Thread by @RisingSayak

What do the Vision Transformers learn? How do they encode anything useful for image recognition? In our latest work, we reimplement a number of works done in this area & investigate various ViT model families (DeiT, DINO, original, etc.).

Done w/ @ariG23498

1/

We also reimplemented different models in #Keras. These were first populated w/ pre-trained parameters & were then evaluated to ensure correctness.

Code, models, a tutorial, interactive demos (w/ @huggingface Spaces), visuals:

github.com/sayakpaul/prob…

2/

We’ve used the following methods for our analysis:

* Attention rollout
* Classic heatmap of the attention weights
* Mean attention distance
* Viz of the positional embeddings & linear projections

We hope our work turns out to be a useful resource for those studying ViTs.

3/

We’ve also built a @huggingface organization around our experiments. The organization holds the Keras pre-trained models & spaces where you can try the visualization on your own images.

huggingface.co/probing-vits

Contributions are welcomed :)

4/

We thank @fchollet for his helpful guidance on the tutorial. We thank @jarvislabsai & @GoogleDevExpert for providing us with credit support that allowed the experiments.

Thanks to @ritwik_raha for helping us with this amazing visual.

5/