diff --git a/Makefile b/Makefile index 1ef1b3bb..e6899417 100644 --- a/Makefile +++ b/Makefile @@ -1,5 +1,5 @@ -GPU=0 -OPENCV=0 +GPU=1 +OPENCV=1 DEBUG=0 ARCH= --gpu-architecture=compute_20 --gpu-code=compute_20 @@ -34,7 +34,7 @@ CFLAGS+= -DGPU LDFLAGS+= -L/usr/local/cuda/lib64 -lcuda -lcudart -lcublas -lcurand endif -OBJ=gemm.o utils.o cuda.o deconvolutional_layer.o convolutional_layer.o list.o image.o activations.o im2col.o col2im.o blas.o crop_layer.o dropout_layer.o maxpool_layer.o softmax_layer.o data.o matrix.o network.o connected_layer.o cost_layer.o parser.o option_list.o darknet.o detection_layer.o imagenet.o captcha.o route_layer.o writing.o box.o nightmare.o normalization_layer.o avgpool_layer.o coco.o dice.o yolo.o layer.o compare.o classifier.o local_layer.o swag.o shortcut_layer.o activation_layer.o rnn_layer.o gru_layer.o rnn.o rnn_vid.o crnn_layer.o coco_demo.o tag.o cifar.o yolo_demo.o go.o batchnorm_layer.o +OBJ=gemm.o utils.o cuda.o deconvolutional_layer.o convolutional_layer.o list.o image.o activations.o im2col.o col2im.o blas.o crop_layer.o dropout_layer.o maxpool_layer.o softmax_layer.o data.o matrix.o network.o connected_layer.o cost_layer.o parser.o option_list.o darknet.o detection_layer.o imagenet.o captcha.o route_layer.o writing.o box.o nightmare.o normalization_layer.o avgpool_layer.o coco.o dice.o yolo2.o layer.o compare.o classifier.o local_layer.o swag.o shortcut_layer.o activation_layer.o rnn_layer.o gru_layer.o rnn.o rnn_vid.o crnn_layer.o coco_demo.o tag.o cifar.o yolo_demo.o go.o batchnorm_layer.o ifeq ($(GPU), 1) LDFLAGS+= -lstdc++ OBJ+=convolutional_kernels.o deconvolutional_kernels.o activation_kernels.o im2col_kernels.o col2im_kernels.o blas_kernels.o crop_layer_kernels.o dropout_layer_kernels.o maxpool_layer_kernels.o softmax_layer_kernels.o network_kernels.o avgpool_layer_kernels.o diff --git a/src/batchnorm_layer.c b/src/batchnorm_layer.c index 6ea40407..9b68277e 100644 --- a/src/batchnorm_layer.c +++ b/src/batchnorm_layer.c @@ -135,6 +135,20 @@ void backward_batchnorm_layer(const layer layer, network_state state) } #ifdef GPU + +void pull_batchnorm_layer(layer l) +{ + cuda_pull_array(l.scales_gpu, l.scales, l.c); + cuda_pull_array(l.rolling_mean_gpu, l.rolling_mean, l.c); + cuda_pull_array(l.rolling_variance_gpu, l.rolling_variance, l.c); +} +void push_batchnorm_layer(layer l) +{ + cuda_push_array(l.scales_gpu, l.scales, l.c); + cuda_push_array(l.rolling_mean_gpu, l.rolling_mean, l.c); + cuda_push_array(l.rolling_variance_gpu, l.rolling_variance, l.c); +} + void forward_batchnorm_layer_gpu(layer l, network_state state) { if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, state.input, 1, l.output_gpu, 1); diff --git a/src/batchnorm_layer.h b/src/batchnorm_layer.h index 61810b68..99d1d0fe 100644 --- a/src/batchnorm_layer.h +++ b/src/batchnorm_layer.h @@ -12,6 +12,8 @@ void backward_batchnorm_layer(layer l, network_state state); #ifdef GPU void forward_batchnorm_layer_gpu(layer l, network_state state); void backward_batchnorm_layer_gpu(layer l, network_state state); +void pull_batchnorm_layer(layer