wilds
wilds copied to clipboard
Waterbirds give 0 worst-group accuracy
Training waterbirds dataset out-of-the-box gives 0 worst group accuracy. Digging deeper, I noticed that all the predictions immediately become the 0 label. Any advice would be helpful. Thanks in advance.
Digging deeper, it seems that setting lr=1e-3 and wd=1e-4 is necessary to get reasonable results (i.e. non-degenerate solutions). It seems these are the hyperparameters recommended by most papers using this dataset (e.g,. https://arxiv.org/pdf/2210.11369.pdf). Perhaps the default values of these hyperparameters should be changed?