I would like to train a dataset for images on the fly adversarial. That means I would have to apply data augmentation on the fly and additionally I could integrate different attacks and sampling methods in the training process.
Unfortunately I don't know yet how to start with the basic on the fly data augmentation. I also have the question, when should the data set be split into training, test and validation data? I think this makes sense before the augmentation or? At what point can the integration of e.g. deepfool attack take place and at what point does it still make sense to use sampling?
Which sampling methods are useful for On the fly basically possible and make sense?
Thanks in advance. :-)
So far I done the adversarial training "offline" and generated and saved the adversarials beforehand. I also implemented and tested different sampling methods in the training process, e.g. random sampling. However, none of them happened on the fly