dumpmemory/fnet

PyTorch implementation of FNet: Mixing Tokens with Fourier transforms

FNet

PyTorch implementation of FNet: Mixing Tokens with Fourier Transforms.

Quickstart

Clone this repository.

git clone https://github.com/jaketae/fnet.git

Navigate to the cloned directory. You can start using the model via

>>> from fnet import FNet
>>> model = FNet()

By default, the model comes with the following parameters:

FNet(
    d_model=256,
    expansion_factor=2,
    dropout=0.5,
    num_layers=6,
)

Summary

While transformers have proven to be successful in various domains, its O(n^2) computation complexity has been considered a structural weakness. Many attempts have been made to optimize the model architecture. The authors of the paper present FNet, a model that replaces self-attention with standard unparametrized Fourier transforms. Not only is FNet faster and computationaly more efficient than the classic transformer, but it also retains 92% of BERT's accuracy on the GLUE benchmark. Given a smaller number of parameters, FNet outperformed transformers.

Resources

• Original Paper
• Rishikesh's implementation
• Image from SyncedReview's Medium article