
| | | // range is 0 <= 1
|
| | | tot_iou += all_ious.iou;
|
| | | tot_iou_loss += 1 - all_ious.iou;
|
| | | args->tot_iou += all_ious.iou; |
| | | args->tot_iou_loss += 1 - all_ious.iou; |
| | | // range is -1 <= giou <= 1
|
| | | tot_giou += all_ious.giou;
|
| | | tot_giou_loss += 1 - all_ious.giou;
|
| | | args->tot_giou_loss += 1 - all_ious.giou; |
| | |
|
| | | tot_diou += all_ious.diou;
|
| | | tot_diou_loss += 1 - all_ious.diou;
|
| | |
| | |
|
| | | 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] = class_multiplier * l.cls_normalizer * (1 - l.output[obj_index]);
|
| | | if (l.objectness_smooth) { |
| | | float delta_obj = class_multiplier * l.obj_normalizer * (1 - l.output[obj_index]); |
| | | if (l.delta[obj_index] == 0) l.delta[obj_index] = delta_obj; |
| | | } |
| | | else l.delta[obj_index] = class_multiplier * l.obj_normalizer * (1 - l.output[obj_index]); |
| | |
|
| | | 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, l.label_smooth_eps, l.classes_multipliers);
|
| | | 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, l.classes_multipliers, l.cls_normalizer); |
| | |
|
| | | //printf(" label: class_id = %d, truth.x = %f, truth.y = %f, truth.w = %f, truth.h = %f \n", class_id, truth.x, truth.y, truth.w, truth.h);
|
| | | //printf(" mask_n = %d, l.output[obj_index] = %f, l.output[class_index + class_id] = %f \n\n", mask_n, l.output[obj_index], l.output[class_index + class_id]);
|
| | |
|
| | | ++count;
|
| | | ++class_count;
|
| | | ++(args->count); |
| | | ++(args->class_count); |
| | | if (all_ious.iou > .5) recall += 1;
|
| | | if (all_ious.iou > .75) recall75 += 1;
|
| | | }
|
| | |
| | | // iou, n
|
| | |
|
| | | if (iou > l.iou_thresh) {
|
| | | int class_id = state.truth[t*(4 + 1) + b*l.truths + 4];
|
| | | int class_id = state.truth[t * l.truth_size + b * l.truths + 4]; |
| | | if (l.map) class_id = l.map[class_id];
|
| | |
|
| | | int box_index = entry_index(l, b, mask_n*l.w*l.h + j*l.w + i, 0);
|
| | | const float class_multiplier = (l.classes_multipliers) ? l.classes_multipliers[class_id] : 1.0f;
|
| | | ious all_ious = delta_yolo_box(truth, l.output, l.biases, 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 * class_multiplier, l.iou_loss, 1, l.max_delta);
|
| | | ious all_ious = delta_yolo_box(truth, l.output, l.biases, 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 * class_multiplier, l.iou_loss, 1, l.max_delta, state.net.rewritten_bbox, l.new_coords); |
| | | (*state.net.total_bbox)++; |
| | |
|
| | | // range is 0 <= 1
|
| | | tot_iou += all_ious.iou;
|
| | | tot_iou_loss += 1 - all_ious.iou;
|
| | | args->tot_iou += all_ious.iou; |
| | | args->tot_iou_loss += 1 - all_ious.iou; |
| | | // range is -1 <= giou <= 1
|
| | | tot_giou += all_ious.giou;
|
| | | tot_giou_loss += 1 - all_ious.giou;
|
| | | args->tot_giou_loss += 1 - all_ious.giou; |
| | |
|
| | | tot_diou += all_ious.diou;
|
| | | tot_diou_loss += 1 - all_ious.diou;
|
| | |
| | |
|
| | | 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] = class_multiplier * l.cls_normalizer * (1 - l.output[obj_index]);
|
| | | if (l.objectness_smooth) { |
| | | float delta_obj = class_multiplier * l.obj_normalizer * (1 - l.output[obj_index]); |
| | | if (l.delta[obj_index] == 0) l.delta[obj_index] = delta_obj; |
| | | } |