l); +void push_batchnorm_layer(layer l); #endif #endif diff --git a/src/data.c b/src/data.c index b0368eeb..fdc4a1db 100644 --- a/src/data.c +++ b/src/data.c @@ -271,78 +271,37 @@ void fill_truth_region(char *path, float *truth, int classes, int num_boxes, int free(boxes); } -void fill_truth_detection(char *path, float *truth, int classes, int num_boxes, int flip, int background, float dx, float dy, float sx, float sy) +void fill_truth_detection(char *path, float *truth, int classes, int flip, float dx, float dy, float sx, float sy) { - char *labelpath = find_replace(path, "JPEGImages", "labels"); + char *labelpath = find_replace(path, "images", "labels"); + labelpath = find_replace(labelpath, "JPEGImages", "labels"); + labelpath = find_replace(labelpath, ".jpg", ".txt"); + labelpath = find_replace(labelpath, ".JPG", ".txt"); labelpath = find_replace(labelpath, ".JPEG", ".txt"); int count = 0; box_label *boxes = read_boxes(labelpath, &count); randomize_boxes(boxes, count); + correct_boxes(boxes, count, dx, dy, sx, sy, flip); + if(count > 17) count = 17; float x,y,w,h; - float left, top, right, bot; int id; int i; - if(background){ - for(i = 0; i < num_boxes*num_boxes*(4+classes+background); i += 4+classes+background){ - truth[i] = 1; - } - } - for(i = 0; i < count; ++i){ - left = boxes[i].left * sx - dx; - right = boxes[i].right * sx - dx; - top = boxes[i].top * sy - dy; - bot = boxes[i].bottom* sy - dy; - id = boxes[i].id; - - if(flip){ - float swap = left; - left = 1. - right; - right = 1. - swap; - } - - left = constrain(0, 1, left); - right = constrain(0, 1, right); - top = constrain(0, 1, top); - bot = constrain(0, 1, bot); - - x = (left+right)/2; - y = (top+bot)/2; - w = (right - left); - h = (bot - top); - - if (x <= 0 || x >= 1 || y <= 0 || y >= 1) continue; - int col = (int)(x*num_boxes); - int row = (int)(y*num_boxes); - - x = x*num_boxes - col; - y = y*num_boxes - row; + for (i = 0; i < count; ++i) { + x = boxes[i].x; + y = boxes[i].y; + w = boxes[i].w; + h = boxes[i].h; + id = boxes[i].id; - /* - float maxwidth = distance_from_edge(i, num_boxes); - float maxheight = distance_from_edge(j, num_boxes); - w = w/maxwidth; - h = h/maxheight; - */ - - w = constrain(0, 1, w); - h = constrain(0, 1, h); if (w < .01 || h < .01) continue; - if(1){ - w = pow(w, 1./2.); - h = pow(h, 1./2.); - } - int index = (col+row*num_boxes)*(4+classes+background); - if(truth[index+classes+background+2]) continue; - if(background) truth[index++] = 0; - truth[index+id] = 1; - index += classes; - truth[index++] = x; - truth[index++] = y; - truth[index++] = w; - truth[index++] = h; + truth[i*5] = id; + truth[i*5+2] = x; + truth[i*5+3] = y; + truth[i*5+4] = w; + truth[i*5+5] = h; } free(boxes); } @@ -485,6 +444,7 @@ data load_data_region(int n, char **paths, int m, int w, int h, int size, int cl d.X.vals = calloc(d.X.rows, sizeof(float*)); d.X.cols = h*w*3; + int k = size*size*(5+classes); d.y = make_matrix(n, k); for(i = 0; i < n; ++i){ @@ -641,7 +601,7 @@ data load_data_swag(char **paths, int n, int classes, float jitter) return d; } -data load_data_detection(int