Helpful or Harmful: Inter-Task Association in Continual Learning

The official code for Helpful or Harmful: Inter-Task Association in Continual Learning

Author: Hyundong Jin and Eunwoo Kim

In Proc. of the European Conference on Computer Vision (ECCV), 2022

Leaderboard of ImageNet with Five Fine grained datasets

Note that the leaderboard compares methods with network expansion, while $H^{2}$ is a method without expanding the network.

Run the code

This repository currently supports the Split CIFAR-10 experiment in the original paper.

You can change the hyper-parmeters in the corresponding file (config/CONFIG.py) if needed.

python3 main.py

The accuracy of each task is the average of 10 independent runs.

T# represents #-th task.

	T1	T2	T3	T4	T5	Avg acc
$H^{2}$	98.82	94.13	96.36	98.90	98.27	97.30

Requirements

Please, find a list with required packages and versions in requirements.txt

Code list

config
  ㄴ CONFIG.py
data_loader
  ㄴ split_cifar10_data.py
model
  ㄴ resnet18.py
src 
  ㄴ main.py
  ㄴ manager.py
  ㄴ pruner.py

Reference

If you find our code useful for your research, please cite our paper.

@inproceedings{jin2022helpful,
  title={Helpful or Harmful: Inter-task Association in Continual Learning},
  author={Hyundong Jin and Eunwoo Kim},
  booktitle={European Conference on Computer Vision},
  pages={519--535},
  year={2022},
  organization={Springer} }