Official PyTorch implementation of our ICLR 2025 paper for general continual learning "Advancing Prompt-Based Methods for Replay-Independent General Continual Learning". (ArXiv/Camera-Ready version is coming soon!)

Our proposed MISA（Mask and Initial-Session Adaptation) consists of the forgetting-aware initial session adaptation and the non-parametric logit mask to facilitate general continual learning, as presented in the following figure:

How to run MISA?

Build the conda environment

Please make sure that necessary packages in the environment.yml file are available.

Two stage of MISA

Initial session adaptation

To warm up the prompt parameters by our initial session adaptation and our proposed forgetting-aware minimization, please run:

. scripts/misa_fam.sh

The warmed-up prompts will be stored in the pretrained_prompt/ folder.

We provide the pretrained prompt parameters in the pretrained_prompt/ folder to faciliate the reproduction of our results.

Training on downstram dataset and test

To test different methods with different datasets, simply run the corresponding script with the specific dataset entry in the file:

. scripts/misa.sh

Acknowledgment

This implementation is developed based on the source code of MVP.

This work was supported in part by the Institute of Information & Communications Technology Planning & Evaluation (IITP) grant funded by the Korean Government (MSIT) (No. RS-2024-00457882, National AI Research Lab Project).

CITATION

If you find our codes or paper useful, please consider giving us a star or citing our work.

@inproceedings{kang2025advancing,
title={Advancing Prompt-Based Methods for Replay-Independent General Continual Learning},
author={Zhiqi KANG and Liyuan Wang and Xingxing Zhang and Karteek Alahari},
booktitle={The Thirteenth International Conference on Learning Representations},
year={2025}
}