n, char **paths, int m, int classes, int w, int h, int num_boxes, int background) +data load_data_detection(int n, int boxes, char **paths, int m, int w, int h, int classes, float jitter) { char **random_paths = get_random_paths(paths, n, m); int i; @@ -652,16 +612,15 @@ data load_data_detection(int n, char **paths, int m, int classes, int w, int h, d.X.vals = calloc(d.X.rows, sizeof(float*)); d.X.cols = h*w*3; - int k = num_boxes*num_boxes*(4+classes+background); - d.y = make_matrix(n, k); + d.y = make_matrix(n, 5*boxes); for(i = 0; i < n; ++i){ image orig = load_image_color(random_paths[i], 0, 0); int oh = orig.h; int ow = orig.w; - int dw = ow/10; - int dh = oh/10; + int dw = (ow*jitter); + int dh = (oh*jitter); int pleft = rand_uniform(-dw, dw); int pright = rand_uniform(-dw, dw); @@ -674,13 +633,6 @@ data load_data_detection(int n, char **paths, int m, int classes, int w, int h, float sx = (float)swidth / ow; float sy = (float)sheight / oh; - /* - float angle = rand_uniform()*.1 - .05; - image rot = rotate_image(orig, angle); - free_image(orig); - orig = rot; - */ - int flip = rand_r(&data_seed)%2; image cropped = crop_image(orig, pleft, ptop, swidth, sheight); @@ -691,7 +643,7 @@ data load_data_detection(int n, char **paths, int m, int classes, int w, int h, if(flip) flip_image(sized); d.X.vals[i] = sized.data; - fill_truth_detection(random_paths[i], d.y.vals[i], classes, num_boxes, flip, background, dx, dy, 1./sx, 1./sy); + fill_truth_detection(random_paths[i], d.y.vals[i], classes, flip, dx, dy, 1./sx, 1./sy); free_image(orig); free_image(cropped); @@ -700,6 +652,7 @@ data load_data_detection(int n, char **paths, int m, int classes, int w, int h, return d; } + void *load_thread(void *ptr) { @@ -717,7 +670,7 @@ void *load_thread(void *ptr) } else if (a.type == STUDY_DATA){ *a.d = load_data_study(a.paths, a.n, a.m, a.labels, a.classes, a.min, a.max, a.size); } else if (a.type == DETECTION_DATA){ - *a.d = load_data_detection(a.n, a.paths, a.m, a.classes, a.w, a.h, a.num_boxes, a.background); + *a.d = load_data_detection(a.n, a.num_boxes, a.paths, a.m, a.classes, a.w, a.h, a.background); } else if (a.type == WRITING_DATA){ *a.d = load_data_writing(a.paths, a.n, a.m, a.w, a.h, a.out_w, a.out_h); } else if (a.type == REGION_DATA){ diff --git a/src/data.h b/src/data.h index 6befeea5..a7347a82 100644 --- a/src/data.h +++ b/src/data.h @@ -25,10 +25,12 @@ typedef struct{ matrix y; int *indexes; int shallow; + int *num_boxes; + box **boxes; } data; typedef enum { - CLASSIFICATION_DATA, DETECTION_DATA, CAPTCHA_DATA, REGION_DATA, IMAGE_DATA, COMPARE_DATA, WRITING_DATA, SWAG_DATA, TAG_DATA, OLD_CLASSIFICATION_DATA, STUDY_DATA + CLASSIFICATION_DATA, DETECTION_DATA, CAPTCHA_DATA, REGION_DATA, IMAGE_DATA, COMPARE_DATA, WRITING_DATA, SWAG_DATA, TAG_DATA, OLD_CLASSIFICATION_DATA, STUDY_DATA, DET_DATA } data_type; typedef struct load_args{ @@ -68,7 +70,7 @@ void print_letters(float *pred, int n); data load_data_captcha(char **paths, int n, int m, int k, int w, int h); data load_data_captcha_encode(char **paths, int n, int m, int w, int h); data load_data(char **paths, int n, int m, char **labels, int k, int w, int h); -data load_data_detection(int n, char **paths, int m, int classes, int w, int h, int num_boxes, int background); +data load_data_detection(int n, int boxes, char **paths, int m, int w, int h, int classes, float jitter); data load_data_tag(char **paths, int n, int m, int k, int min, int max, int size); data load_data_augment(char **paths, int n, int m, char **labels, int k, int min, int max, int size); data load_data_study(char **paths, int n, int m, char **labels, int k, int min, int max, int size); diff --git a/src/parser.c b/src/parser.c index 6c88fd5c..b900ad78 100644 --- a/src/parser.c +++ b/src/parser.c @@ -852,6 +852,18 @@ void save_convolutional_weights(layer l, FILE *fp) fwrite(l.filters, sizeof(float), num, fp); } +void save_batchnorm_weights(layer l, FILE *fp) +{ +#ifdef GPU + if(gpu_index >= 0){ + pull_batchnorm_layer(l); + } +#endif + fwrite(l.scales, sizeof(float), l.c, fp); + fwrite(l.rolling_mean, sizeof(float), l.c, fp); + fwrite(l.rolling_variance, sizeof(float), l.c, fp); +} + void save_connected_weights(layer l, FILE *fp) { #ifdef GPU @@ -889,6 +901,8 @@ void save_weights_upto(network net, char *filename, int cutoff) save_convolutional_weights(l, fp); } if(l.type == CONNECTED){ save_connected_weights(l, fp); + } if(l.type == BATCHNORM){ + save_batchnorm_weights(l, fp); } if(l.type == RNN){ save_connected_weights(*(l.input_layer), fp); save_connected_weights(*(l.self_layer), fp); @@ -943,8 +957,8 @@ void load_connected_weights(layer l, FILE *fp, int transpose) if(transpose){ transpose_matrix(l.weights, l.inputs, l.outputs); } - //printf("Biases: %f mean %f variance\n", mean_array(l.biases, l.outputs), variance_array(l.biases, l.outputs)); - //printf("Weights: %f mean %f variance\n", mean_array(l.weights, l.outputs*l.inputs), variance_array(l.weights, l.outputs*l.inputs)); + //printf("Biases: %f mean %f variance\n", mean_array(l.biases, l.outputs), variance_array(l.biases, l.outputs)); + //printf("Weights: %f mean %f variance\n", mean_array(l.weights, l.outputs*l.inputs), variance_array(l.weights, l.outputs*l.inputs)); if (l.batch_normalize && (!l.dontloadscales)){ fread(l.scales, sizeof(float), l.outputs, fp); fread(l.rolling_mean, sizeof(float), l.outputs, fp); @@ -960,6 +974,18 @@ void load_connected_weights(layer l, FILE *fp, int transpose) #endif } +void load_batchnorm_weights(layer l, FILE *fp) +{ + fread(l.scales, sizeof(float), l.c, fp); + fread(l.rolling_mean, sizeof(float), l.c, fp); + fread(l.rolling_variance, sizeof(float), l.c, fp); +#ifdef GPU + if(gpu_index >= 0){ + push_batchnorm_layer(l); + } +#endif +} + void load_convolutional_weights_binary(layer l, FILE *fp) { fread(l.biases, sizeof(float), l.n, fp); @@ -1053,6 +1079,9 @@ void load_weights_upto(network *net, char *filename, int cutoff) if(l.type == CONNECTED){ load_connected_weights(l, fp, transpose); } + if(l.type == BATCHNORM){ + load_batchnorm_weights(l, fp); + } if(l.type == CRNN){ load_convolutional_weights(*(l.input_layer), fp); load_convolutional_weights(*(l.self_layer), fp); diff --git