See comment [here](https://github.com/facebookresearch/fastMRI/issues/343?fbclid=IwZXh0bgNhZW0CMTAAAR1R6TFzlfldCnI0pZs7WANlNRVgSv9bmIh50BQ1e_CteMGOXp9FjCO-amo_aem_XJama4Um6ddw6ew_17KxAg#issuecomment-2419101865). We should update this so that newer versions of PyTorch do not complain. Could also reimplement the inference to use a model zoo.