@patrickvonplaten I would like to start with Maskformer for Tensorflow/Pytorch. Catch up with how the event goes.

Awesome! Let me know if you have any questions :-)

Hello! I'd like to take on Longformer for Tensorflow/Pytorch please.

@patrickvonplaten I would like to start with T5 for pytorch and tensorflow

Sounds great!

LayoutLM is also taken as mentioned by a contributor on Discord!

@patrickvonplaten I would take GPT and GPT-J (TensorFlow editions) if those are still available.

I'm guessing GPT is GPT2?

I will take Bert, Albert, and Bigbird for both Tensorflow/Pytorch

I'll take Swin and ViT for Tensorflow

I'd like DistilBERT for both TF and PT please

@patrickvonplaten I would take GPT and GPT-J (TensorFlow editions) if those are still available.

I'm guessing GPT is GPT2?

@cakiki You can go for GPT2 (I updated the name in the test)

Can I try GPT2 and GPTJ for Pytorch? if @ydshieh you are not doing so?

I would like to try CLIP for Tensorflow and PyTorch.

I'll take CANINE and TAPAS.

Can I try GPT2 and GPTJ for Pytorch? if @ydshieh you are not doing so?

@ArEnSc
No, you can work on these 2 models :-) Thank you!

@ydshieh Since the MobileBertForSequenceClassification is the copy of BertForSequenceClassification, so I think I will do check doc-test of MobileBert as well to overcome the error from make fixup

I'll take FlauBERT and CamemBERT.

@abdouaziz Awesome! Do you plan to work on both PyTorch and TensorFlow versions, or only one of them?

I would like to work on LUKE model for both TF and PT

@Tegzes you're lucky because there's no LUKE in TF ;) the list above actually just duplicates all models, but many models aren't available yet in TF.

In this case, I will also take DeBERTa and DeBERTa-v2 for PyTorch

@ydshieh

I plan to work only with PyTorch

@Tegzes you're lucky because there's no LUKE in TF ;) the list above actually just duplicates all models, but many models aren't available yet in TF.

True - sorry I've been lazy at creating this list!

Happy to work on TrOCR (pytorch and TF)

I take RoBERTa in PT and TF

I would like to pick up XLM-RoBERTa in PT and TF.

I can work on ELECTRA for PT and TF

@ydshieh
I got the same error with hf-internal-testing working on Reformer. And I couldn't find any checkpoints without randomly initialized.
#16292 (comment)

@ydshieh I got the same error with hf-internal-testing working on Reformer. And I couldn't find any checkpoints without randomly initialized. #16292 (comment)

Hi, @hiromu166

Looks like we didn't upload a tokenizer for hf-internal-testing/tiny-random-reformer. These tiny models are created for some internal testing purpose, and we didn't have the intention to use it for doctest until now.

I will upload a tokenizer today! Thank you very much for pointing out this.

@hiromu166

I uploaded the tokenizer file spiece.model to hf-internal-testing/tiny-random-reformer.

Let me know if you have any problem using it in the doctest 🙏, thanks!

@ydshieh
Thank you for dealing with it quickly!!
I'll check it soon.

Hi @ydshieh,

For TFElectra i am getting answer index issue

doc.py file TF_QUESTION_ANSWERING_SAMPLE has target index at 14 and 15, while i am getting my prediction at 10, 12 index

@ydshieh quick question. @jessecambon and I are working on the DistilBert model but we're seeing many of the examples are actually drawn and built from the examples in src/transformers/utils/doc.py which seems to be a global location, meaning it seems multiple models get their doc tests from this single file. So my question is whether someone else has edited those already or if we're free to edit them as part of our work, which will then flow out to other model architectures.

Hi @bhadreshpsavani , Yes there is an issue regarding the target indices. Please follow this discussion

#16523 (comment)
#16523 (comment)

I will discuss the team.

