Hacks to get nightmare to not break gridsizing

Joseph Redmon

2015-08-11 d0b9326a352ed2fbc3ae66fdef40b4533a2f211d

 src/network_kernels.cu

@@ -7,6 +7,7 @@
#include "data.h"
#include "utils.h"
#include "params.h"
#include "parser.h"

#include "crop_layer.h"
#include "connected_layer.h"
@@ -14,10 +15,13 @@
#include "convolutional_layer.h"
#include "deconvolutional_layer.h"
#include "maxpool_layer.h"
#include "avgpool_layer.h"
#include "normalization_layer.h"
#include "cost_layer.h"
#include "softmax_layer.h"
#include "dropout_layer.h"
#include "route_layer.h"
#include "blas.h"
}

float * get_network_output_gpu_layer(network net, int i);
@@ -29,6 +33,9 @@
    int i;
    for(i = 0; i < net.n; ++i){
        layer l = net.layers[i];
        if(l.delta_gpu){
            scal_ongpu(l.outputs * l.batch, 0, l.delta_gpu, 1);
        }
        if(l.type == CONVOLUTIONAL){
            forward_convolutional_layer_gpu(l, state);
        } else if(l.type == DECONVOLUTIONAL){
@@ -43,8 +50,12 @@
            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 == MAXPOOL){
            forward_maxpool_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){
@@ -58,11 +69,12 @@
{
    int i;
    float * original_input = state.input;
    float * original_delta = state.delta;
    for(i = net.n-1; i >= 0; --i){
        layer l = net.layers[i];
        if(i == 0){
            state.input = original_input;
            state.delta = 0;
            state.delta = original_delta;
        }else{
            layer prev = net.layers[i-1];
            state.input = prev.output_gpu;
@@ -74,10 +86,14 @@
            backward_deconvolutional_layer_gpu(l, state);
        } else if(l.type == MAXPOOL){
            if(i != 0) backward_maxpool_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 == NORMALIZATION){
            backward_normalization_layer_gpu(l, state);
        } else if(l.type == SOFTMAX){
            if(i != 0) backward_softmax_layer_gpu(l, state);
        } else if(l.type == CONNECTED){
@@ -119,6 +135,7 @@
        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);
@@ -132,23 +149,8 @@
float *get_network_output_layer_gpu(network net, int i)
{
    layer l = net.layers[i];
    if(l.type == CONVOLUTIONAL){
        return l.output;
    } else if(l.type == DECONVOLUTIONAL){
        return l.output;
    } else if(l.type == CONNECTED){
        cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
        return l.output;
    } else if(l.type == DETECTION){
        cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
        return l.output;
    } else if(l.type == MAXPOOL){
        return l.output;
    } else if(l.type == SOFTMAX){
        pull_softmax_layer_output(l);
        return l.output;
    }
    return 0;
    cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
    return l.output;
}

float *get_network_output_gpu(network net)