Prompt based continual learning with OOD Score based replay buffer

This repository contains the implementation of KSC2022 paper: Prompt based continual learning with OOD Score based replay buffer

Environment

The system I used and tested in

• Ubuntu 20.04.4 LTS
• Slurm 21.08.1
• NVIDIA GeForce RTX 3090
• Python 3.8

Usage

First, clone the repository locally:

git clone https://github.com/TY-LEE-KR/Capstone_Design_2.git
cd Capstone_Designe_2

Then, install the packages below:

pytorch==1.12.1
torchvision==0.13.1
timm==0.6.7
pillow==9.2.0
matplotlib==3.5.3

These packages can be installed easily by

pip install -r requirements.txt

Data preparation

If you already have CIFAR-100 datasets, pass your dataset path to --data-path.

If the dataset isn't ready, change the download argument in continual_dataloader.py as follows

datasets.CIFAR100(download=True)

Train

To train a model on CIFAR-100, set the --data-path (path to dataset) and --output-dir (result logging directory) and other options in train.sh properly and run in Slurm system.

Training

To train a model on CIFAR-100 via command line:

python main.py --model vit_base_patch16_224 --batch-size 24 --data-path /local_datasets/ --output_dir ./output --epochs 5

Or you can simply use 'train.sh' if you use slurm system.

sbatch train.sh

Evaluation

To evaluate a trained model:

python main.py --eval

License

This repository is released under the Apache 2.0 license as found in the LICENSE file.