[BUG_FIX] RobertaTokenizerFast object has no attributes max_len

[davebulaval]

Quick fix for the transfomers update V4.0.0

Regarding the tokenizer classes:

The tokenizer attribute max_len becomes model_max_length.

[davebulaval]

Also, I've noticed some dead import, no styling rules (Pylint and YAPF) some dead code and the build is impossible to pass on Travis. If the maintainers are interested, I can push a PR including some fix for it.

And the example is version outdated (bert_Score and Transformers)

Also, I would recommend not trying to make the test pass in Travis since it is confusing.

[felixgwu]

Hi @davebulaval, thank you for pointing this out.
However, your changes would break the backward compatibility with the older transformers versions. Would you please modify your PR to make sure that the code works with version 2.0.0 and 3.0.0?
If you don't have time, we can fix it ourselves.

[kalyankumarp]

Hi guys,

Thanks for your awesome work on the bert_score library. Its definitely a better evaluation metric than rouge. I have a text summarization project submission on Tuesday. Until a few days ago, I was using the rouge scores for hyper parameter tuning. I believe that I will get better summaries when tuned with Bert scores.
I would really appreciate if you guys fix this compatibility issue soon.

[davebulaval]

Hi guys,

Thanks for your awesome work on the bert_score library. Its definitely a better evaluation metric than rouge. I have a text summarization project submission on Tuesday. Until a few days ago, I was using the rouge scores for hyper parameter tuning. I believe that I will get better summaries when tuned with Bert scores.
I would really appreciate if you guys fix this compatibility issue soon.

For a quick fix, just pip install -U git+https://github.com/davebulaval/bert_score.git@master

[kalyankumarp]

Hi guys,
Thanks for your awesome work on the bert_score library. Its definitely a better evaluation metric than rouge. I have a text summarization project submission on Tuesday. Until a few days ago, I was using the rouge scores for hyper parameter tuning. I believe that I will get better summaries when tuned with Bert scores.
I would really appreciate if you guys fix this compatibility issue soon.

For a quick fix, just pip install -U git+https://github.com/davebulaval/bert_score.git@master

Thanks. Will try that.

[felixgwu]

Hi @kalyankumarp,

We recommend that you use transformers with version before 4.0.0 currently because neither this PR or #86 passed our unittest. Before we figure out where the bugs are, it's recommended to use an old version. Otherwise, the numbers you report in your paper won't be comparable to the ones in other papers.

Best, Felix

[kalyankumarp]

Hi Felix,

I thought of using transformers of version 3.0.0 but I am using BART model of SimpleTransformers library for text summarization. This library needs transformers of 4.0.0 version.

Thanks,
Kalyan.

[felixgwu]

Hi @kalyankumarp,
We have fixed the problem and the code in the master branch now supports transformers version 4.0.0 and has passed all our unittests. Please feel free to pull the newest commit.

[felixgwu]

Hi @davebulaval,
Thanks again for being the first person pointing out this problem and submitting this PR. Given that you didn't respond to our request for modification and that supporting version 4.0.0 is urgent, we decided to take #86. The reason is that it is compatible with the old version and fixed the problem ourselves. However, we still really appreciate your help.

[kalyankumarp]

Thanks, Felix. Will use the master code.

[kalyankumarp]

Hi @kalyankumarp,
We have fixed the problem and the code in the master branch now supports transformers version 4.0.0 and has passed all our unittests. Please feel free to pull the newest commit.

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. It is still going to the block "if LooseVersion(transformers.version) >= LooseVersion("3.0.0"):"