diff --git a/src/layer.h b/src/layer.h index 9aff67fa..c149f29d 100644 --- a/src/layer.h +++ b/src/layer.h @@ -67,6 +67,7 @@ struct layer{ int size; int side; int stride; + int reverse; int pad; int sqrt; int flip; @@ -118,6 +119,7 @@ struct layer{ int bias_match; int random; float thresh; + int classfix; int dontload; int dontloadscales; diff --git a/src/parser.c b/src/parser.c index 26f45d36..4e71fe5e 100644 --- a/src/parser.c +++ b/src/parser.c @@ -268,6 +268,7 @@ layer parse_region(list *options, size_params params) l.rescore = option_find_int_quiet(options, "rescore",0); l.thresh = option_find_float(options, "thresh", .5); + l.classfix = option_find_int_quiet(options, "classfix", 0); l.coord_scale = option_find_float(options, "coord_scale", 1); l.object_scale = option_find_float(options, "object_scale", 1); @@ -357,6 +358,7 @@ crop_layer parse_crop(list *options, size_params params) layer parse_reorg(list *options, size_params params) { int stride = option_find_int(options, "stride",1); + int reverse = option_find_int_quiet(options, "reverse",0); int batch,h,w,c; h = params.h; @@ -365,7 +367,7 @@ layer parse_reorg(list *options, size_params params) batch=params.batch; if(!(h && w && c)) error("Layer before reorg layer must output image."); - layer layer = make_reorg_layer(batch,w,h,c,stride); + layer layer = make_reorg_layer(batch,w,h,c,stride,reverse); return layer; } diff --git a/src/region_layer.c b/src/region_layer.c index 2702636f..269be1f3 100644 --- a/src/region_layer.c +++ b/src/region_layer.c @@ -89,6 +89,31 @@ float delta_region_box(box truth, float *x, float *biases, int n, int index, int return iou; } +void delta_region_class(float *output, float *delta, int index, int class, int classes, tree *hier, float scale, float *avg_cat) +{ + int i, n; + if(hier){ + float pred = 1; + while(class >= 0){ + pred *= output[index + class]; + int g = hier->group[class]; + int offset = hier->group_offset[g]; + for(i = 0; i < hier->group_size[g]; ++i){ + delta[index + offset + i] = scale * (0 - output[index + offset + i]); + } + delta[index + class] = scale * (1 - output[index + class]); + + class = hier->parent[class]; + } + *avg_cat += pred; + } else { + for(n = 0; n < classes; ++n){ + delta[index + n] = scale * (((n == class)?1 : 0) - output[index + n]); + if(n == class) *avg_cat += output[index + n]; + } + } +} + float logit(float x) { return log(x/(1.-x)); @@ -125,6 +150,7 @@ void forward_region_layer(const region_layer l, network_state state) 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) { for (j = 0; j < l.h; ++j) { @@ -133,15 +159,28 @@ void forward_region_layer(const region_layer l, network_state state) int index = size*(j*l.w*l.n + i*l.n + n) + b*l.outputs; box pred = get_region_box(l.output, l.biases, n, index, i, j, l.w, l.h); float best_iou = 0; + int best_class = -1; for(t = 0; t < 30; ++t){ box truth = float_to_box(state.truth + t*5 + b*l.truths); if(!truth.x) break; float iou = box_iou(pred, truth); - if (iou > best_iou) best_iou = iou; + if (iou > best_iou) { + best_class = state.truth[t*5 + b*l.truths + 4]; + best_iou = iou; + } } avg_anyobj += l.output[index + 4]; l.delta[index + 4] = l.noobject_scale * ((0 - l.output[index + 4]) * logistic_gradient(l.output[index + 4])); - if(best_iou > l.thresh) l.delta[index + 4] = 0; + if(l.classfix == -1) l.delta[index + 4] = l.noobject_scale * ((best_iou - l.output[index + 4]) * logistic_gradient(l.output[index + 4])); + else{ + if (best_iou > l.thresh) { + l.delta[index + 4] = 0; + if(l.classfix > 0){ + delta_region_class(l.output, l.delta, index + 5, best_class, l.classes, l.softmax_tree, l.class_scale*(l.classfix == 2 ? l.output[index + 4] : 1), &avg_cat); + ++class_count; + } + } + } if(*(state.net.seen) < 12800){ box truth = {0}; @@ -205,35 +244,15 @@ void forward_region_layer(const region_layer l, network_state state) int class = state.truth[t*5 + b*l.truths + 4]; if (l.map) class = l.map[class]; - if(l.softmax_tree){ - float pred = 1; - while(class >= 0){ - pred *= l.output[best_index + 5 + class]; - int g = l.softmax_tree->group[class]; - int i; - int offset = l.softmax_tree->group_offset[g]; - for(i = 0; i < l.softmax_tree->group_size[g]; ++i){ - int index = best_index + 5 + offset + i; - l.delta[index] = l.class_scale * (0 - l.output[index]); - } - l.delta[best_index + 5 + class] = l.class_scale * (1 - l.output[best_index + 5 + class]); - - class = l.softmax_tree->parent[class]; - } - avg_cat += pred; - } else { - for(n = 0; n < l.classes; ++n){ - l.delta[best_index + 5 + n] = l.class_scale * (((n == class)?1 : 0) - l.output[best_index + 5 + n]); - if(n == class) avg_cat += l.output[best_index + 5 + n]; - } - } + delta_region_class(l.output, l.delta, best_index + 5, class, l.classes, l.softmax_tree, l.class_scale, &avg_cat); ++count; + ++class_count; } } //printf("\n"); reorg(l.delta, l.w*l.h, size*l.n, l.batch, 0); *(l.cost) = pow(mag_array(l.delta, l.outputs * l.batch), 2); - printf("Region Avg IOU: %f, Class: %f, Obj: %f, No Obj: %f, Avg Recall: %f, count: %d\n", avg_iou/count, avg_cat/count, avg_obj/count, avg_anyobj/(l.w*l.h*l.n*l.batch), recall/count, count); + printf("Region Avg IOU: %f, Class: %f, Obj: %f, No Obj: %f, Avg Recall: %f, count: %d\n", avg_iou/count, avg_cat/class_count, avg_obj/count, avg_anyobj/(l.w*l.h*l.n*l.batch), recall/count, count); } void backward_region_layer(const region_layer l, network_state state) @@ -245,7 +264,6 @@ void get_region_boxes(layer l, int w, int h, float thresh, float **probs, box *b { int i,j,n; float *predictions = l.output; - //int per_cell = 5*num+classes; for (i = 0; i < l.w*l.h; ++i){ int row = i / l.w; int col = i % l.w; @@ -253,6 +271,7 @@ void get_region_boxes(layer l, int w, int h, float thresh, float **probs, box *b int index = i*l.n + n; int p_index = index * (l.classes + 5) + 4; float scale = predictions[p_index]; + if(l.classfix == -1 && scale < .5) scale = 0; int box_index = index * (l.classes + 5); boxes[index] = get_region_box(predictions, l.biases, n, box_index, col, row, l.w, l.h); boxes[index].x *= w; @@ -262,7 +281,7 @@ void get_region_boxes(layer l, int w, int h, float thresh, float **probs, box *b int class_index = index * (l.classes + 5) + 5; if(l.softmax_tree){ - + hierarchy_predictions(predictions + class_index, l.classes, l.softmax_tree, 0); int found = 0; for(j = l.classes - 1; j >= 0; --j){ diff --git a/src/reorg_layer.c b/src/reorg_layer.c index 5bc257a3..0f2a1c21 100644 --- a/src/reorg_layer.c +++ b/src/reorg_layer.c @@ -4,7 +4,7 @@ #include -layer make_reorg_layer(int batch, int h, int w, int c, int stride) +layer make_reorg_layer(int batch, int h, int w, int c, int stride, int reverse) { layer l = {0}; l.type = REORG; @@ -13,9 +13,15 @@ layer make_reorg_layer(int batch, int h, int w, int c, int stride) l.h = h; l.w = w; l.c = c; - l.out_w = w*stride; - l.out_h = h*stride; - l.out_c = c/(stride*stride); + if(reverse){ + l.out_w = w*stride; + l.out_h = h*stride; + l.out_c = c/(stride*stride); + }else{ + l.out_w = w/stride; + l.out_h = h/stride; + l.out_c = c*(stride*stride); + } fprintf(stderr, "Reorg Layer: %d x %d x %d image -> %d x %d x %d image, \n", w,h,c,l.out_w, l.out_h, l.out_c); l.outputs = l.out_h * l.out_w * l.out_c; l.inputs = h*w*c; @@ -25,13 +31,13 @@ layer make_reorg_layer(int batch, int h, int w, int c, int stride) l.forward = forward_reorg_layer; l.backward = backward_reorg_layer; - #ifdef GPU +#ifdef GPU l.forward_gpu = forward_reorg_layer_gpu; l.backward_gpu = backward_reorg_layer_gpu; l.output_gpu = cuda_make_array(l.output, output_size); l.delta_gpu = cuda_make_array(l.delta, output_size); - #endif +#endif return l; } @@ -52,12 +58,12 @@ void resize_reorg_layer(layer *l, int w, int h) l->output = realloc(l->output, output_size * sizeof(float)); l->delta = realloc(l->delta, output_size * sizeof(float)); - #ifdef GPU +#ifdef GPU cuda_free(l->output_gpu); cuda_free(l->delta_gpu); l->output_gpu = cuda_make_array(l->output, output_size); l->delta_gpu = cuda_make_array(l->delta, output_size); - #endif +#endif } void forward_reorg_layer(const layer l, network_state state) @@ -107,11 +113,19 @@ void backward_reorg_layer(const layer l, network_state state) #ifdef GPU void forward_reorg_layer_gpu(layer l, network_state state) { - reorg_ongpu(state.input, l.w, l.h, l.c, l.batch, l.stride, 1, l.output_gpu); + if(l.reverse){ + reorg_ongpu(state.input, l.w, l.h, l.c, l.batch, l.stride, 1, l.output_gpu); + }else { + reorg_ongpu(state.input, l.w, l.h, l.c, l.batch, l.stride, 0, l.output_gpu); + } } void backward_reorg_layer_gpu(layer l, network_state state) { - reorg_ongpu(l.delta_gpu, l.w, l.h, l.c, l.batch, l.stride, 0, state.delta); + if(l.reverse){ + reorg_ongpu(l.delta_gpu, l.w, l.h, l.c, l.batch, l.stride, 0, state.delta); + }else{ + reorg_ongpu(l.delta_gpu, l.w, l.h, l.c, l.batch, l.stride, 1, state.delta); + } } #endif diff --git a/src/reorg_layer.h b/src/reorg_layer.h index 659bc7cc..21c22cd8 100644 --- a/src/reorg_layer.h +++ b/src/reorg_layer.h @@ -6,7 +6,7 @@ #include "layer.h" #include "network.h" -layer make_reorg_layer(int batch, int h, int w, int c, int stride); +layer make_reorg_layer(int batch, int h, int w, int c, int stride, int reverse); void resize_reorg_layer(layer *l, int w, int h); void forward_reorg_layer(const layer l, network_state state); void backward_reorg_layer(const layer l, network_state state);