NielsRogge

Yeah I meant my notebook. Your code works just fine with my PR: ``` Iter 199/200 (0.0013 s) train_loss: 0.4010, train_acc: 0.9857, val_loss: 0.9357, val_acc: 0.7880 Test set acc: 0.8130000233650208...

Hi @gcucurull, my GCN implementation in FLAX is working. However, I've also added the Graph Attention Network, but there's probably a mistake in there, could you please have a look?...

Awesome work @unography - are you planning to add support for GroupViT and OWL-ViT as well?

This PR seems almost ready, I'd just update: * all code examples to use either `LayoutLMTokenizer` or `AutoTokenizer` * add a working code example of `LayoutLMForQuestionAnswering`/`TFLayoutLMForQuestionAnswering`, with an expected output

> I actually don't have a pre-trained TFLayoutLMForQuestionAnswering (i.e. one with tensorflow weights), but I could use the same code and just reference the base model? The Transformers library makes...

@ankrgyl normally if you run `make fixup` and it complains about a model not being in any auto mapping, you can add it to utils/check_repo.py in the IGNORE_NON_AUTO_CONFIGURED mapping. Then,...

To make all tests pass, you need to overwrite the `_prepare_for_class` method defined in `test_modeling_common.py`, to make sure the targets are prepared correctly for `LayoutLMForQuestionAnswering`. It can be defined as...

Regarding the failing tests - you might need to rebase with the main branch. Also note that sometimes, tests which are totally unrelated to your PR fail, in which case...

I've pinged @sgugger for a final review, however he's off this week so will be merged next week :)

I'd like to communicate on that once the pipeline is merged, because the Space above is using that right? Also, the doc tests don't seem to pass: ``` _ [doctest]...