Data augmentation - ever 'added' to original data set?

[mayadiCS]

Thanks for the really excellent content in section 3 Computer vision. Question: I noticed that when using ImageDataGenerator to augment the food 101 (subset) input data, we only end up using the augmented data when training the model. Is that fairly common practice in the industry? (I am relatively new). Meaning, do we never for example use both the augmented + original data (i.e. "double" our data) as the input feed during model training? Many thanks

PS: and just to clarify - I do not mean modify the original input data folder. Agnostic of how the data is consuming could we supplement with the augmented data rather than use it instead of the non augmented data.

[mrdbourke]

Hey @mayadiCS ,

Data augmentation is used to artificially increase the diversity of a training set (different views of similar images).

The usual practice is to use only augmented images or not augmented images.

You can read more on this here: https://machinelearningmastery.com/how-to-configure-image-data-augmentation-when-training-deep-learning-neural-networks/