Optimize pytorch data loader for reading small patches in full HD images

Question

I\'m training my neural network using PyTorch framework. The data is full HD images (1920x1080). But in each iteration, I just need to crop out a random 256x256 patch from t