Vision Transformers on TF Hub

Ever wanted to use Vision Transformers with TFHub and Keras? Well, pull your socks up now and get started. 16 different models are available for classification and fine-tuning. More details:

The story does not end here. Using this notebook you can convert any supported model from the AugReg pool (> 50,000 models!) and use that inside TFHub and Keras: https://colab.research.google.com/github/sayakpaul/ViT-jax2tf/blob/main/conversion.ipynb.

5 Likes

Well done Sayak! great work as usual!!!

Just to complement, link to the the TFHub collection: TensorFlow Hub

1 Like

@lgusm It could be nice to have also the official garden implementations in TFHub when they will be finalized:

https://github.com/tensorflow/models/tree/master/official/vision/beta/projects/vit

I think soon or later we also need to understand if we need to handle a duplicates policy in TFHUB. If not at some point we could confuse users.

1 Like

I agree, that’s a good point!

1 Like

In that specific case it was introduced by a community PR so I think that we could start to coordinate with the model garden review team to suggest to the users an extendend contribution to TFhub when It is merged in TF model garden.

1 Like