Detection is back, baby\!

Joseph Redmon

2015-02-11 0f645836f193e75c4c3b718369e6fab15b5d19c5

 src/parser.c

@@ -7,6 +7,7 @@
#include "crop_layer.h"
#include "cost_layer.h"
#include "convolutional_layer.h"
#include "deconvolutional_layer.h"
#include "connected_layer.h"
#include "maxpool_layer.h"
#include "normalization_layer.h"
@@ -16,7 +17,6 @@
#include "list.h"
#include "option_list.h"
#include "utils.h"
#include "opencl.h"

typedef struct{
    char *type;
@@ -24,6 +24,7 @@
}section;

int is_convolutional(section *s);
int is_deconvolutional(section *s);
int is_connected(section *s);
int is_maxpool(section *s);
int is_dropout(section *s);
@@ -66,6 +67,49 @@
    }
}

deconvolutional_layer *parse_deconvolutional(list *options, network *net, int count)
{
    int h,w,c;
    float learning_rate, momentum, decay;
    int n = option_find_int(options, "filters",1);
    int size = option_find_int(options, "size",1);
    int stride = option_find_int(options, "stride",1);
    char *activation_s = option_find_str(options, "activation", "sigmoid");
    ACTIVATION activation = get_activation(activation_s);
    if(count == 0){
        learning_rate = option_find_float(options, "learning_rate", .001);
        momentum = option_find_float(options, "momentum", .9);
        decay = option_find_float(options, "decay", .0001);
        h = option_find_int(options, "height",1);
        w = option_find_int(options, "width",1);
        c = option_find_int(options, "channels",1);
        net->batch = option_find_int(options, "batch",1);
        net->learning_rate = learning_rate;
        net->momentum = momentum;
        net->decay = decay;
        net->seen = option_find_int(options, "seen",0);
    }else{
        learning_rate = option_find_float_quiet(options, "learning_rate", net->learning_rate);
        momentum = option_find_float_quiet(options, "momentum", net->momentum);
        decay = option_find_float_quiet(options, "decay", net->decay);
        image m =  get_network_image_layer(*net, count-1);
        h = m.h;
        w = m.w;
        c = m.c;
        if(h == 0) error("Layer before deconvolutional layer must output image.");
    }
    deconvolutional_layer *layer = make_deconvolutional_layer(net->batch,h,w,c,n,size,stride,activation,learning_rate,momentum,decay);
    char *weights = option_find_str(options, "weights", 0);
    char *biases = option_find_str(options, "biases", 0);
    parse_data(weights, layer->filters, c*n*size*size);
    parse_data(biases, layer->biases, n);
    #ifdef GPU
    if(weights || biases) push_deconvolutional_layer(*layer);
    #endif
    option_unused(options);
    return layer;
}

convolutional_layer *parse_convolutional(list *options, network *net, int count)
{
    int h,w,c;
@@ -87,6 +131,7 @@
        net->learning_rate = learning_rate;
        net->momentum = momentum;
        net->decay = decay;
        net->seen = option_find_int(options, "seen",0);
    }else{
        learning_rate = option_find_float_quiet(options, "learning_rate", net->learning_rate);
        momentum = option_find_float_quiet(options, "momentum", net->momentum);
@@ -103,7 +148,7 @@
    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 +182,7 @@
    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;
@@ -149,6 +194,7 @@
    if(count == 0){
        input = option_find_int(options, "input",1);
        net->batch = option_find_int(options, "batch",1);
        net->seen = option_find_int(options, "seen",0);
    }else{
        input =  get_network_output_size_layer(*net, count-1);
    }
@@ -163,6 +209,7 @@
    if(count == 0){
        input = option_find_int(options, "input",1);
        net->batch = option_find_int(options, "batch",1);
        net->seen = option_find_int(options, "seen",0);
    }else{
        input =  get_network_output_size_layer(*net, count-1);
    }
@@ -191,6 +238,7 @@
        net->learning_rate = learning_rate;
        net->momentum = momentum;
        net->decay = decay;
        net->seen = option_find_int(options, "seen",0);
    }else{
        image m =  get_network_image_layer(*net, count-1);
        h = m.h;
@@ -213,6 +261,7 @@
        w = option_find_int(options, "width",1);
        c = option_find_int(options, "channels",1);
        net->batch = option_find_int(options, "batch",1);
        net->seen = option_find_int(options, "seen",0);
    }else{
        image m =  get_network_image_layer(*net, count-1);
        h = m.h;
@@ -225,6 +274,7 @@
    return layer;
}

