rwightman / efficientnet-jax Goto Github PK

Thanks for the great work! I have a few questions:

1. I saw you released a collections of pre-trained model checkpoints, do you have any sample code on how to load them in JAX?
2. Does this pretrained models have comparable performance against that from PT/ TF?
3. Also you mentioned the model weights are ported from timm and tensorflow, I am curious to know how did you do the porting? (i.e. conversion from PT/ TF model weights to JAX's)

Hi, I think this is an interesting repo and give a good contribution for jax community. Currently, I am trying to run the vision transformr code on tpu based on jax. This is a official code (https://github.com/google-research/vision_transformer) and provides the example code on colab tpu (https://colab.research.google.com/github/google-research/vision_transformer/blob/master/vit_jax.ipynb).
However, I find I cannot run the code on colab tpu and maybe you can give me some advice. I think the error is segmentation fault. This question is strange since I don't know who can help me. Thank s very much and have a nice weekend!

Hi, @rwightman thanks for your awesome and helpful work as always!

Do you have any plans to publish jeffnet to PyPI? I've seen that jimm exists at https://pypi.org/project/jimm/, so I assume that you're planning to release a bigger package for JAX/Flax/Objax pre-trained models similar to timm, but could you please in the meantime publish jeffnet? Is there any counterpart on doing so?

Thanks! 😄