I’m using some data augmentation on my training set during the training. After the training ends, I’d like to measure my train accuracy using the original, non-augmented training set.
To do so, do I have to initialize dataset again without augmentation, or is there a better way to do it?
Re-creating a new dataset without any transformations sounds like a proper approach assuming you are lazily loading the samples (which would then be cheap).
Otherwise, you could also try to set the internal .transform attribute to False assuming self.transform is used to apply the transformations on the samples (check the Dataset source to make sure that’s indeed the case).