/*
freeweight_layer *parse_freeweight(list *options, network *net, int count)
{
    int input;
@@ -238,6 +288,7 @@
    option_unused(options);
    return layer;
}
*/

dropout_layer *parse_dropout(list *options, network *net, int count)
{
@@ -252,6 +303,7 @@
        net->learning_rate = learning_rate;
        net->momentum = momentum;
        net->decay = decay;
        net->seen = option_find_int(options, "seen",0);
    }else{
        input =  get_network_output_size_layer(*net, count-1);
    }
@@ -272,6 +324,7 @@
        w = option_find_int(options, "width",1);
        c = option_find_int(options, "channels",1);
        net->batch = option_find_int(options, "batch",1);
        net->seen = option_find_int(options, "seen",0);
    }else{
        image m =  get_network_image_layer(*net, count-1);
        h = m.h;
@@ -298,6 +351,10 @@
            convolutional_layer *layer = parse_convolutional(options, &net, count);
            net.types[count] = CONVOLUTIONAL;
            net.layers[count] = layer;
        }else if(is_deconvolutional(s)){
            deconvolutional_layer *layer = parse_deconvolutional(options, &net, count);
            net.types[count] = DECONVOLUTIONAL;
            net.layers[count] = layer;
        }else if(is_connected(s)){
            connected_layer *layer = parse_connected(options, &net, count);
            net.types[count] = CONNECTED;
@@ -327,9 +384,10 @@
            net.types[count] = DROPOUT;
            net.layers[count] = layer;
        }else if(is_freeweight(s)){
            freeweight_layer *layer = parse_freeweight(options, &net, count);
            net.types[count] = FREEWEIGHT;
            net.layers[count] = layer;
            //freeweight_layer *layer = parse_freeweight(options, &net, count);
            //net.types[count] = FREEWEIGHT;
            //net.layers[count] = layer;
            fprintf(stderr, "Type not recognized: %s\n", s->type);
        }else{
            fprintf(stderr, "Type not recognized: %s\n", s->type);
        }
@@ -351,6 +409,11 @@
{
    return (strcmp(s->type, "[cost]")==0);
}
int is_deconvolutional(section *s)
{
    return (strcmp(s->type, "[deconv]")==0
            || strcmp(s->type, "[deconvolutional]")==0);
}
int is_convolutional(section *s)
{
    return (strcmp(s->type, "[conv]")==0
@@ -429,7 +492,7 @@
                break;
            default:
                if(!read_option(line, current->options)){
                    printf("Config file error line %d, could parse: %s\n", nu, line);
                    fprintf(stderr, "Config file error line %d, could parse: %s\n", nu, line);
                    free(line);
                }
                break;
@@ -442,7 +505,7 @@
void print_convolutional_cfg(FILE *fp, convolutional_layer *l, network net, int count)
{
    #ifdef GPU
    if(gpu_index >= 0) pull_convolutional_layer(*l);
    if(gpu_index >= 0)  pull_convolutional_layer(*l);
    #endif
    int i;
    fprintf(fp, "[convolutional]\n");
@@ -453,8 +516,9 @@
                "channels=%d\n"
                "learning_rate=%g\n"
                "momentum=%g\n"
                "decay=%g\n",
                l->batch,l->h, l->w, l->c, l->learning_rate, l->momentum, l->decay);
                "decay=%g\n"
                "seen=%d\n",
                l->batch,l->h, l->w, l->c, l->learning_rate, l->momentum, l->decay, net.seen);
    } else {
        if(l->learning_rate != net.learning_rate)
            fprintf(fp, "learning_rate=%g\n", l->learning_rate);
@@ -478,6 +542,45 @@
    fprintf(fp, "\n\n");
}