| | | else l.delta[obj_index] = class_multiplier * l.obj_normalizer * (1 - l.output[obj_index]); |
| | |
|
| | | 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, l.label_smooth_eps, l.classes_multipliers);
|
| | | 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, l.classes_multipliers, l.cls_normalizer); |
| | |
|
| | | ++count;
|
| | | ++class_count;
|
| | | ++(args->count); |
| | | ++(args->class_count); |
| | | if (all_ious.iou > .5) recall += 1;
|
| | | if (all_ious.iou > .75) recall75 += 1;
|
| | | }
|
| | |
| | | }
|
| | | }
|
| | |
|
| | | if (l.iou_thresh < 1.0f) { |
| | | // averages the deltas obtained by the function: delta_yolo_box()_accumulate
|
| | | for (j = 0; j < l.h; ++j) {
|
| | | for (i = 0; i < l.w; ++i) {
|
| | | for (n = 0; n < l.n; ++n) {
|
| | | int obj_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4); |
| | | int box_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 0);
|
| | | int class_index = entry_index(l, b, n*l.w*l.h + j*l.w + i, 4 + 1);
|
| | | const int stride = l.w*l.h;
|
| | |
|
| | | if (l.delta[obj_index] != 0) |
| | | averages_yolo_deltas(class_index, box_index, stride, l.classes, l.delta);
|
| | | }
|
| | | } |
| | | } |
| | | } |
| | | |
| | | } |
| | | |
| | | return 0; |
| | | } |
| | | |
| | | |
| | | |
| | | void forward_yolo_layer(const layer l, network_state state) |
| | | { |
| | | //int i, j, b, t, n; |
| | | memcpy(l.output, state.input, l.outputs*l.batch * sizeof(float)); |
| | | int b, n; |
| | | |
| | | #ifndef GPU |
| | | for (b = 0; b < l.batch; ++b) { |
| | | for (n = 0; n < l.n; ++n) { |
| | | int bbox_index = entry_index(l, b, n*l.w*l.h, 0); |
| | | if (l.new_coords) { |
| | | //activate_array(l.output + bbox_index, 4 * l.w*l.h, LOGISTIC); // x,y,w,h |
| | | } |
| | | else { |
| | | activate_array(l.output + bbox_index, 2 * l.w*l.h, LOGISTIC); // x,y, |
| | | int obj_index = entry_index(l, b, n*l.w*l.h, 4); |
| | | activate_array(l.output + obj_index, (1 + l.classes)*l.w*l.h, LOGISTIC); |
| | | } |
| | | scal_add_cpu(2 * l.w*l.h, l.scale_x_y, -0.5*(l.scale_x_y - 1), l.output + bbox_index, 1); // scale x,y |
| | | } |
| | | } |
| | | #endif |
| | | |
| | | // delta is zeroed |
| | | memset(l.delta, 0, l.outputs * l.batch * sizeof(float)); |
| | | if (!state.train) return; |
| | | |
| | | int i; |
| | | for (i = 0; i < l.batch * l.w*l.h*l.n; ++i) l.labels[i] = -1; |
| | | for (i = 0; i < l.batch * l.w*l.h*l.n; ++i) l.class_ids[i] = -1; |
| | | //float avg_iou = 0; |
| | | float tot_iou = 0; |
| | | float tot_giou = 0; |
| | | float tot_diou = 0; |
| | | float tot_ciou = 0; |
| | | float tot_iou_loss = 0; |
| | | float tot_giou_loss = 0; |
| | | float tot_diou_loss = 0; |
| | | float tot_ciou_loss = 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; |
| | | |
| | | |
| | | int num_threads = l.batch; |
| | | pthread_t* threads = (pthread_t*)calloc(num_threads, sizeof(pthread_t)); |
| | | |
| | | struct train_yolo_args* yolo_args = (train_yolo_args*)xcalloc(l.batch, sizeof(struct train_yolo_args)); |
| | | |
| | | for (b = 0; b < l.batch; b++) |
| | | { |
| | | yolo_args[b].l = l; |
| | | yolo_args[b].state = state; |
| | | yolo_args[b].b = b; |
| | | |
| | | yolo_args[b].tot_iou = 0; |
| | | yolo_args[b].tot_iou_loss = 0; |
| | | yolo_args[b].tot_giou_loss = 0; |
