BartForQuestionAnswering

[patil-suraj]

This PR adds BartForQuestionAnswering.

Decided to add this models as BART is intended for both NLU and NLG tasks and also achieves comparable performance to ROBERTa on SQuAD.

Also fine-tuned the model here. The metrics are slightly worse than given in the paper. Got following metrics on SQuADv1
{'exact_match': 86.80227057710502, 'f1': 92.73424907872341}

@sshleifer , @patrickvonplaten

[codecov]

Codecov Report

Merging #4908 into master will increase coverage by 0.03%.
The diff coverage is 94.11%.

@@            Coverage Diff             @@
##           master    #4908      +/-   ##
==========================================
+ Coverage   76.99%   77.02%   +0.03%     
==========================================
  Files         128      128              
  Lines       21602    21635      +33     
==========================================
+ Hits        16633    16665      +32     
- Misses       4969     4970       +1     

Impacted Files	Coverage Δ
src/transformers/__init__.py	99.13% <ø> (ø)
src/transformers/modeling_bart.py	96.26% <93.93%> (-0.15%)	⬇️
src/transformers/modeling_auto.py	78.40% <100.00%> (ø)
src/transformers/modeling_utils.py	90.49% <0.00%> (-0.12%)	⬇️
src/transformers/modeling_tf_utils.py	87.42% <0.00%> (+0.15%)	⬆️
src/transformers/file_utils.py	73.49% <0.00%> (+0.40%)	⬆️

---

Continue to review full report at Codecov.

Legend - Click here to learn more
Δ = absolute <relative> (impact), ø = not affected, ? = missing data
Powered by Codecov. Last update ac99217...63eb191. Read the comment docs.

[sshleifer]

LGTM

[sshleifer]

Thanks for the contribution @patil-suraj !

[LysandreJik]

Hi! Very cool @patil-suraj.

Could you also add BartForQuestionAnswering to the all_model_classes in test_modeling_bart.py?

[patil-suraj]

Hi! Very cool @patil-suraj.

Could you also add BartForQuestionAnswering to the all_model_classes in test_modeling_bart.py?

Hi, @LysandreJik
After adding BartForQuestionAnswering in all_model_classes I also had to add output_attention parameter to forward.

Now for some reason test_attention_outputs is failing, I am not sure why, could you help me fix it ?
Thanks !

[patrickvonplaten]

Awesome work @patil-suraj - I can help you with this test :-)

[patrickvonplaten]

I see what the problem is...it's actually not related to your PR at all. Can we you for now just remove BartForQuestionAnswering from the all_models tuples in the tests. @LysandreJik @sshleifer I will open a new PR after this one to fix it :-)

[patil-suraj]

I see what the problem is...it's actually not related to your PR at all. Can we you for now just remove BartForQuestionAnswering from the all_models tuples in the tests. @LysandreJik @sshleifer I will open a new PR after this one to fix it :-)

Thank you @patrickvonplaten . I've removed it from all_models tuple for now