previously there was a single augmentation pipeline that was applied to train, val, and test data. This meant that validation metrics reported during training were affected by augmentations, and that early stopping was also dependent on augmentations. This PR solves this issue by applying whatever augmentation pipeline is requested by the user through the config file to the train data; the val and test data are only ever resized.
previously there was a single augmentation pipeline that was applied to train, val, and test data. This meant that validation metrics reported during training were affected by augmentations, and that early stopping was also dependent on augmentations. This PR solves this issue by applying whatever augmentation pipeline is requested by the user through the config file to the train data; the val and test data are only ever resized.