| | | yolo_args[b].count = 0; |
| | | yolo_args[b].class_count = 0; |
| | | |
| | | if (pthread_create(&threads[b], 0, process_batch, &(yolo_args[b]))) error("Thread creation failed"); |
| | | } |
| | | |
| | | for (b = 0; b < l.batch; b++) |
| | | { |
| | | pthread_join(threads[b], 0); |
| | | |
| | | tot_iou += yolo_args[b].tot_iou; |
| | | tot_iou_loss += yolo_args[b].tot_iou_loss; |
| | | tot_giou_loss += yolo_args[b].tot_giou_loss; |
| | | count += yolo_args[b].count; |
| | | class_count += yolo_args[b].class_count; |
| | | } |
| | | |
| | | free(yolo_args); |
| | | free(threads); |
| | | |
| | | // Search for an equidistant point from the distant boundaries of the local minimum |
| | | int iteration_num = get_current_iteration(state.net); |
| | | const int start_point = state.net.max_batches * 3 / 4; |
| | | //printf(" equidistant_point ep = %d, it = %d \n", state.net.equidistant_point, iteration_num); |
| | | |
| | | if ((state.net.badlabels_rejection_percentage && start_point < iteration_num) || |
| | | (state.net.num_sigmas_reject_badlabels && start_point < iteration_num) || |
| | | (state.net.equidistant_point && state.net.equidistant_point < iteration_num)) |
| | | { |
| | | const float progress_it = iteration_num - state.net.equidistant_point; |
| | | const float progress = progress_it / (state.net.max_batches - state.net.equidistant_point); |
| | | float ep_loss_threshold = (*state.net.delta_rolling_avg) * progress * 1.4; |
| | | |
| | | float cur_max = 0; |
| | | float cur_avg = 0; |
| | | float counter = 0; |
| | | for (i = 0; i < l.batch * l.outputs; ++i) { |
| | | |
| | | if (l.delta[i] != 0) { |
| | | counter++; |
| | | cur_avg += fabs(l.delta[i]); |
| | | |
| | | if (cur_max < fabs(l.delta[i])) |
| | | cur_max = fabs(l.delta[i]); |
| | | } |
| | | } |
| | | |
| | | cur_avg = cur_avg / counter; |
| | | |
| | | if (*state.net.delta_rolling_max == 0) *state.net.delta_rolling_max = cur_max; |
| | | *state.net.delta_rolling_max = *state.net.delta_rolling_max * 0.99 + cur_max * 0.01; |
| | | *state.net.delta_rolling_avg = *state.net.delta_rolling_avg * 0.99 + cur_avg * 0.01; |
| | | |
| | | // reject high loss to filter bad labels |
| | | if (state.net.num_sigmas_reject_badlabels && start_point < iteration_num) |
| | | { |
| | | const float rolling_std = (*state.net.delta_rolling_std); |
| | | const float rolling_max = (*state.net.delta_rolling_max); |
| | | const float rolling_avg = (*state.net.delta_rolling_avg); |
| | | const float progress_badlabels = (float)(iteration_num - start_point) / (start_point); |
| | | |
| | | float cur_std = 0; |
| | | float counter = 0; |
| | | for (i = 0; i < l.batch * l.outputs; ++i) { |
| | | if (l.delta[i] != 0) { |
| | | counter++; |
| | | cur_std += pow(l.delta[i] - rolling_avg, 2); |
| | | } |
| | | } |
| | | cur_std = sqrt(cur_std / counter); |
| | | |
| | | *state.net.delta_rolling_std = *state.net.delta_rolling_std * 0.99 + cur_std * 0.01; |
| | | |
| | | float final_badlebels_threshold = rolling_avg + rolling_std * state.net.num_sigmas_reject_badlabels; |
| | | float badlabels_threshold = rolling_max - progress_badlabels * fabs(rolling_max - final_badlebels_threshold); |
| | | badlabels_threshold = max_val_cmp(final_badlebels_threshold, badlabels_threshold); |
| | | for (i = 0; i < l.batch * l.outputs; ++i) { |
