Faster batch normalization

Joseph Redmon

2015-12-08 c2738835f0a2435ab03f411af3d168aec389d2a6

 src/connected_layer.c

@@ -103,7 +103,7 @@
    b = l.weights;
    c = state.delta;

    if(c) gemm(0,1,m,n,k,1,a,k,b,k,0,c,n);
    if(c) gemm(0,1,m,n,k,1,a,k,b,k,1,c,n);
}

#ifdef GPU
@@ -148,6 +148,12 @@
    float * c = l.output_gpu;
    gemm_ongpu(0,0,m,n,k,1,a,k,b,n,1,c,n);
    activate_array_ongpu(l.output_gpu, l.outputs*l.batch, l.activation);

/*
    cuda_pull_array(l.output_gpu, l.output, l.outputs*l.batch);
    float avg = mean_array(l.output, l.outputs*l.batch);
    printf("%f\n", avg);
    */
}

void backward_connected_layer_gpu(connected_layer l, network_state state)
@@ -173,6 +179,6 @@
    b = l.weights_gpu;
    c = state.delta;

    if(c) gemm_ongpu(0,1,m,n,k,1,a,k,b,k,0,c,n);
    if(c) gemm_ongpu(0,1,m,n,k,1,a,k,b,k,1,c,n);
}
#endif