How to convert models with two vocab files to PyTorch?

[phiradet]

Hi,

I would like to get translation result from the eng-kor model with transformers.MarianMTModel and transformers.MarianTokenizer. I understand we need to first convert the model to PyTorch format with convert_marian_tatoeba_to_pytorch.py first.

The eng-kor has two different vocab sets for encoder and decoder. How can we use transformers.models.marian.convert_marian_to_pytorch.convert function to do the conversion?
Because there is no vocab.yml file in the zip file, I found the line 381 throws IndexError: list index out of range error.

Thanks

[jorgtied]

This model uses two separate vocabularies and it does not properly convert to pytorch and huggingface at the moment. Hopefully, this will be added soon to the conversion procedures.

[phiradet]

Thanks @jorgtied
I plan to use the model for backtranslation. While waiting for the update, I am ok to use the zho-kor model instead. But, the link https://object.pouta.csc.fi/Tatoeba-MT-models/zho-kor/opus-2020-09-10.zip shown in the zho-kor README.md doesn't work. Could you provide me the correct link?

Thanks

[jorgtied]

The latest conversion scripts in the transformer library support the conversion of models with two vocabs. You may also check my recipes in https://github.com/Helsinki-NLP/Opus-MT/tree/master/hf

[jorgtied]

Thanks @jorgtied I plan to use the model for backtranslation. While waiting for the update, I am ok to use the zho-kor model instead. But, the link https://object.pouta.csc.fi/Tatoeba-MT-models/zho-kor/opus-2020-09-10.zip shown in the zho-kor README.md doesn't work. Could you provide me the correct link?

Thanks

I removed that model because it was so poor (at least according to the scores). I should create new ones for this language pair.

[zhiqihuang]

The latest conversion scripts in the transformer library support the conversion of models with two vocabs. You may also check my recipes in https://github.com/Helsinki-NLP/Opus-MT/tree/master/hf

Hi, I still got the same error IndexError: list index out of range because of the two vocab files.

I used the script from transformers, i.e., python3 -m transformers.models.marian.convert_marian_to_pytorch and my transformers version is transformers-4.21.1.

I also tried the convert_to_pytorch.py script you suggested, same error.

Can you show me the command to convert such two vocab model pytorch?

Thanks

[Limess]

More resources on these split vocab models would be helpful. I'm also trying to compile these to CTranslate2 and having difficulties due to the split vocabs.