Similar to ImageCleaner dropping validation data?, I am wondering whether it wouldn’t make sense to include the valid_ds in the output of ImageCleaner. Especially, if you are using the procedure with cleaning and removing duplicates you end up with 64% of your original data (assuming you use 20% split and minus the images you deleted) or is the recommendation to keep the original valid_ds? What would be the reason for that?
I naively tried:
clean_data.valid_ds = data.valid_ds
clean_data.valid_dl = data.valid_dl
Is there a way to reuse the validation from a different databunch?