In the training/deepspeech_training/util/feeding.py file, The graph augmentations are applied only on the train set (as indicated by the train_phase flag).
But when it comes to applying sample augmentations, no such train_phase check is present
so augmentations can be applied to dev and test sets as well.
Is there any specific reason as to why only sample augmentations can be applied to dev and test set but not graph augmentations.