Hi Harley

Joseph Redmon

2016-03-15 e3ee7b7cd64894144a594c0caaa202f7aeac6723

 src/parser.c

@@ -11,6 +11,8 @@
#include "normalization_layer.h"
#include "deconvolutional_layer.h"
#include "connected_layer.h"
#include "rnn_layer.h"
#include "crnn_layer.h"
#include "maxpool_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
@@ -34,6 +36,8 @@
int is_local(section *s);
int is_deconvolutional(section *s);
int is_connected(section *s);
int is_rnn(section *s);
int is_crnn(section *s);
int is_maxpool(section *s);
int is_avgpool(section *s);
int is_dropout(section *s);
@@ -85,6 +89,7 @@
    int w;
    int c;
    int index;
    int time_steps;
} size_params;

deconvolutional_layer parse_deconvolutional(list *options, size_params params)
@@ -151,9 +156,11 @@
    batch=params.batch;
    if(!(h && w && c)) error("Layer before convolutional layer must output image.");
    int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
    int binary = option_find_int_quiet(options, "binary", 0);

    convolutional_layer layer = make_convolutional_layer(batch,h,w,c,n,size,stride,pad,activation, batch_normalize);
    convolutional_layer layer = make_convolutional_layer(batch,h,w,c,n,size,stride,pad,activation, batch_normalize, binary);
    layer.flipped = option_find_int_quiet(options, "flipped", 0);
    layer.dot = option_find_float_quiet(options, "dot", 0);

    char *weights = option_find_str(options, "weights", 0);
    char *biases = option_find_str(options, "biases", 0);
@@ -165,13 +172,45 @@
    return layer;
}

layer parse_crnn(list *options, size_params params)
{
    int output_filters = option_find_int(options, "output_filters",1);
    int hidden_filters = option_find_int(options, "hidden_filters",1);
    char *activation_s = option_find_str(options, "activation", "logistic");
    ACTIVATION activation = get_activation(activation_s);
    int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);

    layer l = make_crnn_layer(params.batch, params.w, params.h, params.c, hidden_filters, output_filters, params.time_steps, activation, batch_normalize);

    l.shortcut = option_find_int_quiet(options, "shortcut", 0);

    return l;
}

layer parse_rnn(list *options, size_params params)
{
    int output = option_find_int(options, "output",1);
    int hidden = option_find_int(options, "hidden",1);
    char *activation_s = option_find_str(options, "activation", "logistic");
    ACTIVATION activation = get_activation(activation_s);
    int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);
    int logistic = option_find_int_quiet(options, "logistic", 0);

    layer l = make_rnn_layer(params.batch, params.inputs, hidden, output, params.time_steps, activation, batch_normalize, logistic);

    l.shortcut = option_find_int_quiet(options, "shortcut", 0);

    return l;
}

connected_layer parse_connected(list *options, size_params params)
{
    int output = option_find_int(options, "output",1);
    char *activation_s = option_find_str(options, "activation", "logistic");
    ACTIVATION activation = get_activation(activation_s);
    int batch_normalize = option_find_int_quiet(options, "batch_normalize", 0);

    connected_layer layer = make_connected_layer(params.batch, params.inputs, output, activation);
    connected_layer layer = make_connected_layer(params.batch, params.inputs, output, activation, batch_normalize);

    char *weights = option_find_str(options, "weights", 0);
    char *biases = option_find_str(options, "biases", 0);
@@ -185,8 +224,9 @@

softmax_layer parse_softmax(list *options, size_params params)
{
    int groups = option_find_int(options, "groups",1);
    int groups = option_find_int_quiet(options, "groups",1);
    softmax_layer layer = make_softmax_layer(params.batch, params.inputs, groups);
    layer.temperature = option_find_float_quiet(options, "temperature", 1);
    return layer;
}

@@ -388,13 +428,16 @@
    net->momentum = option_find_float(options, "momentum", .9);
    net->decay = option_find_float(options, "decay", .0001);
    int subdivs = option_find_int(options, "subdivisions",1);
    net->time_steps = option_find_int_quiet(options, "time_steps",1);
    net->batch /= subdivs;
    net->batch *= net->time_steps;
    net->subdivisions = subdivs;

    net->h = option_find_int_quiet(options, "height",0);
    net->w = option_find_int_quiet(options, "width",0);
    net->c = option_find_int_quiet(options, "channels",0);
    net->inputs = option_find_int_quiet(options, "inputs", net->h * net->w * net->c);
    net->max_crop = option_find_int_quiet(options, "max_crop",net->w*2);

    if(!net->inputs && !(net->h && net->w && net->c)) error("No input parameters supplied");

