im2col slightly faster

Joseph Redmon

2014-10-30 1c0fd9bb4726f28b5ccf4491b8d108b00c884ec3

 src/network.c

@@ -1,4 +1,5 @@
#include <stdio.h>
#include <time.h>
#include "network.h"
#include "image.h"
#include "data.h"
@@ -24,16 +25,20 @@
    net.outputs = 0;
    net.output = 0;
    #ifdef GPU
    net.input_cl = 0;
    net.input_cl = calloc(1, sizeof(cl_mem));
    net.truth_cl = calloc(1, sizeof(cl_mem));
    #endif
    return net;
}

#ifdef GPU

void forward_network_gpu(network net, cl_mem input, cl_mem truth, int train)
{
    //printf("start\n");
    int i;
    for(i = 0; i < net.n; ++i){
        clock_t time = clock();
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            forward_convolutional_layer_gpu(layer, input);
@@ -43,33 +48,34 @@
            cost_layer layer = *(cost_layer *)net.layers[i];
            forward_cost_layer_gpu(layer, input, truth);
        }
        /*
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            forward_connected_layer(layer, input, train);
            input = layer.output;
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            forward_softmax_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            forward_crop_layer(layer, input);
            input = layer.output;
            forward_connected_layer_gpu(layer, input);
            input = layer.output_cl;
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            forward_maxpool_layer(layer, input);
            input = layer.output;
            forward_maxpool_layer_gpu(layer, input);
            input = layer.output_cl;
        }
        else if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
            forward_normalization_layer(layer, input);
            input = layer.output;
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            forward_softmax_layer_gpu(layer, input);
            input = layer.output_cl;
        }
        */
        printf("%d %f\n", i, sec(clock()-time));
        /*
           else if(net.types[i] == CROP){
           crop_layer layer = *(crop_layer *)net.layers[i];
           forward_crop_layer(layer, input);
           input = layer.output;
           }
           else if(net.types[i] == NORMALIZATION){
           normalization_layer layer = *(normalization_layer *)net.layers[i];
           forward_normalization_layer(layer, input);
           input = layer.output;
           }
         */
    }
}

@@ -79,6 +85,7 @@
    cl_mem prev_input;
    cl_mem prev_delta;
    for(i = net.n-1; i >= 0; --i){
        clock_t time = clock();
        if(i == 0){
            prev_input = input;
            prev_delta = 0;
@@ -94,6 +101,19 @@
            cost_layer layer = *(cost_layer *)net.layers[i];
            backward_cost_layer_gpu(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            backward_connected_layer_gpu(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            backward_maxpool_layer_gpu(layer, prev_delta);
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            backward_softmax_layer_gpu(layer, prev_delta);
        }
        printf("back: %d %f\n", i, sec(clock()-time));
    }
}

@@ -105,18 +125,9 @@
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            update_convolutional_layer_gpu(layer);
        }
        else if(net.types[i] == MAXPOOL){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        }
        else if(net.types[i] == SOFTMAX){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        }
        else if(net.types[i] == NORMALIZATION){
            //maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            update_connected_layer(layer);
            update_connected_layer_gpu(layer);
        }
    }
}
@@ -127,6 +138,18 @@
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.output_cl;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.output_cl;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.output_cl;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.output_cl;
    }
    return 0;
}

@@ -136,6 +159,18 @@
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.delta_cl;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.delta_cl;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.delta_cl;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta_cl;
    }
    return 0;
}

@@ -229,6 +264,8 @@
        return layer.output;
    } else if(net.types[i] == DROPOUT){
        return get_network_output_layer(net, i-1);
    } else if(net.types[i] == FREEWEIGHT){
        return get_network_output_layer(net, i-1);
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.output;
@@ -258,6 +295,8 @@
        return layer.delta;
    } else if(net.types[i] == DROPOUT){
        return get_network_delta_layer(net, i-1);
    } else if(net.types[i] == FREEWEIGHT){
        return get_network_delta_layer(net, i-1);
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.delta;
@@ -322,7 +361,7 @@
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            if(i != 0) backward_maxpool_layer(layer, prev_input, prev_delta);
            if(i != 0) backward_maxpool_layer(layer, prev_delta);
        }
        else if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
@@ -330,7 +369,7 @@
        }
        else if(net.types[i] == SOFTMAX){
            softmax_layer layer = *(softmax_layer *)net.layers[i];
            if(i != 0) backward_softmax_layer(layer, prev_input, prev_delta);
            if(i != 0) backward_softmax_layer(layer, prev_delta);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
@@ -343,6 +382,73 @@
    }
}


