Readme.md - When should I stop training

Alexey

2017-03-01 a360f694093f7748e3232f4ed74d40446d735fc3

 src/convolutional_kernels.cu

@@ -2,6 +2,10 @@
#include "curand.h"
#include "cublas_v2.h"

#ifdef CUDNN
#pragma comment(lib, "cudnn.lib")  
#endif

extern "C" {
#include "convolutional_layer.h"
#include "batchnorm_layer.h"
@@ -133,6 +137,9 @@

    if(l.batch_normalize){
        backward_batchnorm_layer_gpu(l, state);
        //axpy_ongpu(l.outputs*l.batch, -state.net.decay, l.x_gpu, 1, l.delta_gpu, 1);
    } else {
        //axpy_ongpu(l.outputs*l.batch, -state.net.decay, l.output_gpu, 1, l.delta_gpu, 1);
    }
    float *original_input = state.input;

@@ -215,6 +222,10 @@
        cuda_pull_array(layer.rolling_mean_gpu, layer.rolling_mean, layer.n);
        cuda_pull_array(layer.rolling_variance_gpu, layer.rolling_variance, layer.n);
    }
    if (layer.adam){
        cuda_pull_array(layer.m_gpu, layer.m, layer.c*layer.n*layer.size*layer.size);
        cuda_pull_array(layer.v_gpu, layer.v, layer.c*layer.n*layer.size*layer.size);
    }
}

void push_convolutional_layer(convolutional_layer layer)
@@ -228,6 +239,10 @@
        cuda_push_array(layer.rolling_mean_gpu, layer.rolling_mean, layer.n);
        cuda_push_array(layer.rolling_variance_gpu, layer.rolling_variance, layer.n);
    }
    if (layer.adam){
        cuda_push_array(layer.m_gpu, layer.m, layer.c*layer.n*layer.size*layer.size);
        cuda_push_array(layer.v_gpu, layer.v, layer.c*layer.n*layer.size*layer.size);
    }
}

void update_convolutional_layer_gpu(convolutional_layer layer, int batch, float learning_rate, float momentum, float decay)