[x] Changed the metadata we store on niftis to be serializable (header is as binary block now)
[x] Changed data structure in test datasets to micic the dict structure in the training datasets
[x] Changed data loading to happen as early as possible (in the preprocessor/dataset) which allows the Torch collate and lightning built-in to device calls to work as intended.
[x] Removed to(device) calls in inference (except canvas which needs to be created, this now uses the data.device as ref)
Everything should work seamlessly with cpu/gpu now
Also cleaned the LightningModule and split functionality into more appropriate methods/hooks
Everything should work seamlessly with cpu/gpu now Also cleaned the LightningModule and split functionality into more appropriate methods/hooks