CUDA so fast

Joseph Redmon

2015-01-23 809f924db2823b9e1eaf3efb9370380edc1f76ed

 src/network.c

@@ -15,6 +15,35 @@
#include "softmax_layer.h"
#include "dropout_layer.h"

char *get_layer_string(LAYER_TYPE a)
{
    switch(a){
        case CONVOLUTIONAL:
            return "convolutional";
        case CONNECTED:
            return "connected";
        case MAXPOOL:
            return "maxpool";
        case SOFTMAX:
            return "softmax";
        case NORMALIZATION:
            return "normalization";
        case DROPOUT:
            return "dropout";
        case FREEWEIGHT:
            return "freeweight";
        case CROP:
            return "crop";
        case COST:
            return "cost";
        default:
            break;
    }
    return "none";
}



network make_network(int n, int batch)
{
    network net;
@@ -24,9 +53,10 @@
    net.types = calloc(net.n, sizeof(LAYER_TYPE));
    net.outputs = 0;
    net.output = 0;
    net.seen = 0;
    #ifdef GPU
    net.input_cl = calloc(1, sizeof(cl_mem));
    net.truth_cl = calloc(1, sizeof(cl_mem));
    net.input_gpu = calloc(1, sizeof(float *));
    net.truth_gpu = calloc(1, sizeof(float *));
    #endif
    return net;
}
@@ -74,12 +104,16 @@
            if(!train) continue;
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            forward_dropout_layer(layer, input);
            input = layer.output;
        }
        else if(net.types[i] == FREEWEIGHT){
            if(!train) continue;
            freeweight_layer layer = *(freeweight_layer *)net.layers[i];
            forward_freeweight_layer(layer, input);
            //freeweight_layer layer = *(freeweight_layer *)net.layers[i];
            //forward_freeweight_layer(layer, input);
        }
        //char buff[256];
        //sprintf(buff, "layer %d", i);
        //cuda_compare(get_network_output_gpu_layer(net, i), input, get_network_output_size_layer(net, i)*net.batch, buff);
    }
}

@@ -102,6 +136,7 @@
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            //secret_update_connected_layer((connected_layer *)net.layers[i]);
            update_connected_layer(layer);
        }
    }
@@ -119,7 +154,8 @@
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == DROPOUT){
        return get_network_output_layer(net, i-1);
        dropout_layer layer = *(dropout_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == FREEWEIGHT){
        return get_network_output_layer(net, i-1);
    } else if(net.types[i] == CONNECTED){
@@ -153,6 +189,7 @@
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta;
    } else if(net.types[i] == DROPOUT){
        if(i == 0) return 0;
        return get_network_delta_layer(net, i-1);
    } else if(net.types[i] == FREEWEIGHT){
        return get_network_delta_layer(net, i-1);
@@ -368,6 +405,10 @@
            cost_layer *layer = (cost_layer *)net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == CROP){
            crop_layer *layer = (crop_layer *)net->layers[i];
            layer->batch = b;
        }
    }
}

@@ -545,7 +586,7 @@
float *network_predict(network net, float *input)
{
    #ifdef GPU
        if(gpu_index >= 0) return network_predict_gpu(net, input);
    if(gpu_index >= 0)  return network_predict_gpu(net, input);
    #endif

    forward_network(net, input, 0, 0);
@@ -645,6 +686,31 @@
    }
}

void compare_networks(network n1, network n2, data test)
{
    matrix g1 = network_predict_data(n1, test);
    matrix g2 = network_predict_data(n2, test);
    int i;
    int a,b,c,d;
    a = b = c = d = 0;
    for(i = 0; i < g1.rows; ++i){
        int truth = max_index(test.y.vals[i], test.y.cols);
        int p1 = max_index(g1.vals[i], g1.cols);
        int p2 = max_index(g2.vals[i], g2.cols);
        if(p1 == truth){
            if(p2 == truth) ++d;
            else ++c;
        }else{
            if(p2 == truth) ++b;
            else ++a;
        }
    }
    printf("%5d %5d\n%5d %5d\n", a, b, c, d);
    float num = pow((abs(b - c) - 1.), 2.);
    float den = b + c;
    printf("%f\n", num/den); 
}

float network_accuracy(network net, data d)
{
    matrix guess = network_predict_data(net, d);