I am trying to train a UNET architecture built with the PyTorch library. I am looking for the best methods to load my data into the model. the dataset has around 500 images (RGB image of around 3000x4000 pixels).

I am planning to extract patches of an image into (256,256,3) shaped patches alone with augmentation while training.

I am looking at methods that do not overload the ram at the same time not compress the data too much to lose the quality of data.