I used 2 Gpus for training, but halved the number of channels instead of the batch size. how should I modify it?