Store the Augmented FFCV Data then Load it?

[realliyifei]

Since some data augmentation techniques in transforms are not supported in FFCV (Ref. the end of #84)

Is it recommended to store the augmented data (e.g. crop, flip, color jitter, gray scale, gaussian blur) as beton first then load it directly? So we can augment them by transforms and at the same time avoid slowing the FFCV loading. But not sure whether there exists other costs that I was not aware of.

[tavisshore]

No because then you'll lose access to the original data - you'll be stuck with the performed transforms.

Have a beton of your original images and then either augment with FFCV techniques or post-loading

[realliyifei]

Thanks for the reply! Yes I agreed and did it later via zipping the FFCV data loader, and using both FFCV and torchvision's data augmentation (it is a pity that both this FFCV and meta's FFCV-SSL don't work well on the SSL's data augmentation)