Performance on the validation set does not improve during training. Simple data augmentations do not contribute to the generalization of the model

Current default training setup does not generalize performance on the training set to the validation and test sets. Simple data augmentation tools available in torchvision, i.e., image transformations, do not seem to improve metric performance nor the learning process on the validation set during training.

Set of experiments changing data augmentation tools:

The selected data augmentations were chosen to produce realistic Xray images: