| | |
| | | |
| | | int get_detection_layer_locations(detection_layer l) |
| | | { |
| | | return l.inputs / (l.classes+l.coords+l.rescore+l.background); |
| | | return l.inputs / (l.classes+l.coords+l.joint+(l.background || l.objectness)); |
| | | } |
| | | |
| | | int get_detection_layer_output_size(detection_layer l) |
| | | { |
| | | return get_detection_layer_locations(l)*(l.background + l.classes + l.coords); |
| | | return get_detection_layer_locations(l)*((l.background || l.objectness) + l.classes + l.coords); |
| | | } |
| | | |
| | | detection_layer make_detection_layer(int batch, int inputs, int classes, int coords, int rescore, int background, int nuisance) |
| | | detection_layer make_detection_layer(int batch, int inputs, int classes, int coords, int joint, int rescore, int background, int objectness) |
| | | { |
| | | detection_layer l = {0}; |
| | | l.type = DETECTION; |
| | |
| | | l.classes = classes; |
| | | l.coords = coords; |
| | | l.rescore = rescore; |
| | | l.nuisance = nuisance; |
| | | l.objectness = objectness; |
| | | l.background = background; |
| | | l.joint = joint; |
| | | l.cost = calloc(1, sizeof(float)); |
| | | l.does_cost=1; |
| | | l.background = background; |
| | | int outputs = get_detection_layer_output_size(l); |
| | | l.outputs = outputs; |
| | | l.output = calloc(batch*outputs, sizeof(float)); |
| | |
| | | return l; |
| | | } |
| | | |
| | | void dark_zone(detection_layer l, int class, int start, network_state state) |
| | | { |
| | | int index = start+l.background+class; |
| | | int size = l.classes+l.coords+l.background; |
| | | int location = (index%(7*7*size)) / size ; |
| | | int r = location / 7; |
| | | int c = location % 7; |
| | | int dr, dc; |
| | | for(dr = -1; dr <= 1; ++dr){ |
| | | for(dc = -1; dc <= 1; ++dc){ |
| | | if(!(dr || dc)) continue; |
| | | if((r + dr) > 6 || (r + dr) < 0) continue; |
| | | if((c + dc) > 6 || (c + dc) < 0) continue; |
| | | int di = (dr*7 + dc) * size; |
| | | if(state.truth[index+di]) continue; |
| | | l.output[index + di] = 0; |
| | | //if(!state.truth[start+di]) continue; |
| | | //l.output[start + di] = 1; |
| | | } |
| | | } |
| | | } |
| | | |
| | | typedef struct{ |
| | | float dx, dy, dw, dh; |
| | | } dbox; |
| | |
| | | wiou = ((1-wiou)*(1-wiou) - iou)/(.00001); |
| | | hiou = ((1-hiou)*(1-hiou) - iou)/(.00001); |
| | | printf("manual %f %f %f %f\n", xiou, yiou, wiou, hiou); |
| | | /* |
| | | |
| | | while(count++ < 300){ |
| | | dbox d = diou(a, b); |
| | | printf("%f %f %f %f\n", a.x, a.y, a.w, a.h); |
| | | a.x += .1*d.dx; |
| | | a.w += .1*d.dw; |
| | | a.y += .1*d.dy; |
| | | a.h += .1*d.dh; |
| | | printf("inter: %f\n", box_intersection(a, b)); |
| | | printf("union: %f\n", box_union(a, b)); |
| | | printf("IOU: %f\n", box_iou(a, b)); |
| | | if(d.dx==0 && d.dw==0 && d.dy==0 && d.dh==0) { |
| | | printf("break!!!\n"); |
| | | break; |
| | | } |
| | | } |
| | | */ |
| | | } |
| | | |
| | | dbox diou(box a, box b) |
| | |
| | | int locations = get_detection_layer_locations(l); |
| | | int i,j; |
| | | for(i = 0; i < l.batch*locations; ++i){ |
| | | int mask = (!state.truth || state.truth[out_i + l.background + l.classes + 2]); |
| | | int mask = (!state.truth || state.truth[out_i + (l.background || l.objectness) + l.classes + 2]); |
| | | float scale = 1; |
| | | if(l.rescore) scale = state.input[in_i++]; |
| | | else if(l.nuisance){ |
| | | if(l.joint) scale = state.input[in_i++]; |
| | | else if(l.objectness){ |
| | | l.output[out_i++] = 1-state.input[in_i++]; |
| | | scale = mask; |
| | | } |
| | |
| | | for(j = 0; j < l.classes; ++j){ |
| | | l.output[out_i++] = scale*state.input[in_i++]; |
| | | } |
| | | if(l.nuisance){ |
| | | if(l.objectness){ |
| | | |
| | | }else if(l.background){ |
| | | softmax_array(l.output + out_i - l.classes-l.background, l.classes+l.background, l.output + out_i - l.classes-l.background); |
| | |
| | | l.output[out_i++] = mask*state.input[in_i++]; |
| | | } |
| | | } |
| | | if(l.does_cost && state.train && 0){ |
| | | int count = 0; |
| | | float avg = 0; |
| | | float avg_iou = 0; |
| | | int count = 0; |
| | | if(l.does_cost && state.train){ |
| | | *(l.cost) = 0; |
| | | int size = get_detection_layer_output_size(l) * l.batch; |
| | | memset(l.delta, 0, size * sizeof(float)); |