a/src/rnn.c b/src/rnn.c index b72fafc7..12f14732 100644 --- a/src/rnn.c +++ b/src/rnn.c @@ -183,7 +183,7 @@ void test_char_rnn(char *cfgfile, char *weightfile, int num, char *seed, float t printf("\n"); } -void valid_char_rnn(char *cfgfile, char *weightfile) +void valid_char_rnn(char *cfgfile, char *weightfile, char *seed) { char *base = basecfg(cfgfile); fprintf(stderr, "%s\n", base); @@ -196,18 +196,22 @@ void valid_char_rnn(char *cfgfile, char *weightfile) int count = 0; int c; + int len = strlen(seed); float *input = calloc(inputs, sizeof(float)); int i; - for(i = 0; i < 100; ++i){ + for(i = 0; i < len; ++i){ + c = seed[i]; + input[(int)c] = 1; network_predict(net, input); + input[(int)c] = 0; } float sum = 0; c = getc(stdin); float log2 = log(2); while(c != EOF){ int next = getc(stdin); - if(next < 0 || next >= 255) error("Out of range character"); if(next == EOF) break; + if(next < 0 || next >= 255) error("Out of range character"); ++count; input[c] = 1; float *out = network_predict(net, input); @@ -218,6 +222,52 @@ void valid_char_rnn(char *cfgfile, char *weightfile) } } +void vec_char_rnn(char *cfgfile, char *weightfile, char *seed) +{ + char *base = basecfg(cfgfile); + fprintf(stderr, "%s\n", base); + + network net = parse_network_cfg(cfgfile); + if(weightfile){ + load_weights(&net, weightfile); + } + int inputs = get_network_input_size(net); + + int c; + int seed_len = strlen(seed); + float *input = calloc(inputs, sizeof(float)); + int i; + char *line; + while((line=fgetl(stdin)) != 0){ + reset_rnn_state(net, 0); + for(i = 0; i < seed_len; ++i){ + c = seed[i]; + input[(int)c] = 1; + network_predict(net, input); + input[(int)c] = 0; + } + strip(line); + int str_len = strlen(line); + for(i = 0; i < str_len; ++i){ + c = line[i]; + input[(int)c] = 1; + network_predict(net, input); + input[(int)c] = 0; + } + c = ' '; + input[(int)c] = 1; + network_predict(net, input); + input[(int)c] = 0; + + layer l = net.layers[0]; + cuda_pull_array(l.output_gpu, l.output, l.outputs); + printf("%s", line); + for(i = 0; i < l.outputs; ++i){ + printf(",%g", l.output[i]); + } + printf("\n"); + } +} void run_char_rnn(int argc, char **argv) { @@ -226,7 +276,7 @@ void run_char_rnn(int argc, char **argv) return; } char *filename = find_char_arg(argc, argv, "-file", "data/shakespeare.txt"); - char *seed = find_char_arg(argc, argv, "-seed", "\n"); + char *seed = find_char_arg(argc, argv, "-seed", "\n\n"); int len = find_int_arg(argc, argv, "-len", 1000); float temp = find_float_arg(argc, argv, "-temp", .7); int rseed = find_int_arg(argc, argv, "-srand", time(0)); @@ -235,6 +285,7 @@ void run_char_rnn(int argc, char **argv) char *cfg = argv[3]; char *weights = (argc > 4) ? argv[4] : 0; if(0==strcmp(argv[2], "train")) train_char_rnn(cfg, weights, filename, clear); - else if(0==strcmp(argv[2], "valid")) valid_char_rnn(cfg, weights); + else if(0==strcmp(argv[2], "valid")) valid_char_rnn(cfg, weights, seed); + else if(0==strcmp(argv[2], "vec")) vec_char_rnn(cfg, weights, seed); else if(0==strcmp(argv[2], "generate")) test_char_rnn(cfg, weights, len, seed, temp, rseed); }