ischlag/fast-weight-transformers

Official code repository of the paper Linear Transformers Are Secretly Fast Weight Programmers.

Linear Transformers Are Secretly Fast Weight Programmers

This repository contains the code accompanying the paper Linear Transformers Are Secretly Fast Weight Programmers which is published at ICML'21. It also contains the logs of all synthetic experiments.

Synthetic Experiments

Requirements

$ cat req.txt 
jupyter==1.0.0
pandas==1.0.1
seaborn==0.10.0
torch==1.6.0
matplotlib==3.1.3
numpy==1.17.2

pip3 install -r req.txt

Rerun Experiments

Logs are provided in the synthetic/logs folder. The files in that folder are a result of running the following commands:

Setting 1 (capacity):

python3 main.py --begin=20 --end=600 --step=20 --attn_name=softmax --update_rule=sum
python3 main.py --begin=20 --end=600 --step=20 --attn_name=linear --update_rule=sum
python3 main.py --begin=20 --end=600 --step=20 --attn_name=dpfp --attn_arg=1 --update_rule=sum
python3 main.py --begin=20 --end=600 --step=20 --attn_name=dpfp --attn_arg=2 --update_rule=sum

python3 main.py --begin=20 --end=600 --step=20 --attn_name=dpfp --attn_arg=3 --update_rule=sum
python3 main.py --begin=20 --end=600 --step=20 --attn_name=favor --attn_arg=64 --update_rule=sum
python3 main.py --begin=20 --end=600 --step=20 --attn_name=favor --attn_arg=128 --update_rule=sum
python3 main.py --begin=20 --end=600 --step=20 --attn_name=favor --attn_arg=512 --update_rule=sum

Setting 2 (update rule):

python3 main.py --begin=20 --end=200 --step=20 --attn_name=dpfp --attn_arg=1 --update_rule=sum --replace
python3 main.py --begin=20 --end=200 --step=20 --attn_name=dpfp --attn_arg=1 --update_rule=ours --replace
python3 main.py --begin=20 --end=200 --step=20 --attn_name=tanh --update_rule=fwm --replace
python3 main.py --begin=20 --end=200 --step=20 --attn_name=dpfp --attn_arg=1 --update_rule=fwm --replace

python3 main.py --begin=20 --end=200 --step=20 --attn_name=dpfp --attn_arg=2 --update_rule=ours --replace
python3 main.py --begin=20 --end=200 --step=20 --attn_name=linear --update_rule=ours --replace
python3 main.py --begin=20 --end=200 --step=20 --attn_name=favor --attn_arg=64 --update_rule=ours --replace
python3 main.py --begin=20 --end=200 --step=20 --attn_name=favor --attn_arg=128 --update_rule=ours --replace

Generate figures from the logs using the following notebooks:

synthetic/setting1_generate_figure.ipynb
synthetic/setting2_generate_figure.ipynb

Language Modelling & Machine Translation

The toolkit and scripts for language modeling experiments can be found at IDSIA/lmtool-fwms.

For machine translation experiments, we ported the different attention functions implemented in the language modeling toolkit to the multi-head attention implementation in FAIRSEQ.

Citation

@inproceedings{schlag2021linear,
      title={Linear Transformers Are Secretly Fast Weight Programmers}, 
      author={Imanol Schlag and Kazuki Irie and J\"urgen Schmidhuber},
      booktitle={Proc. Int. Conf. on Machine Learning (ICML)},
      address = {Virtual only},
      month = jul,
      year={2021}
}