Can we make inference on the whole image according to the training process? Will there be obvious artifacts between patches?