Size mismatch when loading state dict

[HuiyuanXie]

Thanks for this amazing work!

I tried running the predict_amrs_from_plaintext.py script but came cross a runtime error. It occurred when loading the state_dict of the checkpoint you released for AMR 3.0 for AMRBartForConditionalGeneration. I saw that you suggested that transformers version < 3 should be used. I experimented version 2.11.0 as suggested in your requirements.txt file, and also version 2.8.0, but the problem persisted. The tokenizers is of version 2.7.0. I was wondering if you have any idea regarding what the reason might be, and how I can fix the problem? Many thanks!

I've attached the full error message below:

RuntimeError: Error(s) in loading state_dict for AMRBartForConditionalGeneration:
size mismatch for final_logits_bias: copying a param with shape torch.Size([1, 53587]) from checkpoint, the shape in current model is torch.Size([1, 53075]).
size mismatch for model.shared.weight: copying a param with shape torch.Size([53587, 1024]) from checkpoint, the shape in current model is torch.Size([53075, 1024]).
size mismatch for model.encoder.embed_tokens.weight: copying a param with shape torch.Size([53587, 1024]) from checkpoint, the shape in current model is torch.Size([53075, 1024]).
size mismatch for model.decoder.embed_tokens.weight: copying a param with shape torch.Size([53587, 1024]) from checkpoint, the shape in current model is torch.Size([53075, 1024]).

[mbevila]

Is #4 relevant?

[HuiyuanXie]

Hi @mbevila , thanks for the useful info! Now I'm able to load the pretrained model weights. Thanks very much for the help!