hf-internal-testing

/

tiny-random-layoutlmv2

like 0

@nielsr could you check here?

Also CC @Narsil who was helping with this. Context is here: https://github.com/huggingface/transformers/pull/18414#issuecomment-1227410573

Hi everyone, just wanted to bump this

Can we confirm hf-internal-testing/tiny-random-layoutlmv2 isn't used yet in HuggingFace Transformers? cc @Narsil

If helpful, the only reference to it in the repo is here:

$ git grep tiny-random-layoutlmv2
tests/deepspeed/test_model_zoo.py:LAYOUTLMV2_TINY = "hf-internal-testing/tiny-random-layoutlmv2"

I took a look and LAYOUTLMV2_TINY is defined but never used.

Ok so discussed this offline with @narsil , it's unclear to both of us why we need to make this "tiny" LayoutLMv2 model a bit bigger. Wouldn't it work with this model already?

I am looking into it( sorry for the delay).

Ok This modification does make things work (the failure is in detectron2 itself without the modification, which means I'm not willing to invest time to understand what's happening).

What's more important is that this configuration change is compatible with the small weights the are here (I thought I had to modify the config to produce smaller weights, but maybe it was not necessary)