what do you even write here?

Joseph Redmon

2016-08-11 aebe937710ced03d03f73ab23f410f29685655c1

 src/network_kernels.cu

@@ -1,99 +1,160 @@
#include "cuda_runtime.h"
#include "curand.h"
#include "cublas_v2.h"

extern "C" {
#include <stdio.h>
#include <time.h>
#include <assert.h>

#include "network.h"
#include "image.h"
#include "data.h"
#include "utils.h"
#include "params.h"
#include "parser.h"

#include "crop_layer.h"
#include "connected_layer.h"
#include "rnn_layer.h"
#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 "cost_layer.h"
#include "reorg_layer.h"
#include "avgpool_layer.h"
#include "normalization_layer.h"
#include "batchnorm_layer.h"
#include "cost_layer.h"
#include "local_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "route_layer.h"
#include "shortcut_layer.h"
#include "blas.h"
}

extern "C" float * get_network_output_gpu_layer(network net, int i);
extern "C" float * get_network_delta_gpu_layer(network net, int i);
float *get_network_output_gpu(network net);
float * get_network_output_gpu_layer(network net, int i);
float * get_network_delta_gpu_layer(network net, int i);
float * get_network_output_gpu(network net);

void forward_network_gpu(network net, network_state state)
{
    state.workspace = net.workspace;
    int i;
    for(i = 0; i < net.n; ++i){
        if(net.types[i] == CONVOLUTIONAL){
            forward_convolutional_layer_gpu(*(convolutional_layer *)net.layers[i], state);
        state.index = i;
        layer l = net.layers[i];
        if(l.delta_gpu){
            fill_ongpu(l.outputs * l.batch, 0, l.delta_gpu, 1);
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            forward_deconvolutional_layer_gpu(*(deconvolutional_layer *)net.layers[i], state);
        if(l.type == CONVOLUTIONAL){
            forward_convolutional_layer_gpu(l, state);
        } else if(l.type == DECONVOLUTIONAL){
            forward_deconvolutional_layer_gpu(l, state);
        } else if(l.type == ACTIVE){
            forward_activation_layer_gpu(l, state);
        } else if(l.type == LOCAL){
            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){
            forward_rnn_layer_gpu(l, state);
        } else if(l.type == GRU){
            forward_gru_layer_gpu(l, state);
        } else if(l.type == CRNN){
            forward_crnn_layer_gpu(l, state);
        } else if(l.type == CROP){
            forward_crop_layer_gpu(l, state);
        } else if(l.type == COST){
            forward_cost_layer_gpu(l, state);
        } else if(l.type == SOFTMAX){
            forward_softmax_layer_gpu(l, state);
        } else if(l.type == NORMALIZATION){
            forward_normalization_layer_gpu(l, state);
        } else if(l.type == BATCHNORM){
            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){
            forward_dropout_layer_gpu(l, state);
        } else if(l.type == ROUTE){
            forward_route_layer_gpu(l, net);
        } else if(l.type == SHORTCUT){
            forward_shortcut_layer_gpu(l, state);
        }
        else if(net.types[i] == COST){
            forward_cost_layer_gpu(*(cost_layer *)net.layers[i], state);
        }
        else if(net.types[i] == CONNECTED){
            forward_connected_layer_gpu(*(connected_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DETECTION){
            forward_detection_layer_gpu(*(detection_layer *)net.layers[i], state);
        }
        else if(net.types[i] == MAXPOOL){
            forward_maxpool_layer_gpu(*(maxpool_layer *)net.layers[i], state);
        }
        else if(net.types[i] == SOFTMAX){
            forward_softmax_layer_gpu(*(softmax_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DROPOUT){
            forward_dropout_layer_gpu(*(dropout_layer *)net.layers[i], state);
        }
        else if(net.types[i] == CROP){
            forward_crop_layer_gpu(*(crop_layer *)net.layers[i], state);
        }
        state.input = get_network_output_gpu_layer(net, i);
        state.input = l.output_gpu;
    }
}

