Data augmentations - when and how are they applied?

[bastlen]

I am curious about when and how data augmentation is applied? I´m not sure if I got it right when using maskrcnn model:

• data is loaded with dataloader and datamapper containing augmentations
• in plain_train_net.py is the train loop where one iter is one fully training data pass
  --> my question: Is data augmentation applied once when raw data is loaded (after loading the whole training is based on the first data augmentation) or is data augmentation applied for every iter