MLPMixer-jax2tf

^{Example usage.}

This repository hosts code for converting the original MLP-Mixer models [1] (JAX) to TensorFlow. The converted models are hosted on TensorFlow Hub and can be found here: https://tfhub.dev/sayakpaul/collections/mlp-mixer/1.

Note that it's a requirement to use TensorFlow 2.6 or greater to use the converted models.

Several model variants are available:

SAM [2] pre-trained (these models were pre-trained on ImageNet-1k):

• B/16 (classification, feature-extractor)
• B/32 (classification, feature-extractor)

ImageNet-1k fine-tuned:

• B/16 (classification, feature-extractor)
• L/16 (classification, feature-extractor)

ImageNet-21k pre-trained:

• B/16 (classification, feature-extractor)
• L/16 (classification, feature-extractor)

For more details on the training protocols, please follow [1, 3].

The original model classes and weights [4] were converted using the jax2tf tool [5]. For details on the conversion process, please refer to the conversion.ipynb notebook.

I independently validated two models on the ImageNet-1k validation set. The table below reports the top-1 accuracies along with their respective logs from tensorboard.dev.

Here is a tensorboard.dev run that logs fine-tuning results (using this model) for the Flowers dataset.

Other notebooks

• classification.ipynb: Shows how to load a Vision Transformer model from TensorFlow Hub and run image classification.
• fine-tune.ipynb: Shows how to fine-tune a Vision Transformer model from TensorFlow Hub on the tf_flowers dataset.

References

[1] MLP-Mixer: An all-MLP Architecture for Vision by Tolstikhin et al.

[2] Sharpness-Aware Minimization for Efficiently Improving Generalization by Foret et al.

[3] When Vision Transformers Outperform ResNets without Pretraining or Strong Data Augmentations by Chen et al.

[4] Vision Transformer GitHub

[5] jax2tf tool

Acknowledgements

Thanks to the ML-GDE program for providing GCP Credit support that helped me execute the experiments for this project.