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;
    for(i = net.n-1; i >= 0; --i){
        state.index = i;
        layer l = net.layers[i];
        if(i == 0){
            state.input = original_input;
            state.delta = 0;
            state.delta = original_delta;
        }else{
            state.input = get_network_output_gpu_layer(net, i-1);
            state.delta = get_network_delta_gpu_layer(net, i-1);
            layer prev = net.layers[i-1];
            state.input = prev.output_gpu;
            state.delta = prev.delta_gpu;
        }
        if(net.types[i] == CONVOLUTIONAL){
            backward_convolutional_layer_gpu(*(convolutional_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            backward_deconvolutional_layer_gpu(*(deconvolutional_layer *)net.layers[i], state);
        }
        else if(net.types[i] == COST){
            backward_cost_layer_gpu(*(cost_layer *)net.layers[i], state);
        }
        else if(net.types[i] == CONNECTED){
            backward_connected_layer_gpu(*(connected_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DETECTION){
            backward_detection_layer_gpu(*(detection_layer *)net.layers[i], state);
        }
        else if(net.types[i] == MAXPOOL){
            backward_maxpool_layer_gpu(*(maxpool_layer *)net.layers[i], state);
        }
        else if(net.types[i] == DROPOUT){
            backward_dropout_layer_gpu(*(dropout_layer *)net.layers[i], state);
        }
        else if(net.types[i] == SOFTMAX){
            backward_softmax_layer_gpu(*(softmax_layer *)net.layers[i], state);
        if(l.type == CONVOLUTIONAL){
            backward_convolutional_layer_gpu(l, state);
        } else if(l.type == DECONVOLUTIONAL){
            backward_deconvolutional_layer_gpu(l, state);
        } else if(l.type == ACTIVE){
            backward_activation_layer_gpu(l, state);
        } else if(l.type == LOCAL){
            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){
            backward_batchnorm_layer_gpu(l, state);
        } else if(l.type == SOFTMAX){
            if(i != 0) backward_softmax_layer_gpu(l, state);
        } else if(l.type == CONNECTED){
            backward_connected_layer_gpu(l, state);
        } else if(l.type == RNN){
            backward_rnn_layer_gpu(l, state);
        } else if(l.type == GRU){
            backward_gru_layer_gpu(l, state);
        } else if(l.type == CRNN){
            backward_crnn_layer_gpu(l, state);
        } else if(l.type == COST){
            backward_cost_layer_gpu(l, state);
        } else if(l.type == ROUTE){
            backward_route_layer_gpu(l, net);
        } else if(l.type == SHORTCUT){
            backward_shortcut_layer_gpu(l, state);
        }
    }
}
@@ -102,88 +163,35 @@
{
    int i;
    int update_batch = net.batch*net.subdivisions;
    float rate = get_current_rate(net);
    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, update_batch, net.learning_rate, net.momentum, net.decay);
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
            update_deconvolutional_layer_gpu(layer, net.learning_rate, net.momentum, net.decay);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            update_connected_layer_gpu(layer, update_batch, net.learning_rate, net.momentum, net.decay);
        layer l = net.layers[i];
        if(l.type == CONVOLUTIONAL){
            update_convolutional_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == DECONVOLUTIONAL){
            update_deconvolutional_layer_gpu(l, rate, net.momentum, net.decay);
        } else if(l.type == CONNECTED){
            update_connected_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == GRU){
            update_gru_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == RNN){
            update_rnn_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == CRNN){
            update_crnn_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
        } else if(l.type == LOCAL){
            update_local_layer_gpu(l, update_batch, rate, net.momentum, net.decay);
        }
    }
}

float * get_network_output_gpu_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        return ((convolutional_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == DECONVOLUTIONAL){
        return ((deconvolutional_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == DETECTION){
        return ((detection_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == CONNECTED){
        return ((connected_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == MAXPOOL){
        return ((maxpool_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == CROP){
        return ((crop_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == SOFTMAX){
        return ((softmax_layer *)net.layers[i]) -> output_gpu;
    }
    else if(net.types[i] == DROPOUT){
        return get_network_output_gpu_layer(net, i-1);
    }
    return 0;
}

float * get_network_delta_gpu_layer(network net, int i)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.delta_gpu;
    }
    else if(net.types[i] == DETECTION){
        detection_layer layer = *(detection_layer *)net.layers[i];
        return layer.delta_gpu;
    }
    else if(net.types[i] == DECONVOLUTIONAL){
        deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
        return layer.delta_gpu;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.delta_gpu;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.delta_gpu;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        return layer.delta_gpu;
    } else if(net.types[i] == DROPOUT){
        if(i == 0) return 0;
        return get_network_delta_gpu_layer(net, i-1);
    }
    return 0;
}

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].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);
@@ -192,53 +200,80 @@
        cuda_push_array(*net.truth_gpu, y, y_size);
    }
    state.input = *net.input_gpu;
    state.delta = 0;
    state.truth = *net.truth_gpu;
    state.train = 1;
    forward_network_gpu(net, state);
    backward_network_gpu(net, state);
    if ((net.seen / net.batch) % net.subdivisions == 0) update_network_gpu(net);
}

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)
{
    if(net.types[i] == CONVOLUTIONAL){
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.output;
    }
    else if(net.types[i] == DECONVOLUTIONAL){
        deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
        return layer.output;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        cuda_pull_array(layer.output_gpu, layer.output, layer.outputs*layer.batch);
        return layer.output;
    }
    else if(net.types[i] == DETECTION){
        detection_layer layer = *(detection_layer *)net.layers[i];
        int outputs = get_detection_layer_output_size(layer);
        cuda_pull_array(layer.output_gpu, layer.output, outputs*layer.batch);
        return layer.output;
    }
    else if(net.types[i] == MAXPOOL){
        maxpool_layer layer = *(maxpool_layer *)net.layers[i];
        return layer.output;
    }
    else if(net.types[i] == SOFTMAX){
        softmax_layer layer = *(softmax_layer *)net.layers[i];
        pull_softmax_layer_output(layer);
        return layer.output;
    }
    return 0;
    layer l = net.layers[i];
    cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
    return l.output;
}

float *get_network_output_gpu(network net)
{
    int i;
    for(i = net.n-1; i > 0; --i) if(net.types[i] != COST) break;
    for(i = net.n-1; i > 0; --i) if(net.layers[i].type != COST) break;
    return get_network_output_layer_gpu(net, i);
}

@@ -246,6 +281,8 @@
{
    int size = get_network_input_size(net) * net.batch;
    network_state state;
    state.index = 0;
    state.net = net;
    state.input = cuda_make_array(input, size);
    state.truth = 0;
    state.train = 0;