Add Vision Transformer to `keras.applications`

If you open a GitHub issue, here is our policy:

Vision Transformer[1] is model which uses transformers-like architecture. A sequence of vectors is created by dividing a picture into fixed-size patches, linearly embedding each one, adding position embeddings, and then feeding the assembled vectors to a conventional Transformer encoder.

Who will benefit from this feature?

Keras users using keras.applications in their projects.

Will this change the current api? How? keras.applications.ViTBase, keras.applications.ViTLarge, ...

Contributing

• Do you want to contribute a PR? (yes/no): Yes
• If yes, please read this page for instructions
• Briefly describe your candidate solution(if contributing): I have implemented base model of vision transformer from scratch here

References

[1] https://arxiv.org/abs/2010.11929

@fchollet @LukeWood

[innat]

Check. https://github.com/keras-team/keras-cv/issues/668

[IMvision12]

closing this issue…