stable

Joseph Redmon

2016-05-11 9942d484122c346650bb5431fd209d9437b5310a

 src/parser.c

@@ -852,6 +852,18 @@
    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 @@
            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 @@
    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 @@
#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 @@
        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);