#ifdef GPU
float train_network_datum_gpu(network net, float *x, float *y)
{
    int x_size = get_network_input_size(net)*net.batch;
    int y_size = get_network_output_size(net)*net.batch;
    clock_t time = clock();
    if(!*net.input_cl){
        *net.input_cl = cl_make_array(x, x_size);
        *net.truth_cl = cl_make_array(y, y_size);
    }else{
        cl_write_array(*net.input_cl, x, x_size);
        cl_write_array(*net.truth_cl, y, y_size);
    }
    //printf("trans %f\n", sec(clock()-time));
    time = clock();
    forward_network_gpu(net, *net.input_cl, *net.truth_cl, 1);
    //printf("forw %f\n", sec(clock()-time));
    time = clock();
    backward_network_gpu(net, *net.input_cl);
    //printf("back %f\n", sec(clock()-time));
    time = clock();
    float error = get_network_cost(net);
    update_network_gpu(net);
    //printf("updt %f\n", sec(clock()-time));
    time = clock();
    return error;
}

float train_network_sgd_gpu(network net, data d, int n)
{
    int batch = net.batch;
    float *X = calloc(batch*d.X.cols, sizeof(float));
    float *y = calloc(batch*d.y.cols, sizeof(float));

    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_random_batch(d, batch, X, y);
        float err = train_network_datum_gpu(net, X, y);
        sum += err;
    }
    free(X);
    free(y);
    return (float)sum/(n*batch);
}

float train_network_data_gpu(network net, data d, int n)
{
    int batch = net.batch;
    float *X = calloc(batch*d.X.cols, sizeof(float));
    float *y = calloc(batch*d.y.cols, sizeof(float));

    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_next_batch(d, batch, i*batch, X, y);
        float err = train_network_datum_gpu(net, X, y);
        sum += err;
    }
    free(X);
    free(y);
    return (float)sum/(n*batch);
}
#endif


float train_network_datum(network net, float *x, float *y)
{
    forward_network(net, x, y, 1);
@@ -363,7 +469,7 @@
    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        get_batch(d, batch, X, y);
        get_random_batch(d, batch, X, y);
        float err = train_network_datum(net, X, y);
        sum += err;
    }
@@ -424,6 +530,10 @@
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == FREEWEIGHT){
        freeweight_layer layer = *(freeweight_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
@@ -451,6 +561,10 @@
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == FREEWEIGHT){
        freeweight_layer layer = *(freeweight_layer *) net.layers[i];
        return layer.inputs;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.inputs;
@@ -538,7 +652,7 @@
    image *prev = 0;
    int i;
    char buff[256];
    show_image(get_network_image_layer(net, 0), "Crop");
    //show_image(get_network_image_layer(net, 0), "Crop");
    for(i = 0; i < net.n; ++i){
        sprintf(buff, "Layer %d", i);
        if(net.types[i] == CONVOLUTIONAL){
@@ -552,6 +666,27 @@
    } 
}

void top_predictions(network net, int n, int *index)
{
    int i,j;
    int k = get_network_output_size(net);
    float *out = get_network_output(net);
    float thresh = FLT_MAX;
    for(i = 0; i < n; ++i){
        float max = -FLT_MAX;
        int max_i = -1;
        for(j = 0; j < k; ++j){
            float val = out[j];
            if(val > max &&  val < thresh){
                max = val;
                max_i = j;
            }
        }
        index[i] = max_i;
        thresh = max;
    }
}

float *network_predict(network net, float *input)
{
    forward_network(net, input, 0, 0);