some more stuff

Joseph Redmon

2016-08-05 845ab7579685b6702c92c1088ec11e71bde51f3c

 src/network_kernels.cu

@@ -19,10 +19,12 @@
#include "gru_layer.h"
#include "crnn_layer.h"
#include "detection_layer.h"
#include "region_layer.h"
#include "convolutional_layer.h"
#include "activation_layer.h"
#include "deconvolutional_layer.h"
#include "maxpool_layer.h"
#include "reorg_layer.h"
#include "avgpool_layer.h"
#include "normalization_layer.h"
#include "batchnorm_layer.h"
@@ -41,6 +43,7 @@

void forward_network_gpu(network net, network_state state)
{
    state.workspace = net.workspace;
    int i;
    for(i = 0; i < net.n; ++i){
        state.index = i;
@@ -58,6 +61,8 @@
            forward_local_layer_gpu(l, state);
        } else if(l.type == DETECTION){
            forward_detection_layer_gpu(l, state);
        } else if(l.type == REGION){
            forward_region_layer_gpu(l, state);
        } else if(l.type == CONNECTED){
            forward_connected_layer_gpu(l, state);
        } else if(l.type == RNN){
@@ -78,6 +83,8 @@
            forward_batchnorm_layer_gpu(l, state);
        } else if(l.type == MAXPOOL){
            forward_maxpool_layer_gpu(l, state);
        } else if(l.type == REORG){
            forward_reorg_layer_gpu(l, state);
        } else if(l.type == AVGPOOL){
            forward_avgpool_layer_gpu(l, state);
        } else if(l.type == DROPOUT){
@@ -93,6 +100,7 @@

void backward_network_gpu(network net, network_state state)
{
    state.workspace = net.workspace;
    int i;
    float * original_input = state.input;
    float * original_delta = state.delta;
@@ -117,12 +125,16 @@
            backward_local_layer_gpu(l, state);
        } else if(l.type == MAXPOOL){
            if(i != 0) backward_maxpool_layer_gpu(l, state);
        } else if(l.type == REORG){
            backward_reorg_layer_gpu(l, state);
        } else if(l.type == AVGPOOL){
            if(i != 0) backward_avgpool_layer_gpu(l, state);
        } else if(l.type == DROPOUT){
            backward_dropout_layer_gpu(l, state);
        } else if(l.type == DETECTION){
            backward_detection_layer_gpu(l, state);
        } else if(l.type == REGION){
            backward_region_layer_gpu(l, state);
        } else if(l.type == NORMALIZATION){
            backward_normalization_layer_gpu(l, state);
        } else if(l.type == BATCHNORM){
@@ -172,14 +184,14 @@
    }
}

float train_network_datum_gpu(network net, float *x, float *y)
void forward_backward_network_gpu(network net, float *x, float *y)
{
    network_state state;
    state.index = 0;
    state.net = net;
    int x_size = get_network_input_size(net)*net.batch;
    int y_size = get_network_output_size(net)*net.batch;
    if(net.layers[net.n-1].type == DETECTION) y_size = net.layers[net.n-1].truths*net.batch;
    if(net.layers[net.n-1].truths) y_size = net.layers[net.n-1].truths*net.batch;
    if(!*net.input_gpu){
        *net.input_gpu = cuda_make_array(x, x_size);
        *net.truth_gpu = cuda_make_array(y, y_size);
@@ -193,12 +205,64 @@
    state.train = 1;
    forward_network_gpu(net, state);
    backward_network_gpu(net, state);
}

float train_network_datum_gpu(network net, float *x, float *y)
{
    forward_backward_network_gpu(net, x, y);
    float error = get_network_cost(net);
    if (((*net.seen) / net.batch) % net.subdivisions == 0) update_network_gpu(net);

    return error;
}

typedef struct {
    network net;
    float *X;
    float *y;
} train_args;

void *train_thread(void *ptr)
{
    train_args args = *(train_args*)ptr;

    cudaError_t status = cudaSetDevice(args.net.gpu_index);
    check_error(status);
    forward_backward_network_gpu(args.net, args.X, args.y);
    free(ptr);
    return 0;
}

pthread_t train_network_in_thread(train_args args)
{
    pthread_t thread;
    train_args *ptr = (train_args *)calloc(1, sizeof(train_args));
    *ptr = args;
    if(pthread_create(&thread, 0, train_thread, ptr)) error("Thread creation failed");
    return thread;
}

float train_networks(network *nets, int n, data d)
{
    int batch = nets[0].batch;
    float **X = (float **) calloc(n, sizeof(float *));
    float **y = (float **) calloc(n, sizeof(float *));
    pthread_t *threads = (pthread_t *) calloc(n, sizeof(pthread_t));

    int i;
    float sum = 0;
    for(i = 0; i < n; ++i){
        X[i] = (float *) calloc(batch*d.X.cols, sizeof(float));
        y[i] = (float *) calloc(batch*d.y.cols, sizeof(float));
        get_next_batch(d, batch, i*batch, X[i], y[i]);
        float err = train_network_datum(nets[i], X[i], y[i]);
        sum += err;
    }
    free(X);
    free(y);
    return (float)sum/(n*batch);
}

float *get_network_output_layer_gpu(network net, int i)
{
    layer l = net.layers[i];