From 2f62fe33c913cd9484fe7f2486889d12292c66e0 Mon Sep 17 00:00:00 2001 From: Joseph Redmon Date: Fri, 6 Feb 2015 18:53:53 -0800 Subject: [PATCH] saving weight files as binaries, hell yeah --- Makefile | 6 +-- src/connected_layer.c | 2 - src/convolutional_layer.c | 3 -- src/darknet.c | 55 +++++++++++++++++++-------- src/parser.c | 80 ++++++++++++++++++++++++++++++++++++++- src/parser.h | 2 + src/server.c | 22 ----------- src/utils.c | 22 +++++++++++ src/utils.h | 2 + 9 files changed, 147 insertions(+), 47 deletions(-) diff --git a/Makefile b/Makefile index f685bb4ae89..879ff8e4482 100644 --- a/Makefile +++ b/Makefile @@ -12,13 +12,13 @@ OPTS=-O3 LDFLAGS=`pkg-config --libs opencv` -lm -pthread COMMON=`pkg-config --cflags opencv` -I/usr/local/cuda/include/ CFLAGS=-Wall -Wfatal-errors -CFLAGS+=$(OPTS) ifeq ($(DEBUG), 1) -COMMON+=-O0 -g -CFLAGS+=-O0 -g +OPTS=-O0 -g endif +CFLAGS+=$(OPTS) + ifeq ($(GPU), 1) COMMON+=-DGPU CFLAGS+=-DGPU diff --git a/src/connected_layer.c b/src/connected_layer.c index 1a5fc2b173f..642570c9f5c 100644 --- a/src/connected_layer.c +++ b/src/connected_layer.c @@ -36,14 +36,12 @@ connected_layer *make_connected_layer(int batch, int inputs, int outputs, ACTIVA float scale = 1./sqrt(inputs); - //scale = .01; for(i = 0; i < inputs*outputs; ++i){ layer->weights[i] = scale*rand_normal(); } for(i = 0; i < outputs; ++i){ layer->biases[i] = scale; - // layer->biases[i] = 1; } #ifdef GPU diff --git a/src/convolutional_layer.c b/src/convolutional_layer.c index 62118e4dad4..6a172aa2ad4 100644 --- a/src/convolutional_layer.c +++ b/src/convolutional_layer.c @@ -66,12 +66,9 @@ convolutional_layer *make_convolutional_layer(int batch, int h, int w, int c, in layer->biases = calloc(n, sizeof(float)); layer->bias_updates = calloc(n, sizeof(float)); float scale = 1./sqrt(size*size*c); - //scale = .01; for(i = 0; i < c*n*size*size; ++i) layer->filters[i] = scale*rand_normal(); for(i = 0; i < n; ++i){ - //layer->biases[i] = rand_normal()*scale + scale; layer->biases[i] = scale; - //layer->biases[i] = 1; } int out_h = convolutional_out_height(*layer); int out_w = convolutional_out_width(*layer); diff --git a/src/darknet.c b/src/darknet.c index cc3fc0724ef..ab4c7ada6cb 100644 --- a/src/darknet.c +++ b/src/darknet.c @@ -222,13 +222,16 @@ char *basename(char *cfgfile) return c; } -void train_imagenet(char *cfgfile) +void train_imagenet(char *cfgfile, char *weightfile) { float avg_loss = -1; srand(time(0)); char *base = basename(cfgfile); printf("%s\n", base); network net = parse_network_cfg(cfgfile); + if(weightfile){ + load_weights(&net, weightfile); + } //test_learn_bias(*(convolutional_layer *)net.layers[1]); //set_learning_network(&net, net.learning_rate, 0, net.decay); printf("Learning Rate: %g, Momentum: %g, Decay: %g\n", net.learning_rate, net.momentum, net.decay); @@ -259,16 +262,19 @@ void train_imagenet(char *cfgfile) free_data(train); if(i%100==0){ char buff[256]; - sprintf(buff, "/home/pjreddie/imagenet_backup/%s_%d.cfg",base, i); - save_network(net, buff); + sprintf(buff, "/home/pjreddie/imagenet_backup/%s_%d.weights",base, i); + save_weights(net, buff); } } } -void validate_imagenet(char *filename) +void validate_imagenet(char *filename, char *weightfile) { int i = 0; network net = parse_network_cfg(filename); + if(weightfile){ + load_weights(&net, weightfile); + } srand(time(0)); char **labels = get_labels("/home/pjreddie/data/imagenet/cls.val.labels.list"); @@ -370,14 +376,14 @@ void test_dog(char *cfgfile) float *X = im.data; network net = parse_network_cfg(cfgfile); set_batch_network(&net, 1); - float *predictions = network_predict(net, X); + network_predict(net, X); image crop = get_network_image_layer(net, 0); - //show_image(crop, "cropped"); - // print_image(crop); - //show_image(im, "orig"); + show_image(crop, "cropped"); + print_image(crop); + show_image(im, "orig"); float * inter = get_network_output(net); pm(1000, 1, inter); - //cvWaitKey(0); + cvWaitKey(0); } void test_imagenet(char *cfgfile) @@ -586,7 +592,6 @@ void test_convolutional_layer() float *in = calloc(size, sizeof(float)); int i; for(i = 0; i < size; ++i) in[i] = rand_normal(); - float *in_gpu = cuda_make_array(in, size); convolutional_layer layer = *(convolutional_layer *)net.layers[0]; int out_size = convolutional_out_height(layer)*convolutional_out_width(layer)*layer.batch; cuda_compare(layer.output_gpu, layer.output, out_size, "nothing"); @@ -703,14 +708,18 @@ void del_arg(int argc, char **argv, int index) { int i; for(i = index; i < argc-1; ++i) argv[i] = argv[i+1]; + argv[i] = 0; } int find_arg(int argc, char* argv[], char *arg) { int i; - for(i = 0; i < argc; ++i) if(0==strcmp(argv[i], arg)) { - del_arg(argc, argv, i); - return 1; + for(i = 0; i < argc; ++i) { + if(!argv[i]) continue; + if(0==strcmp(argv[i], arg)) { + del_arg(argc, argv, i); + return 1; + } } return 0; } @@ -719,6 +728,7 @@ int find_int_arg(int argc, char **argv, char *arg, int def) { int i; for(i = 0; i < argc-1; ++i){ + if(!argv[i]) continue; if(0==strcmp(argv[i], arg)){ def = atoi(argv[i+1]); del_arg(argc, argv, i); @@ -729,6 +739,20 @@ int find_int_arg(int argc, char **argv, char *arg, int def) return def; } +void scale_rate(char *filename, float scale) +{ + // Ready for some weird shit?? + FILE *fp = fopen(filename, "r+b"); + if(!fp) file_error(filename); + float rate = 0; + fread(&rate, sizeof(float), 1, fp); + printf("Scaling learning rate from %f to %f\n", rate, rate*scale); + rate = rate*scale; + fseek(fp, 0, SEEK_SET); + fwrite(&rate, sizeof(float), 1, fp); + fclose(fp); +} + int main(int argc, char **argv) { //test_convolutional_layer(); @@ -765,12 +789,12 @@ int main(int argc, char **argv) else if(0==strcmp(argv[1], "ctrain")) train_cifar10(argv[2]); else if(0==strcmp(argv[1], "nist")) train_nist(argv[2]); else if(0==strcmp(argv[1], "ctest")) test_cifar10(argv[2]); - else if(0==strcmp(argv[1], "train")) train_imagenet(argv[2]); + else if(0==strcmp(argv[1], "train")) train_imagenet(argv[2], (argc > 3)? argv[3] : 0); //else if(0==strcmp(argv[1], "client")) train_imagenet_distributed(argv[2]); else if(0==strcmp(argv[1], "detect")) test_detection(argv[2]); else if(0==strcmp(argv[1], "init")) test_init(argv[2]); else if(0==strcmp(argv[1], "visualize")) test_visualize(argv[2]); - else if(0==strcmp(argv[1], "valid")) validate_imagenet(argv[2]); + else if(0==strcmp(argv[1], "valid")) validate_imagenet(argv[2], (argc > 3)? argv[3] : 0); else if(0==strcmp(argv[1], "testnist")) test_nist(argv[2]); else if(0==strcmp(argv[1], "validetect")) validate_detection_net(argv[2]); else if(argc < 4){ @@ -778,6 +802,7 @@ int main(int argc, char **argv) return 0; } else if(0==strcmp(argv[1], "compare")) compare_nist(argv[2], argv[3]); + else if(0==strcmp(argv[1], "scale")) scale_rate(argv[2], atof(argv[3])); fprintf(stderr, "Success!\n"); return 0; } diff --git a/src/parser.c b/src/parser.c index a00feec111d..6a107ccb007 100644 --- a/src/parser.c +++ b/src/parser.c @@ -103,7 +103,7 @@ convolutional_layer *parse_convolutional(list *options, network *net, int count) parse_data(weights, layer->filters, c*n*size*size); parse_data(biases, layer->biases, n); #ifdef GPU - push_convolutional_layer(*layer); + if(weights || biases) push_convolutional_layer(*layer); #endif option_unused(options); return layer; @@ -137,7 +137,7 @@ connected_layer *parse_connected(list *options, network *net, int count) parse_data(biases, layer->biases, output); parse_data(weights, layer->weights, input*output); #ifdef GPU - push_connected_layer(*layer); + if(weights || biases) push_connected_layer(*layer); #endif option_unused(options); return layer; @@ -597,6 +597,82 @@ void print_cost_cfg(FILE *fp, cost_layer *l, network net, int count) fprintf(fp, "\n"); } +void save_weights(network net, char *filename) +{ + printf("Saving weights to %s\n", filename); + FILE *fp = fopen(filename, "w"); + if(!fp) file_error(filename); + + fwrite(&net.learning_rate, sizeof(float), 1, fp); + fwrite(&net.momentum, sizeof(float), 1, fp); + fwrite(&net.decay, sizeof(float), 1, fp); + fwrite(&net.seen, sizeof(int), 1, fp); + + int