| | | if (fabs(l.delta[i]) > badlabels_threshold) |
| | | l.delta[i] = 0; |
| | | } |
| | | printf(" rolling_std = %f, rolling_max = %f, rolling_avg = %f \n", rolling_std, rolling_max, rolling_avg); |
| | | printf(" badlabels loss_threshold = %f, start_it = %d, progress = %f \n", badlabels_threshold, start_point, progress_badlabels *100); |
| | | |
| | | ep_loss_threshold = min_val_cmp(final_badlebels_threshold, rolling_avg) * progress; |
| | | } |
| | | |
| | | |
| | | // reject some percent of the highest deltas to filter bad labels |
| | | if (state.net.badlabels_rejection_percentage && start_point < iteration_num) { |
| | | if (*state.net.badlabels_reject_threshold == 0) |
| | | *state.net.badlabels_reject_threshold = *state.net.delta_rolling_max; |
| | | |
| | | printf(" badlabels_reject_threshold = %f \n", *state.net.badlabels_reject_threshold); |
| | | |
| | | const float num_deltas_per_anchor = (l.classes + 4 + 1); |
| | | float counter_reject = 0; |
| | | float counter_all = 0; |
| | | for (i = 0; i < l.batch * l.outputs; ++i) { |
| | | if (l.delta[i] != 0) { |
| | | counter_all++; |
| | | if (fabs(l.delta[i]) > (*state.net.badlabels_reject_threshold)) { |
| | | counter_reject++; |
| | | l.delta[i] = 0; |
| | | } |
| | | } |
| | | } |
| | | float cur_percent = 100 * (counter_reject*num_deltas_per_anchor / counter_all); |
| | | if (cur_percent > state.net.badlabels_rejection_percentage) { |
| | | *state.net.badlabels_reject_threshold += 0.01; |
| | | printf(" increase!!! \n"); |
| | | } |
| | | else if (*state.net.badlabels_reject_threshold > 0.01) { |
| | | *state.net.badlabels_reject_threshold -= 0.01; |
| | | printf(" decrease!!! \n"); |
| | | } |
| | | |
| | | printf(" badlabels_reject_threshold = %f, cur_percent = %f, badlabels_rejection_percentage = %f, delta_rolling_max = %f \n", |
| | | *state.net.badlabels_reject_threshold, cur_percent, state.net.badlabels_rejection_percentage, *state.net.delta_rolling_max); |
| | | } |
| | | |
| | | |
| | | // reject low loss to find equidistant point |
| | | if (state.net.equidistant_point && state.net.equidistant_point < iteration_num) { |
| | | printf(" equidistant_point loss_threshold = %f, start_it = %d, progress = %3.1f %% \n", ep_loss_threshold, state.net.equidistant_point, progress * 100); |
| | | for (i = 0; i < l.batch * l.outputs; ++i) { |
| | | if (fabs(l.delta[i]) < ep_loss_threshold) |
| | | l.delta[i] = 0; |
| | | }
|
| | | }
|
| | | }
|
| | |
| | | if (count == 0) count = 1;
|
| | | if (class_count == 0) class_count = 1;
|
| | |
|
| | | //*(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", state.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);
|
| | | if (l.show_details == 0) { |
| | | float loss = pow(mag_array(l.delta, l.outputs * l.batch), 2); |
| | | *(l.cost) = loss; |
| | | |
| | | loss /= l.batch; |
| | | |
| | | fprintf(stderr, "v3 (%s loss, Normalizer: (iou: %.2f, obj: %.2f, cls: %.2f) Region %d Avg (IOU: %f), count: %d, total_loss = %f \n", |
| | | (l.iou_loss == MSE ? "mse" : (l.iou_loss == GIOU ? "giou" : "iou")), l.iou_normalizer, l.obj_normalizer, l.cls_normalizer, state.index, tot_iou / count, count, loss); |
| | | } |
| | | else { |
| | | // show detailed output |
| | |
|
| | | int stride = l.w*l.h;
|
