Is your feature request related to a problem? Please describe.
Patch-based dataset: set random sampling mode, and set the number of patches from each image (volume) (will be extracted when iterating dataloader).
Could we have this to train with large images and large datasets? Thank you.