GR-LoRA: Gradient-Recycling Low-Rank Adaptation for Class-Incremental Learning

May 12, 2026 ยท View on GitHub

This repository contains the official implementation of our ICML 2026 paper, "GR-LoRA: Gradient-Recycling Low-Rank Adaptation for Class-Incremental Learning."

cover

1.Requisite

This code is implemented in PyTorch, and we perform the experiments under the following environment settings:

  • python = 3.10
  • torch = 2.1.2+cu121
  • torchvision = 0.16.2+cu121
  • timm = 0.6.12

The code has been tested on Linux Platform with a GPU (RTX4090).

2.Dataset

  • Create a folder data/
  • CIFAR 100: will be automatically downloaded by the code.
  • ImageNet-R: retrieve from Google Drive and unzipping, place it into data/ folder.
  • ImageNet-A: retrieve from Google Drive and unzipping, place it into data/ folder.
  • CUB: retrieve from Google Drive and unzipping , place it into data/ folder.

3.Training

The JSON configuration files in configs/ are preconfigured for 20-task scenarios. You can modify init_cls, increment and total_sessions parameters in configs/[dataset].json to configure different CIL settings.

  • CIFAR100: 
    python main.py --device '0' --config configs/cifar100.json
  • ImageNet-R: 
    python main.py --device '0' --config configs/imagenetr.json

4.Citation

@inproceedings{
    lin2026grlora,
    title={GR-LoRA: Gradient-Recycling Low-Rank Adaptation for Class-Incremental Learning},
    author={Yipeng Lin, Fengqiang Wan and Yang Yang},
    booktitle={Forty-third International Conference on Machine Learning},
    year={2026},
}

5.Reference

We appreciate the following repositories for their contributions of useful components and functions to our work.