| | | for (i = 0; i < l.batch*locations; ++i) { |
| | | int classes = l.nuisance+l.classes; |
| | | int classes = l.objectness+l.classes; |
| | | int offset = i*(classes+l.coords); |
| | | for (j = offset; j < offset+classes; ++j) { |
| | | *(l.cost) += pow(state.truth[j] - l.output[j], 2); |
| | | l.delta[j] = state.truth[j] - l.output[j]; |
| | | } |
| | | box truth; |
| | | truth.x = state.truth[j+0]; |
| | | truth.y = state.truth[j+1]; |
| | | truth.w = state.truth[j+2]; |
| | | truth.h = state.truth[j+3]; |
| | | box out; |
| | | out.x = l.output[j+0]; |
| | | out.y = l.output[j+1]; |
| | | out.w = l.output[j+2]; |
| | | out.h = l.output[j+3]; |
| | | if(!(truth.w*truth.h)) continue; |
| | | //printf("iou: %f\n", iou); |
| | | dbox d = diou(out, truth); |
| | | l.delta[j+0] = d.dx; |
| | | l.delta[j+1] = d.dy; |
| | | l.delta[j+2] = d.dw; |
| | | l.delta[j+3] = d.dh; |
| | | |
| | | int sqr = 1; |
| | | if(sqr){ |
| | | truth.w *= truth.w; |
| | | truth.h *= truth.h; |
| | | out.w *= out.w; |
| | | out.h *= out.h; |
| | | } |
| | | float iou = box_iou(truth, out); |
| | | *(l.cost) += pow((1-iou), 2); |
| | | avg += iou; |
| | | box truth; |
| | | truth.x = state.truth[j+0]/7; |
| | | truth.y = state.truth[j+1]/7; |
| | | truth.w = pow(state.truth[j+2], 2); |
| | | truth.h = pow(state.truth[j+3], 2); |
| | | box out; |
| | | out.x = l.output[j+0]/7; |
| | | out.y = l.output[j+1]/7; |
| | | out.w = pow(l.output[j+2], 2); |
| | | out.h = pow(l.output[j+3], 2); |
| | | |
| | | if(!(truth.w*truth.h)) continue; |
| | | float iou = box_iou(out, truth); |
| | | avg_iou += iou; |
| | | ++count; |
| | | dbox delta = diou(out, truth); |
| | | |
| | | l.delta[j+0] = 10 * delta.dx/7; |
| | | l.delta[j+1] = 10 * delta.dy/7; |
| | | l.delta[j+2] = 10 * delta.dw * 2 * sqrt(out.w); |
| | | l.delta[j+3] = 10 * delta.dh * 2 * sqrt(out.h); |
| | | |
| | | |
| | | *(l.cost) += pow((1-iou), 2); |
| | | l.delta[j+0] = 4 * (state.truth[j+0] - l.output[j+0]); |
| | | l.delta[j+1] = 4 * (state.truth[j+1] - l.output[j+1]); |
| | | l.delta[j+2] = 4 * (state.truth[j+2] - l.output[j+2]); |
| | | l.delta[j+3] = 4 * (state.truth[j+3] - l.output[j+3]); |
| | | if(l.rescore){ |
| | | for (j = offset; j < offset+classes; ++j) { |
| | | if(state.truth[j]) state.truth[j] = iou; |
| | | l.delta[j] = state.truth[j] - l.output[j]; |
| | | } |
| | | } |
| | | } |
| | | fprintf(stderr, "Avg IOU: %f\n", avg/count); |
| | | printf("Avg IOU: %f\n", avg_iou/count); |
| | | } |
| | | /* |
| | | int count = 0; |
| | | for(i = 0; i < l.batch*locations; ++i){ |
| | | for(j = 0; j < l.classes+l.background; ++j){ |
| | | printf("%f, ", l.output[count++]); |
| | | } |
| | | printf("\n"); |
| | | for(j = 0; j < l.coords; ++j){ |
| | | printf("%f, ", l.output[count++]); |
| | | } |
| | | printf("\n"); |
| | | } |
| | | */ |
| | | /* |
| | | if(l.background || 1){ |
| | | for(i = 0; i < l.batch*locations; ++i){ |
| | | int index = i*(l.classes+l.coords+l.background); |
| | | for(j= 0; j < l.classes; ++j){ |
| | | if(state.truth[index+j+l.background]){ |
| | | //dark_zone(l, j, index, state); |
| | | } |
| | | } |
| | | } |
| | | } |
| | | */ |
| | | } |
| | | |
| | | void backward_detection_layer(const detection_layer l, network_state state) |
| | |
| | | for(i = 0; i < l.batch*locations; ++i){ |
| | | float scale = 1; |
| | | float latent_delta = 0; |
| | | if(l.rescore) scale = state.input[in_i++]; |
| | | else if (l.nuisance) state.delta[in_i++] = -l.delta[out_i++]; |
| | | if(l.joint) scale = state.input[in_i++]; |
| | | else if (l.objectness) state.delta[in_i++] = -l.delta[out_i++]; |
| | | else if (l.background) state.delta[in_i++] = scale*l.delta[out_i++]; |
| | | for(j = 0; j < l.classes; ++j){ |
| | | latent_delta += state.input[in_i]*l.delta[out_i]; |
| | | state.delta[in_i++] = scale*l.delta[out_i++]; |
| | | } |
| | | |
| | | if (l.nuisance) { |
| | | if (l.objectness) { |
| | | |
| | | }else if (l.background) gradient_array(l.output + out_i, l.coords, LOGISTIC, l.delta + out_i); |
| | | for(j = 0; j < l.coords; ++j){ |
| | | state.delta[in_i++] = l.delta[out_i++]; |
| | | } |
| | | if(l.rescore) state.delta[in_i-l.coords-l.classes-l.rescore-l.background] = latent_delta; |
| | | if(l.joint) state.delta[in_i-l.coords-l.classes-l.joint] = latent_delta; |
| | | } |
| | | } |
| | | |