Awesome

RealFormer

PyTorch implementation of RealFormer: Transformer Likes Residual Attention.

Quickstart

Clone this repository.

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

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

>>> from realformer import RealFormerEncoder
>>> model = RealFormerEncoder()

By default, the model comes with the following parameters:

RealFormerEncoder(
    d_model=512,
    num_heads=8,
    expansion_factor=2,
    dropout=0.5,
    max_len=512,
    num_layers=6,
)

Summary

Residual Attention Layer Transformer, shortened as RealFormer, is a transformer variant that incorporatess residual skip connections to allow previous attention scores to pass through the entire network. It outperforms canonical transformers on a variety of tasks and datasets, including masked language modeling (MLM), GLUE, and SQuAD.

Implementation Notes

• Just like torch.nn.TransformerEncoder, the RealFormerEncoder does not include any embedding layers. It is recommended that you implemenet positional encoding schemes (e.g. sinusodial tables, learnable embeddings) as needed.
• The authors mention that RealFormer layers can be used as drop-in replacements for any transformer model, whether they be autoencoding (encoders) or auto-regressive (decoders). We closely follow the flow of the paper and include only an encoder version of the implementation for now.

Resources

• Original Paper