Working on subset of data

[vijayakuruba]

Hi,

Thank you for providing the WILDS benchmark. It has been really helpful, very clean and easily reproducible benchmarks.

I have a question regarding working on fraction of dataset for fast development on a small dataset.
When we take fraction of dataset, does the original classes/species and proportion of camera traps are preserved? How do we ensure this? Looks like the code is randomly sampling the split. Please let me know.
Thank you!

[ssagawa]

Thank you for your kind words and your interest! Yes, we randomly sample examples when we take a fraction of the dataset:

wilds/wilds/datasets/wilds_dataset.py

Line 58 in a7a452c

def get_subset(self, split, frac=1.0, transform=None):

. So the proportions of classes and cameras are roughly preserved, modulo random variation.

[vijayakuruba]

Thank you!