void print_deconvolutional_cfg(FILE *fp, deconvolutional_layer *l, network net, int count)
{
    #ifdef GPU
    if(gpu_index >= 0)  pull_deconvolutional_layer(*l);
    #endif
    int i;
    fprintf(fp, "[deconvolutional]\n");
    if(count == 0) {
        fprintf(fp,   "batch=%d\n"
                "height=%d\n"
                "width=%d\n"
                "channels=%d\n"
                "learning_rate=%g\n"
                "momentum=%g\n"
                "decay=%g\n"
                "seen=%d\n",
                l->batch,l->h, l->w, l->c, l->learning_rate, l->momentum, l->decay, net.seen);
    } else {
        if(l->learning_rate != net.learning_rate)
            fprintf(fp, "learning_rate=%g\n", l->learning_rate);
        if(l->momentum != net.momentum)
            fprintf(fp, "momentum=%g\n", l->momentum);
        if(l->decay != net.decay)
            fprintf(fp, "decay=%g\n", l->decay);
    }
    fprintf(fp, "filters=%d\n"
            "size=%d\n"
            "stride=%d\n"
            "activation=%s\n",
            l->n, l->size, l->stride,
            get_activation_string(l->activation));
    fprintf(fp, "biases=");
    for(i = 0; i < l->n; ++i) fprintf(fp, "%g,", l->biases[i]);
    fprintf(fp, "\n");
    fprintf(fp, "weights=");
    for(i = 0; i < l->n*l->c*l->size*l->size; ++i) fprintf(fp, "%g,", l->filters[i]);
    fprintf(fp, "\n\n");
}

void print_freeweight_cfg(FILE *fp, freeweight_layer *l, network net, int count)
{
    fprintf(fp, "[freeweight]\n");
@@ -508,8 +611,9 @@
                "input=%d\n"
                "learning_rate=%g\n"
                "momentum=%g\n"
                "decay=%g\n",
                l->batch, l->inputs, l->learning_rate, l->momentum, l->decay);
                "decay=%g\n"
                "seen=%d\n",
                l->batch, l->inputs, l->learning_rate, l->momentum, l->decay, net.seen);
    } else {
        if(l->learning_rate != net.learning_rate)
            fprintf(fp, "learning_rate=%g\n", l->learning_rate);
@@ -540,8 +644,9 @@
                "channels=%d\n"
                "learning_rate=%g\n"
                "momentum=%g\n"
                "decay=%g\n",
                l->batch,l->h, l->w, l->c, net.learning_rate, net.momentum, net.decay);
                "decay=%g\n"
                "seen=%d\n",
                l->batch,l->h, l->w, l->c, net.learning_rate, net.momentum, net.decay, net.seen);
    }
    fprintf(fp, "crop_height=%d\ncrop_width=%d\nflip=%d\n\n", l->crop_height, l->crop_width, l->flip);
}
@@ -585,6 +690,104 @@
    fprintf(fp, "\n");
}

void save_weights(network net, char *filename)
{
    fprintf(stderr, "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] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *) net.layers[i];
            #ifdef GPU
            if(gpu_index >= 0){
                pull_deconvolutional_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)
{
    fprintf(stderr, "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] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_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_deconvolutional_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)
{
@@ -595,6 +798,8 @@
    {
        if(net.types[i] == CONVOLUTIONAL)
            print_convolutional_cfg(fp, (convolutional_layer *)net.layers[i], net, i);
        else if(net.types[i] == DECONVOLUTIONAL)
            print_deconvolutional_cfg(fp, (deconvolutional_layer *)net.layers[i], net, i);
        else if(net.types[i] == CONNECTED)
            print_connected_cfg(fp, (connected_layer *)net.layers[i], net, i);
        else if(net.types[i] == CROP)