aft-pytorch

Unofficial PyTorch implementation of Attention Free Transformer's layers by Zhai, et al. [abs, pdf] from Apple Inc.

Installation

You can install aft_pt via pip:

pip install aft_pt

Usage

You can import the AFT-Full or AFT-Simple layer (as described in the paper) from the package like so:

AFTFull

from aft_pt import AFTFull

layer = AFTFull(
    max_seqlen=20,
    dim=512,
    hidden_dim=64
)

# a batch of sequences with 10 timesteps of length 512 each
x = torch.rand(32, 10, 512)
y = layer(x) # [32, 10, 512]

AFTSimple

from aft_pt import AFTSimple

layer = AFTSimple(
    max_seqlen=20,
    dim=512,
    hidden_dim=64
)

# a batch of sequences with 10 timesteps of length 512 each
x = torch.rand(32, 10, 512)
y = layer(x) # [32, 10, 512]

AFTLocal

from aft_pt import AFTLocal

layer = AFTLocal(
    max_seqlen=20,
    dim=512,
    hidden_dim=64
)

# a batch of sequences with 10 timesteps of length 512 each
x = torch.rand(32, 10, 512)
y = layer(x) # [32, 10, 512]

This layer wrapper is a 'plug-and-play' with your existing networks / Transformers. You can swap out the Self-Attention layer with the available layers in this package with minimal changes.

TODO

• Add full AFT architecture
• Add variants like, AFTConv
• Benchmark using Karpathy's minGPT

Contributing

If you like this repo, please leave a star! If there are any amends or suggestions, feel free to raise a PR/issue.

Credits

@misc{attention-free-transformer,
title = {An Attention Free Transformer},
author = {Shuangfei Zhai and Walter Talbott and Nitish Srivastava and Chen Huang and Hanlin Goh and Ruixiang Zhang and Josh Susskind},
year = {2021},
URL = {https://arxiv.org/pdf/2105.14103.pdf}
}

License

MIT