upgrade transformers to 4.13.0

[julian-risch]

Proposed changes:

• Transformers is still limited to pytorch < 1.10 https://github.com/huggingface/transformers/blob/6200fd7bbcccc34c60d386af931cfea4f8b2dd53/setup.py#L152 and i found the scatter method in its tapas modeling to break with 1.10 because the method is now called scatter_

The failing test cases are caused by the changes in the following PR in transformers: huggingface/transformers#13873

There is a fix for this one now here on transformers' master branch but no new release yet: huggingface/transformers@24b30d4
However, another problem remains, which is about the scatter method in torch used by TAPAS.
Maybe it can be resolved with:
pip install torch-scatter -f https://data.pyg.org/whl/torch-1.10.0+${CUDA}.html
https://github.com/rusty1s/pytorch_scatter

closes #1658

Status (please check what you already did):

• First draft (up for discussions & feedback)
• Final code
• Added tests
• Updated documentation

[tholor]

Good timing - let's move directly to 4.12: https://github.com/huggingface/transformers/releases/tag/v4.12.0

[lalitpagaria]

Good timing - let's move directly to 4.12: https://github.com/huggingface/transformers/releases/tag/v4.12.0

I would not suggest to use ..0 version better to wait for at least 3 patch version. Of the late seeing many bugs in their ..0 version which gets resolved in patch versions. 😅

[julian-risch]

Good timing - let's move directly to 4.12: https://github.com/huggingface/transformers/releases/tag/v4.12.0

I would not suggest to use ..0 version better to wait for at least 3 patch version. Of the late seeing many bugs in their ..0 version which gets resolved in patch versions. 😅

Good point! Luckily there is 4.12.2 already. https://github.com/huggingface/transformers/releases?page=1 Guess why. ;)
I'm still trying out different versions to find out which change causes some problems. Maybe something about changes of the default values of padding.

[lalitpagaria]

I'm still trying out different versions to find out which change causes some problems. Maybe something about changes of the default values of padding.

I see your are performing sequential search. How about binary search? :)

[julian-risch]

I see your are performing sequential search. How about binary search? :)

That would assume that whenever the tests fail there is something that needs to be changed in haystack. However, transformers introduces some bugs from time to time on their side, for example in 4.11.0 (as you mentioned).

What I found out now is that 4.11.3 let's our tests test_table_reader, test_reader and test_extractive_qa_pipeline fail whereas 4.11.1 works well (besides a small error in zero shot classification, which I assume has been fixed on transformers side).

[lalitpagaria]

I see your are performing sequential search. How about binary search? :)

That would assume that whenever the tests fail there is something that needs to be changed in haystack. However, transformers introduces some bugs from time to time on their side, for example in 4.11.0 (as you mentioned).

What I found out now is that 4.11.3 let's our tests test_table_reader, test_reader and test_extractive_qa_pipeline fail whereas 4.11.1 works well (besides a small error in zero shot classification, which I assume has been fixed on transformers side).

No, I mean to speed up finding a rouge version.
We can try the first mid of (4.12.2 and 4.7.0) if build pass then select next version between (mid, 4.12.2) otherwise between (mid, 4.7.0) and so on. :)

[tholor]

Quick update: PyTorch has been upgraded to 1.10 already in a separate PR (#1789 ). For transformers, we still wait for one bug fix. The fix is already merged, but it didn't make it to their last release, unfortunately. We should be ready to bump the transformers version with their next release.