InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning

This is the official implementation of our CVPR 2024 paper "Interference-Free Low-Rank Adaptation for Continual Learning". In this paper, we propose a new parameter-efficient continual learning method called interference-free low-rank adaptation (InfLoRA).

Introduction

Continual learning requires the model to learn multiple tasks sequentially. In continual learning, the model should possess the ability to maintain its performance on old tasks (stability) and the ability to adapt to new tasks continuously (plasticity). Recently, parameter-efficient fine-tuning (PEFT), which involves freezing a pre-trained model and injecting a small number of learnable parameters to adapt to downstream tasks, has gained increasing popularity in continual learning. Although existing continual learning methods based on PEFT have demonstrated superior performance compared to those not based on PEFT, most of them do not consider how to eliminate the interference of the new task on the old tasks, which inhibits the model from making a good trade-off between stability and plasticity. In this work, we propose a new PEFT method, called interference-free low-rank adaptation (InfLoRA), for continual learning. InfLoRA injects a small number of parameters to reparameterize the pre-trained weights and shows that fine-tuning these injected parameters is equivalent to fine-tuning the pre-trained weights within a subspace. Furthermore, InfLoRA designs this subspace to eliminate the interference of the new task on the old tasks, making a good trade-off between stability and plasticity. Experimental results show that InfLoRA outperforms existing state-of-the-art continual learning methods on multiple datasets.

Requisite

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

• python = 3.8
• torch = 1.10.0
• torchvision = 0.11.1
• timm = 0.6.7

I think the code can run under other versions of the environment, but I haven't tried.

Dataset preparation

• Create a folder data/
• CIFAR 100: should automatically be downloaded
• ImageNet-R: download dataset from https://people.eecs.berkeley.edu/~hendrycks/imagenet-r.tar. After unzipping, place it into data/ folder
• DomainNet: download from http://ai.bu.edu/M3SDA/, place it into data/ folder

Training

All commands should be run under the project root directory. Currently, the code has been validated on 1 A6000 GPU (48G) and 4 2080ti GPUs (11G).

CIFAR100:

For InfLoRA

python main.py --device your_device --config configs/cifar100_inflora.json 

For InfLoRA-b5

python main.py --device your_device --config configs/cifar100_inflorab5.json 

For InfLoRA+CA

python main.py --device your_device --config configs/cifar100_infloraca.json 

ImageNet-R (10 Tasks):

For InfLoRA

python main.py --device your_device --config configs/mimg10_inflora.json 

For InfLoRA-b5

python main.py --device your_device --config configs/mimg10_inflorab5.json 

For InfLoRA+CA

python main.py --device your_device --config configs/mimg10_infloraca.json 

ImageNet-R (20 Tasks):

For InfLoRA

python main.py --device your_device --config configs/mimg20_inflora.json 

For InfLoRA-b5

python main.py --device your_device --config configs/mimg20_inflorab5.json 

ImageNet-R (5 Tasks):

For InfLoRA

python main.py --device your_device --config configs/mimg5_inflora.json 

For InfLoRA-b5

python main.py --device your_device --config configs/mimg5_inflorab5.json 

DomainNet:

For InfLoRA

python main.py --device your_device --config configs/domainnet_inflora.json 

For InfLoRA-b5

python main.py --device your_device --config configs/domainnet_inflorab5.json 

Citation

@inproceedings{liang2024inflora,
  title={InfLoRA: Interference-Free Low-Rank Adaptation for Continual Learning},
  author={Liang, Yan-Shuo and Li, Wu-Jun},
  booktitle={Proceedings of the IEEE/CVF Conference on Computer Vision and Pattern Recognition},
  pages={23638--23647},
  year={2024}
}

Acknoledgements

We thank the following repos providing helpful components/functions in our work.

• PyCIL
• S-Prompts