ConvNeXt is added to HuggingFace Transformers

[gante]

Hi ConvNeXt team 👋

We've recently added ConvNeXt to HuggingFace Transformers: https://huggingface.co/docs/transformers/master/en/model_doc/convnext.

All checkpoints are on the hub: https://huggingface.co/models?other=convnext. We are also close to merging its TensorFlow equivalent, as well as corresponding pre-trained weights in a TF format - ConvNeXt is expected to be one of the main features in our next transformers release.

The cool thing about the hub is that it's entirely git-based: each model has its own Git repository, to which you can git add, git commit and git push. The model card is just a README that can be added, similar to how you would add a README to a Github repo. You can check the files (and git history) of a particular checkpoint here for instance.

We've added the models as part of the Facebook organization (to which other models, like BART, DPR, Wav2vec 2.0 etc. also belong). We were wondering if you're interested in becoming part of the Facebook organization on the hub. This allows you to have write access to all model repos of Facebook, which means you can easily upload new checkpoints, write model cards, etc.

Let me know what you think!

Kind regards,

João & HuggingFace team

[johnypark]

Hi @gante, any updates on tensorflow equivalent? I would like to locate corresponding pre-trained weights for tensorflow models. It seems that quite of bit of work is already been done here:

https://github.com/sayakpaul/ConvNeXt-TF/blob/main/models/model_configs.py

[gante]

Hey @johnypark -- The pre-trained weights were converted to TF and uploaded to the same Hugging Face repository as the original weights (e.g. here). You should be able to load the pre-trained TF model using Hugging Face's tools (see an example here) :)