OpenKiwi always download the tokenizer files for XLMRoberta even if a local path is configured.

[yym6472]

When I am training the XLM-Roberta based QE system, I pre-downloaded the pre-trained XLM-Roberta model from huggingface's library and modified the field system.model.encoder.model_name in xlmroberta.yaml from the default xlm-roberta-base to my local path that contains the pre-downloaded XLM-Roberta model. However, when running the code, I found OpenKiwi will always download the files config.json and sentencepiece.bpe.model rather than directly use the pre-downloaded ones.

Finally I found this is caused by the following code in kiwi/systems/encoders/xlmroberta.py:48~49:

        if tokenizer_name not in XLM_ROBERTA_PRETRAINED_MODEL_ARCHIVE_LIST:
            tokenizer_name = 'xlm-roberta-base'

which means if model_name is configured to some local path, it will be rewrite to xlm-roberta-base. However, for Bert and XLM, there is no such issue. Is that a bug or under some consideration?

[captainvera]

Hello @yym6472,

Sorry I took a while to respond.

I tested what you are saying, and you are indeed correct. This version of Openkiwi was initially built when transformers was in version < 2 or 3. The transformers library still had a lot of quirks when working with models other than BERT and a couple others.

This was probably introduced to avoid bugs when loading finetuned versions of these models (like you're trying to do here) as we assumed the tokenizer part wouldn't change. (Could it be possible that it didn't used to get saved with the model? I really can't recall anymore)

So as of today with transformers in a much more mature state, I'd probably classify this as a bug. We would really appreciate if you could do a PR to fix it :)

[yym6472]

Thanks for your reply! I have proposed a PR #105 that simply removes the above-mentioned two lines.