void forward_batchnorm_layer_gpu(layer l, network_state state)
{
if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, state.input, 1, l.output_gpu, 1);
if(l.type == CONNECTED){
l.out_c = l.outputs;
l.out_h = l.out_w = 1;
}
if (state.train) {
fast_mean_gpu(l.output_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.mean_gpu);
fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.variance_gpu);
scal_ongpu(l.out_c, .99, l.rolling_mean_gpu, 1);
axpy_ongpu(l.out_c, .01, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
scal_ongpu(l.out_c, .99, l.rolling_variance_gpu, 1);
axpy_ongpu(l.out_c, .01, l.variance_gpu, 1, l.rolling_variance_gpu, 1);
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.out_c, l.out_h*l.out_w);
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
} else {
normalize_gpu(l.output_gpu, l.rolling_mean_gpu, l.rolling_variance_gpu, l.batch, l.out_c, l.out_h*l.out_w);
}
scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
}
void forward_batchnorm_layer_gpu(layer l, network_state state)
{
if(l.type == BATCHNORM) copy_ongpu(l.outputs*l.batch, state.input, 1, l.output_gpu, 1);
if(l.type == CONNECTED){
l.out_c = l.outputs;
l.out_h = l.out_w = 1;
}
if (state.train) {
#ifdef CUDNN
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_gpu, 1);
float one = 1;
float zero = 0;
cudnnBatchNormalizationForwardTraining(cudnn_handle(),
CUDNN_BATCHNORM_SPATIAL,
&one,
&zero,
l.dstTensorDesc,
l.x_gpu,
l.dstTensorDesc,
l.output_gpu,
l.normTensorDesc,
l.scales_gpu,
l.biases_gpu,
.01,
l.rolling_mean_gpu,
l.rolling_variance_gpu,
.00001,
l.mean_gpu,
l.variance_gpu);
#else
fast_mean_gpu(l.output_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.mean_gpu);
fast_variance_gpu(l.output_gpu, l.mean_gpu, l.batch, l.out_c, l.out_h*l.out_w, l.variance_gpu);
scal_ongpu(l.out_c, .99, l.rolling_mean_gpu, 1);
axpy_ongpu(l.out_c, .01, l.mean_gpu, 1, l.rolling_mean_gpu, 1);
scal_ongpu(l.out_c, .99, l.rolling_variance_gpu, 1);
axpy_ongpu(l.out_c, .01, l.variance_gpu, 1, l.rolling_variance_gpu, 1);
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.out_c, l.out_h*l.out_w);
copy_ongpu(l.outputs*l.batch, l.output_gpu, 1, l.x_norm_gpu, 1);
scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.out_c, l.out_w*l.out_h);
#endif
} else {
normalize_gpu(l.output_gpu, l.rolling_mean_gpu, l.rolling_variance_gpu, l.batch, l.out_c, l.out_h*l.out_w);
scale_bias_gpu(l.output_gpu, l.scales_gpu, l.batch, l.out_c, l.out_h*l.out_w);
add_bias_gpu(l.output_gpu, l.biases_gpu, l.batch, l.out_c, l.out_w*l.out_h);
}
}