This is the Pytorch Implementation for the paper:
LoRA-Based Continual Learning with Constraints on Critical Parameter Changes
Shimou Ling, Liang Zhang, Jiangwei Zhao, Lili Pan, Hongliang Li
Pattern Recognition (submitted)
Abstract: LoRA-based continual learning represents a promising avenue for leveraging pre-trained models in downstream continual learning tasks. Recent studies have shown that orthogonal LoRA tuning effectively mitigates forgetting. However, this work unveils that under orthogonal LoRA tuning, the critical parameters for pre-tasks still change notably after learning post-tasks.To address this problem, we directly propose freezing the most critical parameter matrices in the Vision Transformer (ViT) for pre-tasks before learning post-tasks. In addition, building on orthogonal LoRA tuning, we propose orthogonal LoRA composition (LoRAC) based on QR decomposition, which may further enhance the plasticity of our method. Elaborate ablation studies and extensive comparisons demonstrate the effectiveness of our proposed method. Our results indicate that our method achieves state-of-the-art (SOTA) performance on several well-known continual learning benchmarks. For instance, on the Split CIFAR-100 dataset, our method shows a 6.35% improvement in accuracy and a 3.24% reduction in forgetting compared to previous methods.
Depedencies
- Python >= 3.8 (Recommend to use Anaconda or Miniconda)
- Pytorch 2.0 or later. See Pytorch for install instructions.
- Linux (Ubuntu 20.04.3)
Installation
First, you can clone this repo using the command:
git clone https://github.com/learninginvision/LoRAC-IPCThen, you can create a virtual environment using conda, as follows:
conda env create -f environment.yaml
conda activate lorac-ipcWe provide source about the datasets we use in our experiment as below:
| Dataset | Dataset file |
|---|---|
| Split CIFAR-100 | CIFAR-100 |
| Split ImageNet-R | ImageNet-R |
| Split DomainNet | DomainNet |
| 5-datasets | automatically be downloaded |
In addition, if you do not want to modify the code, please store the dataset in the path: ../dataset/ and make sure that the folder structure under this path is as follows:
├── CIFAR100
│ └── cifar-100-python
├── 5-datasets
│ ├── MNIST
│ ├── FashionMNIST
│ ├── notMNIST
│ ├── cifar-10-python
│ └── svhn
└── ImageNet-R
│ └── imagenet-r
│ └── train
│ └── test
└── Domain
└── sketch
└── real
└── quickdraw
└── painting
└── infograph
└── clipart
Following previous work, the files used to divide DomainNet have been placed under the path ./continual_datasets/splits.
Run the following command to train the model sequentially:
bash ./training_scripts/train_{dataset}_vit_compilora_mask_{ptm}.sh{dataset}: The name of the dataset. (e.g.cifar100){ptm}: The name of the pre-trained model. (e.g.in21k)
Additionally, the hyperparameters for training the model can be adjusted in the ./training_scripts/*.sh file.
After training, you can get model checkpoints in the folder ./output/exp_name/checkpoints, where exp_name is the name of the experiment including different hyperparameters.
After completing training, the model's performance can be tested using the following command:
Without Task ID inference
bash ./testing_scripts/test_{dataset}_vit_compilora_{ptm}.shWith Task ID inference
bash ./testing_scripts/test_{dataset}_vit_compilora_tii_{ptm}.shThe result will be saved in the folder ./output/exp_name/eval_summary.txt.
Results on Split CIFAR-100 and Split ImageNet-R in rehearsal-free class-incremental learning.
This code is built upon thu-ml/HiDe-Prompt, and refers to the following four projects:
[1] L2P
[2] Dual-Prompt
We thank the authors for releasing their code.