@@ -456,6 +499,7 @@
    params.c = net.c;
    params.inputs = net.inputs;
    params.batch = net.batch;
    params.time_steps = net.time_steps;

    n = n->next;
    int count = 0;
@@ -474,6 +518,10 @@
            l = parse_activation(options, params);
        }else if(is_deconvolutional(s)){
            l = parse_deconvolutional(options, params);
        }else if(is_rnn(s)){
            l = parse_rnn(options, params);
        }else if(is_crnn(s)){
            l = parse_crnn(options, params);
        }else if(is_connected(s)){
            l = parse_connected(options, params);
        }else if(is_crop(s)){
@@ -564,6 +612,14 @@
    return (strcmp(s->type, "[net]")==0
            || strcmp(s->type, "[network]")==0);
}
int is_crnn(section *s)
{
    return (strcmp(s->type, "[crnn]")==0);
}
int is_rnn(section *s)
{
    return (strcmp(s->type, "[rnn]")==0);
}
int is_connected(section *s)
{
    return (strcmp(s->type, "[conn]")==0
@@ -674,6 +730,39 @@
    fclose(fp);
}

void save_convolutional_weights(layer l, FILE *fp)
{
#ifdef GPU
    if(gpu_index >= 0){
        pull_convolutional_layer(l);
    }
#endif
    int num = l.n*l.c*l.size*l.size;
    fwrite(l.biases, sizeof(float), l.n, fp);
    if (l.batch_normalize){
        fwrite(l.scales, sizeof(float), l.n, fp);
        fwrite(l.rolling_mean, sizeof(float), l.n, fp);
        fwrite(l.rolling_variance, sizeof(float), l.n, fp);
    }
    fwrite(l.filters, sizeof(float), num, fp);
}

void save_connected_weights(layer l, FILE *fp)
{
#ifdef GPU
    if(gpu_index >= 0){
        pull_connected_layer(l);
    }
#endif
    fwrite(l.biases, sizeof(float), l.outputs, fp);
    fwrite(l.weights, sizeof(float), l.outputs*l.inputs, fp);
    if (l.batch_normalize){
        fwrite(l.scales, sizeof(float), l.outputs, fp);
        fwrite(l.rolling_mean, sizeof(float), l.outputs, fp);
        fwrite(l.rolling_variance, sizeof(float), l.outputs, fp);
    }
}

void save_weights_upto(network net, char *filename, int cutoff)
{
    fprintf(stderr, "Saving weights to %s\n", filename);
@@ -692,27 +781,17 @@
    for(i = 0; i < net.n && i < cutoff; ++i){
        layer l = net.layers[i];
        if(l.type == CONVOLUTIONAL){
#ifdef GPU
            if(gpu_index >= 0){
                pull_convolutional_layer(l);
            }
#endif
            int num = l.n*l.c*l.size*l.size;
            fwrite(l.biases, sizeof(float), l.n, fp);
            if (l.batch_normalize){
                fwrite(l.scales, sizeof(float), l.n, fp);
                fwrite(l.rolling_mean, sizeof(float), l.n, fp);
                fwrite(l.rolling_variance, sizeof(float), l.n, fp);
            }
            fwrite(l.filters, sizeof(float), num, fp);
            save_convolutional_weights(l, fp);
        } if(l.type == CONNECTED){
#ifdef GPU
            if(gpu_index >= 0){
                pull_connected_layer(l);
            }
#endif
            fwrite(l.biases, sizeof(float), l.outputs, fp);
            fwrite(l.weights, sizeof(float), l.outputs*l.inputs, fp);
            save_connected_weights(l, fp);
        } if(l.type == RNN){
            save_connected_weights(*(l.input_layer), fp);
            save_connected_weights(*(l.self_layer), fp);
            save_connected_weights(*(l.output_layer), fp);
        } if(l.type == CRNN){
            save_convolutional_weights(*(l.input_layer), fp);
            save_convolutional_weights(*(l.self_layer), fp);
            save_convolutional_weights(*(l.output_layer), fp);
        } if(l.type == LOCAL){
#ifdef GPU
            if(gpu_index >= 0){
@@ -745,11 +824,59 @@
    free(transpose);
}

