Support for TFDS Sub-datasets such as cycle_gan's horse2zebra or iphone2dslr_flower

[ferreirafabio]

May it be possible that there is currently no support to load pre-downloaded tfds datasets from a dir (in below example: ./datasets)? For example:

python train.py datasets --dataset tfds/cycle_gan/iphone2dslr_flower --model edgenext_small

produces the below error:

No builder could be found in the directory: iphone2dslr_flower for the builder: iphone2dslr_flower.

Replacing the /cycle_gan/iphone2dslr_flower with /cycle_gan_iphone2dslr_flower does not work either. I assume this is because the reader_factory.py (see below) splits the / twice but takes only the last split which is iphone2dslr_flower.

def create_reader(name, root, split='train', **kwargs):
    name = name.lower()
    name = name.split('/', 2)
    prefix = ''
    if len(name) > 1:
        prefix = name[0]
    name = name[-1]
    ...

Or are such sub-datasets indeed supported and I overlooked how to use or load them?
Thanks in advance!

[rwightman]

@ferreirafabio I was literally looking at this today for HF datasets name handling, can you try

changing the split line to name = name.split('/', 1)

[ferreirafabio]

@rwightman hehe, that change then allows using --dataset tfds/cycle_gan/iphone2dslr_flower indeed! Thank you 👍