Hi @ydshieh,

For TFElectra i am getting answer index issue

doc.py file TF_QUESTION_ANSWERING_SAMPLE has target index at 14 and 15, while i am getting my prediction at 10, 12 index

Hi, @jmwoloso & @jessecambon

doc.py is not meant to be modified by the sprint contributor :-)

In order to customize, you can use add_code_sample_docstrings, and provide it with expected_output=..., expected_loss=..., checkpoint=..., etc.

You can looks this change on Roberta as a reference.

Let me know if you have any difficulty using this approach.

@ydshieh quick question. @jessecambon and I are working on the DistilBert model but we're seeing many of the examples are actually drawn and built from the examples in src/transformers/utils/doc.py which seems to be a global location, meaning it seems multiple models get their doc tests from this single file. So my question is whether someone else has edited those already or if we're free to edit them as part of our work, which will then flow out to other model architectures.

Hi @bhadreshpsavani , Yes there is an issue regarding the target indices. Please follow this discussion

#16523 (comment) #16523 (comment)

I will discuss the team.

Hi @ydshieh,
For TFElectra i am getting answer index issue
doc.py file TF_QUESTION_ANSWERING_SAMPLE has target index at 14 and 15, while i am getting my prediction at 10, 12 index

By the way, you don't have this issue for Electra (PyTorch version)? Do you use different checkpoint for PyTorch/TensorFlow Electra?

Hi @bhadreshpsavani , Yes there is an issue regarding the target indices. Please follow this discussion
#16523 (comment) #16523 (comment)
I will discuss the team.

Hi @ydshieh,
For TFElectra i am getting answer index issue
doc.py file TF_QUESTION_ANSWERING_SAMPLE has target index at 14 and 15, while i am getting my prediction at 10, 12 index

By the way, you don't have this issue for Electra (PyTorch version)? Do you use different checkpoint for PyTorch/TensorFlow Electra?

Hi @ydshieh,
Actually, I am using the same checkpoint.

Hi @ydshieh, Actually, I am using the same checkpoint. In python, we use where to find index but here its hardcoded i guess

What I saw in PT_QUESTION_ANSWERING_SAMPLE

    ```python
    >>> # target is "nice puppet"
    >>> target_start_index, target_end_index = torch.tensor([14]), torch.tensor([15])

    >>> outputs = model(**inputs, start_positions=target_start_index, end_positions=target_end_index)
    >>> loss = outputs.loss
    >>> round(loss.item(), 2)
    {expected_loss}

So if you use the same checkpoint for PT/TF Electra, the PyTorch should have the same issue I think.
Could you try run the doc test (for PT Electra) and see if it pass?

Also, let's move this discussion to your PR.

@ydshieh ok, thank you for the clarification. so is the goal then to move away from doc.py and have individual doc tests in the specific model architecture files?

Hi, @jmwoloso not exactly. We still use doc.py, but in the individual model file(s), we tried to use a method add_code_sample_docstrings and provide some arguments.

The method add_code_sample_docstrings itself will use the example in doc.py and customize the doc examples using the arguments you provide.

ok, awesome. thanks @ydshieh!

Hi,

Is anyone working on BART TF version? If not I can try and contribute WRT that model.
Else can you suggest some model from above list(with slightly lower complexity as this will be my first attempt at contribution to transformers library :) ) which can be taken up?

Thanks

Not sure if someone is working on TF Bart. You can search the Pull request list.

Otherwise, TF Wav2Vec2 should be a good one to try. Should be quite easy with the recent change in

#16494 (comment)

Since no one seems to have taken layoutlmv2 and the last activity on this thread was 25 days ago, I’m going to start working on layoutlmv2 :)

Edit: The PR is ready for review now

@patrickvonplaten I would like to start with RoFormer for TF and PT.

@nandwalritik - that's great! Do you want to open a PR for this?

Hi @patrickvonplaten , I am interested in working on data2VecText. Thanks!