Skip to content
/ FACIL Public

Framework for Analysis of Class-Incremental Learning with 12 state-of-the-art methods and 3 baselines.

License

Notifications You must be signed in to change notification settings

mmasana/FACIL

Folders and files

NameName
Last commit message
Last commit date

Latest commit

 

History

11 Commits
 
 
 
 
 
 
 
 
 
 
 
 
 
 

Repository files navigation

Framework for Analysis of Class-Incremental Learning


What is FACILKey FeaturesHow To UseApproachesDatasetsNetworksLicenseCite


What is FACIL

FACIL started as code for the paper:
Class-incremental learning: survey and performance evaluation
Marc Masana, Xialei Liu, Bartlomiej Twardowski, Mikel Menta, Andrew D. Bagdanov, Joost van de Weijer
(TPAMI) (arxiv)

It allows to reproduce the results in the paper as well as provide a (hopefully!) helpful framework to develop new methods for incremental learning and analyse existing ones. Our idea is to expand the available approaches and tools with the help of the community. To help FACIL grow, don't forget to star this github repository and share it to friends and coworkers!

Key Features

We provide a framework based on class-incremental learning. However, task-incremental learning is also fully supported. Experiments by default provide results on both task-aware and task-agnostic evaluation. Furthermore, if an experiment runs with one task on one dataset, results would be equivalent to 'common' supervised learning.

Setting task-ID at train time task-ID at test time # of tasks
class-incremental learning yes no ≥1
task-incremental learning yes yes ≥1
non-incremental supervised learning yes yes 1

Current available approaches include:

Finetuning • Freezing • Joint

LwF • iCaRL • EWC • PathInt • MAS • RWalk • EEIL • LwM • DMC • BiC • LUCIR • IL2M

How To Use

Clone this github repository:

git clone https://github.com/mmasana/FACIL.git
cd FACIL
Optionally, create an environment to run the code (click to expand).

Using a requirements file

The library requirements of the code are detailed in requirements.txt. You can install them using pip with:

python3 -m pip install -r requirements.txt

Using a conda environment

Development environment based on Conda distribution. All dependencies are in environment.yml file.

Create env

To create a new environment check out the repository and type:

conda env create --file environment.yml --name FACIL

Notice: set the appropriate version of your CUDA driver for cudatoolkit in environment.yml.

Environment activation/deactivation

conda activate FACIL
conda deactivate

To run the basic code:

python3 -u src/main_incremental.py

More options are explained in the src, including GridSearch usage. Also, more specific options on approaches, loggers, datasets and networks.

Scripts

We provide scripts to reproduce the specific scenarios proposed in Class-incremental learning: survey and performance evaluation:

  • CIFAR-100 (10 tasks) with ResNet-32 without exemplars
  • CIFAR-100 (10 tasks) with ResNet-32 with fixed and growing memory
  • MORE COMING SOON...

All scripts run 10 times to later calculate mean and standard deviation of the results. Check out all available in the scripts folder.

License

Please check the MIT license that is listed in this repository.

Cite

If you want to cite the framework feel free to use this preprint citation while we await publication:

@article{masana2022class,
  title={Class-Incremental Learning: Survey and Performance Evaluation on Image Classification},
  author={Masana, Marc and Liu, Xialei and Twardowski, Bartłomiej and Menta, Mikel and Bagdanov, Andrew D. and van de Weijer, Joost},
  journal={IEEE Transactions on Pattern Analysis and Machine Intelligence},
  doi={10.1109/TPAMI.2022.3213473},
  year={2023},
  volume={45},
  number={5},
  pages={5513-5533}}
}

The basis of FACIL is made possible thanks to Marc Masana, Xialei Liu, Bartlomiej Twardowski and Mikel Menta. Code structure is inspired by HAT. Feel free to contribute or propose new features by opening an issue!