void load_connected_weights(layer l, FILE *fp, int transpose)
{
    fread(l.biases, sizeof(float), l.outputs, fp);
    fread(l.weights, sizeof(float), l.outputs*l.inputs, fp);
    if(transpose){
        transpose_matrix(l.weights, l.inputs, l.outputs);
    }
    if (l.batch_normalize && (!l.dontloadscales)){
        fread(l.scales, sizeof(float), l.outputs, fp);
        fread(l.rolling_mean, sizeof(float), l.outputs, fp);
        fread(l.rolling_variance, sizeof(float), l.outputs, fp);
    }
#ifdef GPU
    if(gpu_index >= 0){
        push_connected_layer(l);
    }
#endif
}

void load_convolutional_weights(layer l, FILE *fp)
{
    int num = l.n*l.c*l.size*l.size;
    fread(l.biases, sizeof(float), l.n, fp);
    if (l.batch_normalize && (!l.dontloadscales)){
        fread(l.scales, sizeof(float), l.n, fp);
        fread(l.rolling_mean, sizeof(float), l.n, fp);
        fread(l.rolling_variance, sizeof(float), l.n, fp);
        /*
        int i;
        for(i = 0; i < l.n; ++i){
            if(l.rolling_mean[i] > 1 || l.rolling_mean[i] < -1 || l.rolling_variance[i] > 1 || l.rolling_variance[i] < -1)
            printf("%f %f\n", l.rolling_mean[i], l.rolling_variance[i]);
        }
        */
    }
    fflush(stdout);
    fread(l.filters, sizeof(float), num, fp);
    if (l.flipped) {
        transpose_matrix(l.filters, l.c*l.size*l.size, l.n);
    }
#ifdef GPU
    if(gpu_index >= 0){
        push_convolutional_layer(l);
    }
#endif
}


void load_weights_upto(network *net, char *filename, int cutoff)
{
    fprintf(stderr, "Loading weights from %s...", filename);
    fflush(stdout);
    FILE *fp = fopen(filename, "r");
    FILE *fp = fopen(filename, "rb");
    if(!fp) file_error(filename);

    int major;
@@ -759,28 +886,14 @@
    fread(&minor, sizeof(int), 1, fp);
    fread(&revision, sizeof(int), 1, fp);
    fread(net->seen, sizeof(int), 1, fp);
    int transpose = (major > 1000) || (minor > 1000);

    int i;
    for(i = 0; i < net->n && i < cutoff; ++i){
        layer l = net->layers[i];
        if (l.dontload) continue;
        if(l.type == CONVOLUTIONAL){
            int num = l.n*l.c*l.size*l.size;
            fread(l.biases, sizeof(float), l.n, fp);
            if (l.batch_normalize && (!l.dontloadscales)){
                fread(l.scales, sizeof(float), l.n, fp);
                fread(l.rolling_mean, sizeof(float), l.n, fp);
                fread(l.rolling_variance, sizeof(float), l.n, fp);
            }
            fread(l.filters, sizeof(float), num, fp);
            if (l.flipped) {
                transpose_matrix(l.filters, l.c*l.size*l.size, l.n);
            }
#ifdef GPU
            if(gpu_index >= 0){
                push_convolutional_layer(l);
            }
#endif
            load_convolutional_weights(l, fp);
        }
        if(l.type == DECONVOLUTIONAL){
            int num = l.n*l.c*l.size*l.size;
@@ -793,16 +906,17 @@
#endif
        }
        if(l.type == CONNECTED){
            fread(l.biases, sizeof(float), l.outputs, fp);
            fread(l.weights, sizeof(float), l.outputs*l.inputs, fp);
            if(major > 1000 || minor > 1000){
                transpose_matrix(l.weights, l.inputs, l.outputs);
            }
#ifdef GPU
            if(gpu_index >= 0){
                push_connected_layer(l);
            }
#endif
            load_connected_weights(l, fp, transpose);
        }
        if(l.type == CRNN){
            load_convolutional_weights(*(l.input_layer), fp);
            load_convolutional_weights(*(l.self_layer), fp);
            load_convolutional_weights(*(l.output_layer), fp);
        }
        if(l.type == RNN){
            load_connected_weights(*(l.input_layer), fp, transpose);
            load_connected_weights(*(l.self_layer), fp, transpose);
            load_connected_weights(*(l.output_layer), fp, transpose);
        }
        if(l.type == LOCAL){
            int locations = l.out_w*l.out_h;