CUDA so fast

Joseph Redmon

2015-01-23 809f924db2823b9e1eaf3efb9370380edc1f76ed

 src/network.c

@@ -1,4 +1,5 @@
#include <stdio.h>
#include <time.h>
#include "network.h"
#include "image.h"
#include "data.h"
@@ -14,6 +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;
@@ -23,123 +53,14 @@
    net.types = calloc(net.n, sizeof(LAYER_TYPE));
    net.outputs = 0;
    net.output = 0;
    net.seen = 0;
    #ifdef GPU
    net.input_cl = 0;
    net.input_gpu = calloc(1, sizeof(float *));
    net.truth_gpu = calloc(1, sizeof(float *));
    #endif
    return net;
}

#ifdef GPU
void forward_network_gpu(network net, cl_mem input, cl_mem truth, int train)
{
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            forward_convolutional_layer_gpu(layer, input);
            input = layer.output_cl;
        }
        else if(net.types[i] == COST){
            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;
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            forward_maxpool_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;
        }
        */
    }
}

void backward_network_gpu(network net, cl_mem input)
{
    int i;
    cl_mem prev_input;
    cl_mem prev_delta;
    for(i = net.n-1; i >= 0; --i){
        if(i == 0){
            prev_input = input;
            prev_delta = 0;
        }else{
            prev_input = get_network_output_cl_layer(net, i-1);
            prev_delta = get_network_delta_cl_layer(net, i-1);
        }
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            backward_convolutional_layer_gpu(layer, prev_delta);
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            backward_cost_layer_gpu(layer, prev_input, prev_delta);
        }
    }
}

void update_network_gpu(network net)
{
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            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);
        }
    }
}

cl_mem get_network_output_cl_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.output_cl;
    }
    return 0;
}

cl_mem get_network_delta_cl_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.delta_cl;
    }
    return 0;
}

#endif

void forward_network(network net, float *input, float *truth, int train)
{
@@ -183,12 +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);
    }
}

@@ -211,6 +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);
        }
    }
@@ -228,10 +154,16 @@
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == DROPOUT){
        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){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *)net.layers[i];
        return layer.output;
    } else if(net.types[i] == NORMALIZATION){
        normalization_layer layer = *(normalization_layer *)net.layers[i];
        return layer.output;
@@ -257,6 +189,9 @@
        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);
    } else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
@@ -318,11 +253,15 @@
        }
        if(net.types[i] == CONVOLUTIONAL){
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            backward_convolutional_layer(layer, prev_delta);
            backward_convolutional_layer(layer, prev_input, prev_delta);
        }
        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] == DROPOUT){
            dropout_layer layer = *(dropout_layer *)net.layers[i];
            backward_dropout_layer(layer, prev_delta);
        }
        else if(net.types[i] == NORMALIZATION){
            normalization_layer layer = *(normalization_layer *)net.layers[i];
@@ -330,7 +269,7 @@
        }
        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];
@@ -345,12 +284,13 @@

float train_network_datum(network net, float *x, float *y)
{
    #ifdef GPU
    if(gpu_index >= 0) return train_network_datum_gpu(net, x, y);
    #endif
    forward_network(net, x, y, 1);
    //int class = get_predicted_class_network(net);
    backward_network(net, x);
    float error = get_network_cost(net);
    update_network(net);
    //return (y[class]?1:0);
    return error;
}

@@ -363,7 +303,7 @@
    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;
    }
@@ -371,6 +311,26 @@
    free(y);
    return (float)sum/(n*batch);
}

float train_network(network net, data d)
{
    int batch = net.batch;
    int n = d.X.rows / 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(net, X, y);
        sum += err;
    }
    free(X);
    free(y);
    return (float)sum/(n*batch);
}

float train_network_batch(network net, data d, int n)
{
    int i,j;
@@ -390,23 +350,69 @@
    return (float)sum/(n*batch);
}


