Converting fairseq NMT to transformers misses model weight

Hi there, question about fairseq NMT model (FSMT) conversion.

I tried to convert my own fairseq-nmt model (transformer_wmt_en_de) based on this conversion script.
However, decoder.embed_out weight is missing after converting fairseq model to transformers FSMT model. This parameter exists when not specifing --share-all-embeddings or --share-decoder-input-output-embed, while official fairseq wmt models do not have decoder.embed_out weight because specifying --share-all-embedding.
pytorch/fairseq#2537

Are there any solution or tips to converting own fairseq model?

1 possible answer(s) on “Converting fairseq NMT to transformers misses model weight