#85 Hotfix/RobertaTokenizerFast object has no attribute max_len

[kirzharov]

Link to the issue: #85

[kirzharov]

I checked score function locally and get results for precision, recall, and F1 measure with transformers versions:

• 4.0.0
• 3.5.1
• 2.8.0

Looks like everything works as before for transformers version 4.0.0

[felixgwu]

Hi @kirzharov
Thank you for this pull request. I've verified that old versions work as before. However, it doesn't pass our test with transformers 4.0.0. It's possible that some other arguments have to be changed as well.

[kirzharov]

Hi @felixgwu, thanks! Did you mean about only score function or others? I'll check 👍

[felixgwu]

When I ran python -m unittest discover, almost all the tests failed. It might be that they changed some default settings and the embeddings changed, but I haven't looked into the details.

[kirzharov]

@felixgwu Ok, thanks! I'll check today 👍

[felixgwu]

Hi @kirzharov, we have fixed the bug and merged your PR. FYI, the mismatch comes from the fact that huggingface's fast tokenizers are inconsistent with the old tokenizers and create different tokens. We disable the fast tokenizers currently. Thanks again for your help!

[kirzharov]

@felixgwu, awesome, thanks! It's great that you managed to fix this quickly. 🔥 I also noticed difference with new Fast tokenizers and use_fast=False 👍

[kalyankumarp]

Hi,
Do I have to wait for the 0.3.7 version of bert score to be able to use transformers 4.0.0? I tried 0.3.6 bert score with 4.0.0 transformers. I am still getting the same error.

Thanks,
Kalyan.

[felixgwu]

Hi @kalyankumarp,
We have released veresion 0.3.7. Can you please share the error messages if it still doesn't work?

[kalyankumarp]

Hi Felix,

It's working. Thanks for the fix.