i; + for(i = 0; i < net.n; ++i){ + if(net.types[i] == CONVOLUTIONAL){ + convolutional_layer layer = *(convolutional_layer *) net.layers[i]; + #ifdef GPU + if(gpu_index >= 0){ + pull_convolutional_layer(layer); + } + #endif + int num = layer.n*layer.c*layer.size*layer.size; + fwrite(layer.biases, sizeof(float), layer.n, fp); + fwrite(layer.filters, sizeof(float), num, fp); + } + if(net.types[i] == CONNECTED){ + connected_layer layer = *(connected_layer *) net.layers[i]; + #ifdef GPU + if(gpu_index >= 0){ + pull_connected_layer(layer); + } + #endif + fwrite(layer.biases, sizeof(float), layer.outputs, fp); + fwrite(layer.weights, sizeof(float), layer.outputs*layer.inputs, fp); + } + } + fclose(fp); +} + +void load_weights(network *net, char *filename) +{ + printf("Loading weights from %s\n", filename); + FILE *fp = fopen(filename, "r"); + if(!fp) file_error(filename); + + fread(&net->learning_rate, sizeof(float), 1, fp); + fread(&net->momentum, sizeof(float), 1, fp); + fread(&net->decay, sizeof(float), 1, fp); + fread(&net->seen, sizeof(int), 1, fp); + set_learning_network(net, net->learning_rate, net->momentum, net->decay); + + int i; + for(i = 0; i < net->n; ++i){ + if(net->types[i] == CONVOLUTIONAL){ + convolutional_layer layer = *(convolutional_layer *) net->layers[i]; + int num = layer.n*layer.c*layer.size*layer.size; + fread(layer.biases, sizeof(float), layer.n, fp); + fread(layer.filters, sizeof(float), num, fp); + #ifdef GPU + if(gpu_index >= 0){ + push_convolutional_layer(layer); + } + #endif + } + if(net->types[i] == CONNECTED){ + connected_layer layer = *(connected_layer *) net->layers[i]; + fread(layer.biases, sizeof(float), layer.outputs, fp); + fread(layer.weights, sizeof(float), layer.outputs*layer.inputs, fp); + #ifdef GPU + if(gpu_index >= 0){ + push_connected_layer(layer); + } + #endif + } + } + fclose(fp); +} void save_network(network net, char *filename) { diff --git a/src/parser.h b/src/parser.h index 891e658b45c..2e8190e34cc 100644 --- a/src/parser.h +++ b/src/parser.h @@ -4,5 +4,7 @@ network parse_network_cfg(char *filename); void save_network(network net, char *filename); +void save_weights(network net, char *filename); +void load_weights(network *net, char *filename); #endif diff --git a/src/server.c b/src/server.c index 788ac876a95..6e5105e8b39 100644 --- a/src/server.c +++ b/src/server.c @@ -50,28 +50,6 @@ typedef struct{ network net; } connection_info; -void read_all(int fd, char *buffer, size_t bytes) -{ - //printf("Want %d\n", bytes); - size_t n = 0; - while(n < bytes){ - int next = read(fd, buffer + n, bytes-n); - if(next <= 0) error("read failed"); - n += next; - } -} - -void write_all(int fd, char *buffer, size_t bytes) -{ - //printf("Writ %d\n", bytes); - size_t n = 0; - while(n < bytes){ - int next = write(fd, buffer + n, bytes-n); - if(next <= 0) error("write failed"); - n += next; - } -} - void read_and_add_into(int fd, float *a, int n) { float *buff = calloc(n, sizeof(float)); diff --git a/src/utils.c b/src/utils.c index 826168224ba..bf02ff3a75c 100644 --- a/src/utils.c +++ b/src/utils.c @@ -2,6 +2,7 @@ #include #include #include +#include #include #include @@ -148,6 +149,27 @@ char *fgetl(FILE *fp) return line; } +void read_all(int fd, char *buffer, size_t bytes) +{ + size_t n = 0; + while(n < bytes){ + int next = read(fd, buffer + n, bytes-n); + if(next <= 0) error("read failed"); + n += next; + } +} + +void write_all(int fd, char *buffer, size_t bytes) +{ + size_t n = 0; + while(n < bytes){ + size_t next = write(fd, buffer + n, bytes-n); + if(next <= 0) error("write failed"); + n += next; + } +} + + char *copy_string(char *s) { char *copy = malloc(strlen(s)+1); diff --git a/src/utils.h b/src/utils.h index daf3a413a5b..e233da853f2 100644 --- a/src/utils.h +++ b/src/utils.h @@ -4,6 +4,8 @@ #include #include "list.h" +void read_all(int fd, char *buffer, size_t bytes); +void write_all(int fd, char *buffer, size_t bytes); char *find_replace(char *str, char *orig, char *rep); void error(const char *s); void malloc_error();