Closed wenyudu closed 4 years ago
Thanks for the PR! Looks good to me. I'll take a closer look at this sometime next week.
Did you get a chance to finetune SpanBERT with the changes on English Coref? If so, could you please share the results? Thanks!
Yeah, sure. I tried SpanBert base model and replicated the reported result (Test F1 ~ 77.7, Dev F1 ~ 77.4). I don't have a 32GB GPU at hand but I will try to run large model when I have a chance.
Fix bugs in _pytorch_totf.py. Now the model can directly use SpanBert PyTorch pretrained models recently released from https://github.com/facebookresearch/SpanBERT.