When I use data augmentation for creating samples for deep learning, how many more samples am I creating when compared to the original dataset?

[ScarletMaguire]

To explain my question more clearly, when I use data augmentation parameters for deep learning samples (e.g. crop, brightness, zoom) is each parameter applied to a different sample each time, or are several parameters applied at the same time? For example, if I have 100 samples and I choose Crop, Brightness and Zoom, does it create 100 extra samples where all three augmentation parameters have been applied at once? Or doe sit create 300 extra samples, where 100 has had crop applied, 100 had brightness modified and 100 had zoom applied?

Thanks is advance for any help!

[PavanYadav]

Hi @ScarletMaguire 

When multiple data augmentations are chosen, a pipeline is created and these transformations are applied in real time while the data is sent to the model for training. We do not create augmentations of the input images and store it anywhere. If you have 100 samples, the final number of samples will still be 100, but some or all the data augmentations are applied on subset of those 100. The samples on which augmentations should apply, and the augmentations itself are randomly chosen and these will change every epoch. This way model will learn from different augmentations applied on different input images.

When you choose multiple or single augmentations on 100 sample input images, the total number of input images will still be 100 and the total samples wont increase to 200 or to 400.