Faster batch normalization

Joseph Redmon

2015-12-08 c2738835f0a2435ab03f411af3d168aec389d2a6

 src/convolutional_kernels.cu

@@ -119,17 +119,14 @@

    if(l.batch_normalize){
        if(state.train){
            fast_mean_gpu(l.output_gpu, l.batch, l.n, l.out_h*l.out_w, l.spatial_mean_gpu, l.mean_gpu);   
            fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.n, l.out_h*l.out_w, l.spatial_variance_gpu, l.variance_gpu);
            fast_mean_gpu(l.output_gpu, l.batch, l.n, l.out_h*l.out_w, l.mean_gpu);
            fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.n, l.out_h*l.out_w, l.variance_gpu);

            scal_ongpu(l.n, .95, l.rolling_mean_gpu, 1);
            axpy_ongpu(l.n, .05, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
            scal_ongpu(l.n, .95, l.rolling_variance_gpu, 1);
            axpy_ongpu(l.n, .05, l.variance_gpu, 1, l.rolling_variance_gpu, 1);

            // cuda_pull_array(l.variance_gpu, l.mean, l.n);
            // printf("%f\n", l.mean[0]);

            copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
            normalize_gpu(l.output_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_h*l.out_w);
            copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
@@ -161,8 +158,8 @@

        scale_bias_gpu(l.delta_gpu, l.scales_gpu, l.batch, l.n, l.out_h*l.out_w);

        fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.spatial_mean_delta_gpu, l.mean_delta_gpu);
        fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.spatial_variance_delta_gpu, l.variance_delta_gpu);
        fast_mean_delta_gpu(l.delta_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.mean_delta_gpu);
        fast_variance_delta_gpu(l.x_gpu, l.delta_gpu, l.mean_gpu, l.variance_gpu, l.batch, l.n, l.out_w*l.out_h, l.variance_delta_gpu);
        normalize_delta_gpu(l.x_gpu, l.mean_gpu, l.variance_gpu, l.mean_delta_gpu, l.variance_delta_gpu, l.batch, l.n, l.out_w*l.out_h, l.delta_gpu);
    }