Probing ViTs

By Aritra Roy Gosthipaty and Sayak Paul (equal contribution)

In this repository, we provide tools to probe into the representations learned by different families of Vision Transformers (supervised pre-training with ImageNet-21k, ImageNet-1k, distillation, self-supervised pre-training):

• Original ViT [1]
• DeiT [2]
• DINO [3]

We hope these tools will prove to be useful for the community. Please follow along with this post on keras.io for a better navigation through the repository.

Updates

• June 3, 2022: The project got the Google OSS Expert Prize.
• May 10, 2022: The project got a mention from Yannic Kilcher in ML News. Thanks, Yannic!
• May 4, 2022: We're glad to receive the #TFCommunitySpotlight award for this project.

Self-attention visualization

Original Image	Attention Maps	Attention Maps Overlayed

output-dino.mp4

Original Video Source

output-dog.mp4

Original Video Source

Supervised salient representations

In the DINO blog post, the authors show a video with the following caption:

The original video is shown on the left. In the middle is a segmentation example generated by a supervised model, and on the right is one generated by DINO.

A screenshot of the video is as follows:

We obtain the attention maps generated with the supervised pre-trained model and find that they are not that salient w.r.t the DINO model. We observe a similar behaviour in our experiments as well. The figure below shows the attention heatmaps extracted with a ViT-B16 model pre-trained (supervised) using ImageNet-1k:

Dinosaur	Dog

We used this Colab Notebook to conduct this experiment.

Hugging Face Spaces

You can now probe into the ViTs with your own input images.

Attention Heat Maps	Attention Rollout

Visualizing mean attention distances

Methods

We don't propose any novel methods of probing the representations of neural networks. Instead we take the existing works and implement them in TensorFlow.

• Mean attention distance [1, 4]
• Attention Rollout [5]
• Visualization of the learned projection filters [1]
• Visualization of the learned positioanl embeddings
• Attention maps from individual attention heads [3]
• Generation of attention heatmaps from videos [3]

Another interesting repository that also visualizes ViTs in PyTorch: https://github.com/jacobgil/vit-explain.

Notes

We first implemented the above-mentioned architectures in TensorFlow and then we populated the pre-trained parameters into them using the official codebases. In order to validate this, we evaluated the implementations on the ImageNet-1k validation set and ensured that the reported top-1 accuracies matched.

We value the spirit of open-source. So, if you spot any bugs in the code or see a scope for improvement don't hesitate to open up an issue or contribute a PR. We'd very much appreciate it.

Navigating through the codebase

Our ViT implementations are in vit. We provide utility notebooks in the notebooks directory which contains the following:

• dino-attention-maps-video.ipynb shows how to generate attention heatmaps from a video. (Visually,) best results were obtained with DINO.
• dino-attention-maps.ipynb shows how to generate attention maps from individual attention heads from the final transformer block. (Visually,) best results were obtained with DINO.
• load-dino-weights-vitb16.ipynb shows how to populate the pre-trained DINO parameters into our implementation (only for ViT B-16 but can easily be extended to others).
• load-jax-weights-vitb16.ipynb shows how to populate the pre-trained ViT parameters into our implementation (only for ViT B-16 but can easily be extended to others).
• mean-attention-distance-1k.ipynb shows how to plot mean attention distances of different transformer blocks of different ViTs computed over 1000 images.
• single-instance-probing.ipynb shows how to compute mean attention distance, attention-rollout map for a single prediction instance.
• visualizing-linear-projections.ipynb shows visualizations of the linear projection filters learned by ViTs.
• visualizing-positional-embeddings.ipynb shows visualizations of the similarities of the positional embeddings learned by ViTs.

DeiT-related code has its separate repository: https://github.com/sayakpaul/deit-tf.

Models

Here are the links to the models where the pre-trained parameters were populated:

• Original ViT model (pretrained on ImageNet-21k and fine-tuned on ImageNet-1k)
• Original ViT model (pretrained on ImageNet-1k)
• DINO model (pretrained on ImageNet-1k)
• DeiT models (pretrained on ImageNet-1k including distilled and non-distilled ones)

Training and visualizing with small datasets

Coming soon!

References

[1] An Image is Worth 16x16 Words: Transformers for Image Recognition at Scale: https://arxiv.org/abs/2010.11929

[2] DeiT: https://arxiv.org/abs/2012.12877

[3] DINO: https://arxiv.org/abs/2104.14294

[4] Do Vision Transformers See Like Convolutional Neural Networks?: https://arxiv.org/abs/2108.08810

[5] Quantifying Attention Flow in Transformers: https://arxiv.org/abs/2005.00928

Acknowledgements

• PyImageSearch
• Jarvislabs.ai
• GDE Program