With a considerable amount of help from Willi Gierke, we were able to port the original .npz
weights of new BiT-ResNet models [1] to TensorFlow and Keras.
Colab Notebook showing the porting and all the other details are available here [2]. The converted models are available on TF-Hub as well:
These models are a part of the recent work done by Google Brain on Knowledge Distillation [3].
References
[1] Big Transfer (BiT): General Visual Representation Learning by Kolesnikov et al.
[2] BiT-jax2tf: GitHub - sayakpaul/BiT-jax2tf: This repository hosts the code to port NumPy model weights of BiT-ResNets to TensorFlow SavedModel format.
[3] Knowledge distillation: A good teacher is patient and consistent by Beyer et al.