| | | float* no_iou_loss_delta = (float *)calloc(l.batch * l.outputs, sizeof(float));
|
| | | memcpy(no_iou_loss_delta, l.delta, l.batch * l.outputs * sizeof(float));
|
| | | |
| | | |
| | | int j, n; |
| | | for (b = 0; b < l.batch; ++b) {
|
| | | for (j = 0; j < l.h; ++j) {
|
| | | for (i = 0; i < l.w; ++i) {
|
| | |
| | | }
|
| | | }
|
| | | }
|
| | | float classification_loss = l.cls_normalizer * pow(mag_array(no_iou_loss_delta, l.outputs * l.batch), 2);
|
| | | |
| | | float classification_loss = l.obj_normalizer * pow(mag_array(no_iou_loss_delta, l.outputs * l.batch), 2); |
| | | free(no_iou_loss_delta);
|
| | | float loss = pow(mag_array(l.delta, l.outputs * l.batch), 2);
|
| | | float iou_loss = loss - classification_loss;
|
| | |
|
| | | float avg_iou_loss = 0;
|
| | | *(l.cost) = loss; |
| | | |
| | | // gIOU loss + MSE (objectness) loss
|
| | | if (l.iou_loss == MSE) {
|
| | | *(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2);
|
| | |
| | | *(l.cost) = avg_iou_loss + classification_loss;
|
| | | }
|
| | |
|
| | | |
| | | loss /= l.batch;
|
| | | classification_loss /= l.batch;
|
| | | iou_loss /= l.batch;
|
| | |
|
| | | fprintf(stderr, "v3 (%s loss, Normalizer: (iou: %.2f, cls: %.2f) Region %d Avg (IOU: %f, GIOU: %f), Class: %f, Obj: %f, No Obj: %f, .5R: %f, .75R: %f, count: %d, class_loss = %f, iou_loss = %f, total_loss = %f \n",
|
| | | (l.iou_loss == MSE ? "mse" : (l.iou_loss == GIOU ? "giou" : "iou")), l.iou_normalizer, l.cls_normalizer, state.index, tot_iou / count, tot_giou / count, avg_cat / class_count, avg_obj / count, avg_anyobj / (l.w*l.h*l.n*l.batch), recall / count, recall75 / count, count,
|
| | | classification_loss, iou_loss, loss);
|
| | | fprintf(stderr, "v3 (%s loss, Normalizer: (iou: %.2f, obj: %.2f, cls: %.2f) Region %d Avg (IOU: %f), count: %d, class_loss = %f, iou_loss = %f, total_loss = %f \n", |
| | | (l.iou_loss == MSE ? "mse" : (l.iou_loss == GIOU ? "giou" : "iou")), l.iou_normalizer, l.obj_normalizer, l.cls_normalizer, state.index, tot_iou / count, count, classification_loss, iou_loss, loss); |
| | | |
| | | //fprintf(stderr, "v3 (%s loss, Normalizer: (iou: %.2f, cls: %.2f) Region %d Avg (IOU: %f, GIOU: %f), Class: %f, Obj: %f, No Obj: %f, .5R: %f, .75R: %f, count: %d, class_loss = %f, iou_loss = %f, total_loss = %f \n", |
| | | // (l.iou_loss == MSE ? "mse" : (l.iou_loss == GIOU ? "giou" : "iou")), l.iou_normalizer, l.obj_normalizer, state.index, tot_iou / count, tot_giou / count, avg_cat / class_count, avg_obj / count, avg_anyobj / (l.w*l.h*l.n*l.batch), recall / count, recall75 / count, count, |
| | | // classification_loss, iou_loss, loss); |
| | | } |
| | | }
|
| | |
|
| | | void backward_yolo_layer(const layer l, network_state state)
|
| | |
| | | if (objectness > thresh) {
|
| | | //printf("\n objectness = %f, thresh = %f, i = %d, n = %d \n", objectness, thresh, i, n);
|
| | | int box_index = entry_index(l, 0, n*l.w*l.h + i, 0);
|
| | | dets[count].bbox = get_yolo_box(predictions, l.biases, l.mask[n], box_index, col, row, l.w, l.h, netw, neth, l.w*l.h);
|
| | | dets[count].bbox = get_yolo_box(predictions, l.biases, l.mask[n], box_index, col, row, l.w, l.h, netw, neth, l.w*l.h, l.new_coords); |
| | | dets[count].objectness = objectness;
|
| | | dets[count].classes = l.classes;
|
| | | if (l.embedding_output) { |
