NeurIPS25-P-M

[NeurIPS 2025] Official Implementation for Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning

Authors

Haomiao Qiu^1,2, Miao Zhang¹*, Ziyue Qiao²*, Liqiang Nie¹

¹ Harbin Institute of Technology (Shenzhen)
² Great Bay University
* Corresponding author

Links

• Paper: Paper Link
• Code Repository: GitHub

---

Updates

• [05/2025] Initial release

---

Introduction

We propose P&M, a continual learning framework that mitigates forgetting via optimal post-training model merging and task-vector perturbations, achieving sota performance. This repository provides the official implementation, train and evaluation scripts.

---

Installation

1. Clone the repository

git clone https://github.com/iLearn-Lab/NeurIPS-P-M_CL.git
cd NeurIPS25-P-M_CL

2. Create environment

python -m venv .venv
source .venv/bin/activate   # Linux / Mac
# .venv\Scripts\activate    # Windows

3. Install dependencies

pip install -r requirements.txt

---

4. Dataset preparation

Download the datasets and uncompress them:

• CIFAR-100: https://www.cs.toronto.edu/~kriz/cifar.html
• ImageNet-R: https://github.com/hendrycks/imagenet-r
• DomainNet: https://ai.bu.edu/M3SDA/

Rearrange the directory structure:

Directory structure for three datasets:

DATA_ROOT
    |- train
    |    |- class_folder_1
    |    |    |- image_file_1
    |    |    |- image_file_2
    |    |- class_folder_2
    |         |- image_file_2
    |         |- image_file_3
    |- val
         |- class_folder_1
         |    |- image_file_5
         |    |- image_file_6
         |- class_folder_2
              |- image_file_7
              |- image_file_8

We provide the scripts split_[dataset].py in the tools folder to rearange the directory structure. Please change the root_dir in each script to the path of the uncompressed dataset.

---

Usage

For three datasets: bash reproduce.sh

---

Citation

@article{qiu2025train,
  title={Train with Perturbation, Infer after Merging: A Two-Stage Framework for Continual Learning},
  author={Qiu, Haomiao and Zhang, Miao and Qiao, Ziyue and Nie, Liqiang},
  journal={arXiv preprint arXiv:2505.22389},
  year={2025}
}

---

Acknowledgement

• Thanks to our supervisor and collaborators for valuable support.
• The code is developed based on https://github.com/zugexiaodui/VPTinNSforCL! We sincerely thank the authors for open-sourcing their code.

---

License

This project is released under the Apache License 2.0.