/cl-as-seq

Official code for NeurIPS 2023 paper "Recasting Continual Learning as Sequence Modeling"

Primary LanguagePython

Recasting Continual Learning as Sequence Modeling

This repository contains the code for our NeurIPS 2023 paper titled Recasting Continual Learning as Sequence Modeling. For a brief overview of the paper, please check this tweet.

Poster

Requirements

  • Python 3.10
  • Pip packages:
pip install -r requirements.txt

Usage

The basic usage of the training script is as follows:

python train.py -mc [model config] -dc [data config] -o [override options] -l [log directory]

In commands.sh, we provide the commands used to train the models in the paper.

Downloading Datasets

All datasets except MS-Celeb-1M are downloaded automatically by the code.

MS-Celeb-1M

Use BitTorrent to download the dataset from Academic Torrents.

transmission-cli https://academictorrents.com/download/9e67eb7cc23c9417f39778a8e06cca5e26196a97.torrent -w data

Citation

@inproceedings{Lee2023Recasting,
  author    = {Soochan Lee and Jaehyeon Son and Gunhee Kim},
  title     = {Recasting Continual Learning as Sequence Modeling},
  booktitle = {NeurIPS},
  year      = {2023},
}