| | | get_embedding(l.embedding_output, l.w, l.h, l.n*l.embedding_size, l.embedding_size, col, row, n, 0, dets[count].embeddings); |
| | | } |
| | | |
| | | for (j = 0; j < l.classes; ++j) {
|
| | | int class_index = entry_index(l, 0, n*l.w*l.h + i, 4 + 1 + j);
|
| | | float prob = objectness*predictions[class_index];
|
| | |
| | | if (objectness > thresh) {
|
| | | //printf("\n objectness = %f, thresh = %f, i = %d, n = %d \n", objectness, thresh, i, n);
|
| | | int box_index = entry_index(l, batch, n*l.w*l.h + i, 0);
|
| | | dets[count].bbox = get_yolo_box(predictions, l.biases, l.mask[n], box_index, col, row, l.w, l.h, netw, neth, l.w*l.h);
|
| | | dets[count].bbox = get_yolo_box(predictions, l.biases, l.mask[n], box_index, col, row, l.w, l.h, netw, neth, l.w*l.h, l.new_coords); |
| | | dets[count].objectness = objectness;
|
| | | dets[count].classes = l.classes;
|
| | | if (l.embedding_output) { |
| | | get_embedding(l.embedding_output, l.w, l.h, l.n*l.embedding_size, l.embedding_size, col, row, n, batch, dets[count].embeddings); |
| | | } |
| | | |
| | | for (j = 0; j < l.classes; ++j) {
|
| | | int class_index = entry_index(l, batch, n*l.w*l.h + i, 4 + 1 + j);
|
| | | float prob = objectness*predictions[class_index];
|
| | |
| | |
|
| | | void forward_yolo_layer_gpu(const layer l, network_state state)
|
| | | {
|
| | | if (l.embedding_output) { |
| | | layer le = state.net.layers[l.embedding_layer_id]; |
| | | cuda_pull_array_async(le.output_gpu, l.embedding_output, le.batch*le.outputs); |
| | | } |
| | | |
| | | //copy_ongpu(l.batch*l.inputs, state.input, 1, l.output_gpu, 1);
|
| | | simple_copy_ongpu(l.batch*l.inputs, state.input, l.output_gpu);
|
| | | int b, n;
|
| | | 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);
|
| | | int bbox_index = entry_index(l, b, n*l.w*l.h, 0); |
| | | // y = 1./(1. + exp(-x))
|
| | | // x = ln(y/(1-y)) // ln - natural logarithm (base = e)
|
| | | // if(y->1) x -> inf
|
| | | // if(y->0) x -> -inf
|
| | | activate_array_ongpu(l.output_gpu + index, 2*l.w*l.h, LOGISTIC); // x,y
|
| | | if (l.scale_x_y != 1) scal_add_ongpu(2 * l.w*l.h, l.scale_x_y, -0.5*(l.scale_x_y - 1), l.output_gpu + index, 1); // scale x,y
|
| | | index = entry_index(l, b, n*l.w*l.h, 4);
|
| | | activate_array_ongpu(l.output_gpu + index, (1+l.classes)*l.w*l.h, LOGISTIC); // classes and objectness
|
| | | if (l.new_coords) { |
| | | //activate_array_ongpu(l.output_gpu + bbox_index, 4 * l.w*l.h, LOGISTIC); // x,y,w,h |
| | | } |
| | | else { |
| | | activate_array_ongpu(l.output_gpu + bbox_index, 2 * l.w*l.h, LOGISTIC); // x,y |
| | | |
| | | int obj_index = entry_index(l, b, n*l.w*l.h, 4); |
| | | activate_array_ongpu(l.output_gpu + obj_index, (1 + l.classes)*l.w*l.h, LOGISTIC); // classes and objectness |
| | | } |
| | | if (l.scale_x_y != 1) scal_add_ongpu(2 * l.w*l.h, l.scale_x_y, -0.5*(l.scale_x_y - 1), l.output_gpu + bbox_index, 1); // scale x,y |
| | | }
|
| | | }
|
| | | if(!state.train || l.onlyforward){
|
| | |
| | |
|
| | | void backward_yolo_layer_gpu(const layer l, network_state state)
|
| | | {
|
| | | axpy_ongpu(l.batch*l.inputs, state.net.loss_scale, l.delta_gpu, 1, state.delta, 1);
|
| | | axpy_ongpu(l.batch*l.inputs, state.net.loss_scale * l.delta_normalizer, l.delta_gpu, 1, state.delta, 1); |
| | | }
|
| | | #endif
|