Open
Description
Hello,
I think there is an issue on LVIS-92i dataloader when it is used in train mode. Specifically at line 71 train_cat_ids = list(train_anno.keys())
there is no check if the categories have at least nshot samples (this is instead done in case of validation).
Proposed fix:
train_cat_ids = [i for i in list(train_anno.keys()) if len(train_anno[i]) > self.shot]
Metadata
Metadata
Assignees
Labels
No labels