renamed sigmoid to logistic

Joseph Redmon

2015-03-08 f047cfff99e00e28c02eb59b6d32386c122f9af6

 src/network_kernels.cu

@@ -9,7 +9,9 @@

#include "crop_layer.h"
#include "connected_layer.h"
#include "detection_layer.h"
#include "convolutional_layer.h"
#include "deconvolutional_layer.h"
#include "maxpool_layer.h"
#include "cost_layer.h"
#include "normalization_layer.h"
@@ -20,6 +22,7 @@

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);

void forward_network_gpu(network net, float * input, float * truth, int train)
{
@@ -31,6 +34,11 @@
            forward_convolutional_layer_gpu(layer, input);
            input = layer.output_gpu;
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
            forward_deconvolutional_layer_gpu(layer, input);
            input = layer.output_gpu;
        }
        else if(net.types[i] == COST){
            cost_layer layer = *(cost_layer *)net.layers[i];
            forward_cost_layer_gpu(layer, input, truth);
@@ -40,6 +48,11 @@
            forward_connected_layer_gpu(layer, input);
            input = layer.output_gpu;
        }
        else if(net.types[i] == DETECTION){
            detection_layer layer = *(detection_layer *)net.layers[i];
            forward_detection_layer_gpu(layer, input, truth);
            input = layer.output_gpu;
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            forward_maxpool_layer_gpu(layer, input);
@@ -58,14 +71,15 @@
        }
        else if(net.types[i] == CROP){
            crop_layer layer = *(crop_layer *)net.layers[i];
            forward_crop_layer_gpu(layer, input);
            forward_crop_layer_gpu(layer, train, input);
            input = layer.output_gpu;
        }
        //cudaDeviceSynchronize();
        //printf("Forward %d %s %f\n", i, get_layer_string(net.types[i]), sec(clock() - time));
    }
}

void backward_network_gpu(network net, float * input)
void backward_network_gpu(network net, float * input, float *truth)
{
    int i;
    float * prev_input;
@@ -83,6 +97,10 @@
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            backward_convolutional_layer_gpu(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
            backward_deconvolutional_layer_gpu(layer, prev_input, 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);
@@ -91,6 +109,10 @@
            connected_layer layer = *(connected_layer *)net.layers[i];
            backward_connected_layer_gpu(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == DETECTION){
            detection_layer layer = *(detection_layer *)net.layers[i];
            backward_detection_layer_gpu(layer, prev_input, prev_delta);
        }
        else if(net.types[i] == MAXPOOL){
            maxpool_layer layer = *(maxpool_layer *)net.layers[i];
            backward_maxpool_layer_gpu(layer, prev_delta);
@@ -115,6 +137,10 @@
            convolutional_layer layer = *(convolutional_layer *)net.layers[i];
            update_convolutional_layer_gpu(layer);
        }
        else if(net.types[i] == DECONVOLUTIONAL){
            deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
            update_deconvolutional_layer_gpu(layer);
        }
        else if(net.types[i] == CONNECTED){
            connected_layer layer = *(connected_layer *)net.layers[i];
            update_connected_layer_gpu(layer);
@@ -128,6 +154,14 @@
        convolutional_layer layer = *(convolutional_layer *)net.layers[i];
        return layer.output_gpu;
    }
    else if(net.types[i] == DECONVOLUTIONAL){
        deconvolutional_layer layer = *(deconvolutional_layer *)net.layers[i];
        return layer.output_gpu;
    }
    else if(net.types[i] == DETECTION){
        detection_layer layer = *(detection_layer *)net.layers[i];
        return layer.output_gpu;
    }
    else if(net.types[i] == CONNECTED){
        connected_layer layer = *(connected_layer *)net.layers[i];
        return layer.output_gpu;
@@ -156,6 +190,14 @@
        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;
@@ -176,6 +218,7 @@

float train_network_datum_gpu(network net, float *x, float *y)
{
  //clock_t time = clock();
    int x_size = get_network_input_size(net)*net.batch;
    int y_size = get_network_output_size(net)*net.batch;
    if(!*net.input_gpu){
@@ -185,10 +228,22 @@
        cuda_push_array(*net.input_gpu, x, x_size);
        cuda_push_array(*net.truth_gpu, y, y_size);
    }
  //printf("trans %f\n", sec(clock() - time));
  //time = clock();
    forward_network_gpu(net, *net.input_gpu, *net.truth_gpu, 1);
    backward_network_gpu(net, *net.input_gpu);
  //printf("forw %f\n", sec(clock() - time));
  //time = clock();
    backward_network_gpu(net, *net.input_gpu, *net.truth_gpu);
  //printf("back %f\n", sec(clock() - time));
  //time = clock();
    update_network_gpu(net);
    float error = get_network_cost(net);

    //print_letters(y, 50);
    //float *out = get_network_output_gpu(net);
    //print_letters(out, 50);
  //printf("updt %f\n", sec(clock() - time));
  //time = clock();
    return error;
}

@@ -198,11 +253,21 @@
        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;