Continual Learning Datasets

This repository contains frequently used CL datasets and simple train and evaluation code.

Environment

The system I used and tested in

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

Datasets

The available datasets are as follows:

Task & Class Incremental

• CIFAR-100
• CIFAR-10
• MNIST
• FashionMNIST
• SVHN
• NotMNIST
• Flower102
• Stanford Cars196
• CUB200
• Indoor Scene67
• TinyImagenet
• Imagenet-R
• MNISTM
• Synthetic Digit

Domain Incremental

• CORe50
• DomainNet
• PermutedMNIST
• iDigits (MNIST, SVHN, MNISTM, Synthetic Digit)

Usage

First, clone the repository locally:

git clone https://github.com/JH-LEE-KR/Continual-datasets.git
cd Continual-datasets

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

Training

It can be used in various scenarios by changing --dataset and --num_tasks as shown below (default: class incremental):

Split-CIFAR100 with 10 tasks

python main.py --dataset Split-CIFAR100 --num_tasks 10

Split-Imagenet-R with 10 tasks

python main.py --dataset Split-Imagenet-R --num_tasks 10

5 datasets, 5 tasks with MNIST, Fashion-MNIST, NotMNIST, CIFAR10, SVHN

python main.py --dataset 5-datasets --num_tasks 5

Sequence of datasets (CUB200,TinyImagenet,Scene67,Cars196,Flower102) with 5 tasks

python main.py --dataset CUB200,TinyImagenet,Scene67,Cars196,Flower102 --num_tasks 5

You can customize scenario, as below: Split-TinyImagenet with 10 tasks

python main.py --dataset Split-TinyImagenet --num_tasks 10

Split-CUB200 with 10 tasks

python main.py --dataset Split-CUB200 --num_tasks 10

Sequence of datasets (CIFAR100,CUB200,TinyImagenet,Scene67,Cars196,Flower102,Imagenet-R) with 7 tasks

python main.py --dataset CIFAR100,CUB200,TinyImagenet,Scene67,Cars196,Flower102,Imagenet-R --num_tasks 7

Domain Incremental CORe50 with 7 tasks

python main.py --dataset CORe50 --num_tasks 7 --domain_inc --no_train_mask

Domain Incremental DomainNet with 6 tasks

python main.py --dataset DomainNet --num_tasks 6 --domain_inc --no_train_mask

Domain Incremental PermutedMNIST with 6 tasks

python main.py --dataset PermutedMNIST --num_tasks 6 --domain_inc --no_train_mask

Domain Incremental iDigits with 4 tasks

python main.py --dataset iDigits --num_tasks 4 --domain_inc --no_train_mask

Options

--train_mask, if using the class mask at training.
--no_train_mask, if domain incremental setting, not using the class mask at training.
--task_inc, if doing task incremental.
--domain_inc, if doing domain incremental.
--shuffle, shuffle the data order.
--label_shuffle, shuffle the label order when split the single dataset.

Also available in Slurm by changing options on train.sh