Make GridPatchDataset cacheable by adding shapes parameter

[schellenchris]

I noticed computing the slices each epoch isn't really fast so I made the GridPatchDataset cacheable by adding a shapes parameter.
I compute the shapes shapes = [nib.load(img).shape for img in imgs] in advance, it's pretty fast. I hand the shapes over to the GridPatchDataset to compute the length of the dataset. For __getitem__ I store the sliced image so I just need to slice it once. For multi dimension slicing and overlaps it could get more complex but I think it's achievable. Handing the GridPatchDataset into CacheDataset managed to speed up my training.

[wyli]

do you mean GridPatchDataset to inherit Dataset in addition to IterableDataset? would be great to have a small example to briefly show the idea...

[schellenchris]

inheriting Dataset would be the clean solution I think. I just implemented __len__ and __getitem__ functions to make it work the quick way. I linked you my GridPatchDataset
https://github.com/schellenchris/MONAIFiles/blob/73f98e197be4ad318b3cb150d6c9270fc504e269/grid_dataset.py#L205-L236