Thanks for the great work!
Data augmentation support with albumentations library during training would be very useful. Typically augmentations and their hyperparams are defined in a yaml and applied only to the training set. This yaml path can maybe passed to configure_dataset in your training code.