Justus Schock

> The thing with gpu support, is that it will not be available in pytorch DataLoader since they are cpu only . (correct ?) This is only partially correct. The...

> ok, If I understand correctly, you say, one should put num_worker=0, to stay in the main process, ... Either that or use the transforms after the ones from the...

Hi @jizongFox , I’m afraid this won’t be possible/sufficient. Torchio heavily uses nibabel and SimpleITK which would still have a reference to normal numpy. But more importantly both of them...

In my case, It's more like: I have a dataset, process the images in the dataset sequentially and save the final images to disk. Since the dataset is so huge,...

Not necessarily, could be a arbitrary iterable because it also just returns images and not full subjects

> Looking at the extension? I might have misunderstood the question. What I mean: If the API is so that the user passes a directory for loading, for what files...

But how would it know attributes like `shape` and `spacing` without actually loading the image? It may not store the information, but it definitely has to load at least parts...

> You can read metadata from the NIfTI header very quickly without loading the image data. Yes but you cannot do that for DICOM data. So it is kind of...

@Borda I think this is certainly useful. Are there plans to tackle this?

I'll try, but I don't think, this will be that easy, since I could not reproduce it with the exact same tests on my local machine. In our CI/CD however,...