void train_network(network net, data d)
void set_learning_network(network *net, float rate, float momentum, float decay)
{
    int i;
    int correct = 0;
    for(i = 0; i < d.X.rows; ++i){
        correct += train_network_datum(net, d.X.vals[i], d.y.vals[i]);
        if(i%100 == 0){
            visualize_network(net);
            cvWaitKey(10);
    net->learning_rate=rate;
    net->momentum = momentum;
    net->decay = decay;
    for(i = 0; i < net->n; ++i){
        if(net->types[i] == CONVOLUTIONAL){
            convolutional_layer *layer = (convolutional_layer *)net->layers[i];
            layer->learning_rate=rate;
            layer->momentum = momentum;
            layer->decay = decay;
        }
        else if(net->types[i] == CONNECTED){
            connected_layer *layer = (connected_layer *)net->layers[i];
            layer->learning_rate=rate;
            layer->momentum = momentum;
            layer->decay = decay;
        }
    }
    visualize_network(net);
    cvWaitKey(100);
    fprintf(stderr, "Accuracy: %f\n", (float)correct/d.X.rows);
}


void set_batch_network(network *net, int b)
{
    net->batch = b;
    int i;
    for(i = 0; i < net->n; ++i){
        if(net->types[i] == CONVOLUTIONAL){
            convolutional_layer *layer = (convolutional_layer *)net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == MAXPOOL){
            maxpool_layer *layer = (maxpool_layer *)net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == CONNECTED){
            connected_layer *layer = (connected_layer *)net->layers[i];
            layer->batch = b;
        } else if(net->types[i] == DROPOUT){
            dropout_layer *layer = (dropout_layer *) net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == FREEWEIGHT){
            freeweight_layer *layer = (freeweight_layer *) net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == SOFTMAX){
            softmax_layer *layer = (softmax_layer *)net->layers[i];
            layer->batch = b;
        }
        else if(net->types[i] == COST){
            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;
        }
    }
}


int get_network_input_size_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
@@ -423,11 +429,19 @@
    } else if(net.types[i] == DROPOUT){
        dropout_layer layer = *(dropout_layer *) net.layers[i];
        return layer.inputs;
    } else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *) net.layers[i];
        return layer.c*layer.h*layer.w;
    }
    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;
    }
    printf("Can't find input size\n");
    return 0;
}

@@ -443,6 +457,10 @@
        image output = get_maxpool_image(layer);
        return output.h*output.w*output.c;
    }
     else if(net.types[i] == CROP){
        crop_layer layer = *(crop_layer *) net.layers[i];
        return layer.c*layer.crop_height*layer.crop_width;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.outputs;
@@ -451,10 +469,15 @@
        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;
    }
    printf("Can't find output size\n");
    return 0;
}

@@ -538,7 +561,7 @@
    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){
@@ -552,8 +575,20 @@
    } 
}

void top_predictions(network net, int k, int *index)
{
    int size = get_network_output_size(net);
    float *out = get_network_output(net);
    top_k(out, size, k, index);
}


float *network_predict(network net, float *input)
{
    #ifdef GPU
    if(gpu_index >= 0)  return network_predict_gpu(net, input);
    #endif

    forward_network(net, input, 0, 0);
    float *out = get_network_output(net);
    return out;
@@ -589,7 +624,7 @@
    int i,j,b;
    int k = get_network_output_size(net);
    matrix pred = make_matrix(test.X.rows, k);
    float *X = calloc(net.batch*test.X.rows, sizeof(float));
    float *X = calloc(net.batch*test.X.cols, sizeof(float));
    for(i = 0; i < test.X.rows; i += net.batch){
        for(b = 0; b < net.batch; ++b){
            if(i+b == test.X.rows) break;
@@ -651,18 +686,54 @@
    }
}

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);
    float acc = matrix_accuracy(d.y, guess);
    float acc = matrix_topk_accuracy(d.y, guess,1);
    free_matrix(guess);
    return acc;
}

float *network_accuracies(network net, data d)
{
    static float acc[2];
    matrix guess = network_predict_data(net, d);
    acc[0] = matrix_topk_accuracy(d.y, guess,1);
    acc[1] = matrix_topk_accuracy(d.y, guess,5);
    free_matrix(guess);
    return acc;
}


float network_accuracy_multi(network net, data d, int n)
{
    matrix guess = network_predict_data_multi(net, d, n);
    float acc = matrix_accuracy(d.y, guess);
    float acc = matrix_topk_accuracy(d.y, guess,1);
    free_matrix(guess);
    return acc;
}