https://github.com/grypesc/seed

ICLR2024 paper on Continual Learning
https://github.com/grypesc/seed

class-incremental class-incremental-learning computer-vision continual-learning facil machine-learning

Last synced: 10 days ago
JSON representation

ICLR2024 paper on Continual Learning

• Host: GitHub
• URL: https://github.com/grypesc/seed
• Owner: grypesc
• License: mit
• Created: 2023-12-18T11:37:48.000Z (11 months ago)
• Default Branch: main
• Last Pushed: 2024-04-21T19:44:38.000Z (7 months ago)
• Last Synced: 2024-04-22T00:56:40.452Z (7 months ago)
• Topics: class-incremental, class-incremental-learning, computer-vision, continual-learning, facil, machine-learning
• Language: Python
• Homepage: https://arxiv.org/abs/2401.10191
• Size: 427 KB
• Stars: 19
• Watchers: 2
• Forks: 3
• Open Issues: 1
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

        
# Divide and not forget: Ensemble of selectively trained experts in Continual Learning: ICLR2024 (Main track)

https://arxiv.org/abs/2401.10191  

https://openreview.net/forum?id=sSyytcewxe  

![image](inference.jpg?raw=true "inference")

This repository contains code for the SEED paper published at the main track of ICLR2024. It is based on FACIL (https://github.com/mmasana/FACIL) benchmark.

To reproduce results run one of provided scripts. 

Setup environment according to readme of FACIL.

Run SEED on CIFAR100 10 tasks, 10 classes each:

```bash

bash cifar10x10.sh

```

Run SEED on CIFAR100 20 tasks, 5 classes each:

```bash

bash cifar20x5.sh

```

Run SEED on CIFAR100 50 tasks, 2 classes each:

```bash

bash cifar50x2.sh

```

To lower the number of parameters as in Tab.5 use ```--network resnet 20 --shared 2```. You can also add parameter pruning as in DER.

To reproduce results for ImageNet Subset download ImageNet subset from https://www.kaggle.com/datasets/arjunashok33/imagenet-subset-for-inc-learn and unzip it in ```../data``` directory.

```bash

bash imagenet10x10.sh

```

To reproduce results for DomainNet download it from http://ai.bu.edu/M3SDA/ and put it in ```../data/domainnet``` directory (unzip it).

Run SEED on DomainNet 36 tasks of different domains, 5 classes each:

```bash

bash domainnet36x5.sh

```

You can add ```--extra-aug fetril``` flag to enable better augmentations.

If you would like to cooperate on improving the method, please contact me via LinkedIn or Facebook, I have several ideas.

If you find this work useful, please consider citing it:

```

@inproceedings{rypesc2023divide,

  title={Divide and not forget: Ensemble of selectively trained experts in Continual Learning},

  author={Rype{\'s}{\'c}, Grzegorz and Cygert, Sebastian and Khan, Valeriya and Trzcinski, Tomasz and Zieli{\'n}ski, Bartosz Micha{\l} and Twardowski, Bart{\l}omiej},

  booktitle={The Twelfth International Conference on Learning Representations},

  year={2023}

}

   ```