perceiver-pytorch

Why we included this project

Researchers and ML engineers who want the Perceiver architecture without reimplementing it will find this a direct, readable PyTorch port of the original 2021 paper, plus the Perceiver IO follow-up for flexible output lengths. A single class handles images, video, or any axis-shaped data, and the hyperparameters for latent count, dimension, and attention depth stay close to the paper's design, so the package is easy to configure and experiment with. It is small, MIT-licensed, and includes an experimental bottom-up attention variant. That makes it handy for prototyping and coursework where you want to trace how iterative attention works before committing to a larger framework.