Closed chunshengshi closed 3 months ago
@zhang-tao-whu Please check this issue. It seems a bug from xtuner.
Thank you very much for your discovery. This is a bug from the xtuner library. After fixing this bug, we attempted retraining, and there was a little performance improvement. For example, the cIoU on refCOCO val without finetuning increased from 65.6 to 66.3.
In the omg_llava_collate_fn.py file, there is a problem with attention_mask[:i] = True. I think it should be attention_mask[:,:i] = True. Hope that answers, thanks!