ihaeyong/softnet-fscil

Winning SubNetwork (WSN), Soft-SubNetwork (SoftNet)

SoftNet-FSCIL

- Train SoftNet on Topic split of CIFAR100 (base session training)

>> ./train_soft_cifar.sh #gpu train

- Train and Evaluation of SoftNet on Topic split of CIFAR100 (few-shot incremental session traing and evaluation)

>> ./train_soft_cifar.sh #gpu test

- Train SoftNet + FSO on Topic split of CIFAR100 (base session training)

>> update soon

Update further soon...

3D Loss Landscapes

@inproceedings{
kang2023on,
title={On the Soft-Subnetwork for Few-Shot Class Incremental Learning},
author={Haeyong Kang and Jaehong Yoon and Sultan Rizky Hikmawan Madjid and Sung Ju Hwang and Chang D. Yoo},
booktitle={The Eleventh International Conference on Learning Representations },
year={2023},
url={https://openreview.net/forum?id=z57WK5lGeHd}
}

Extended Work: Forget-free Continual Learning with Soft-Winning SubNetworks

@misc{kang2023forgetfree,
      title={Forget-free Continual Learning with Soft-Winning SubNetworks}, 
      author={Haeyong Kang and Jaehong Yoon and Sultan Rizky Madjid and Sung Ju Hwang and Chang D. Yoo},
      year={2023},
      eprint={2303.14962},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

softnet-fscil's Issues

Data issue in Paper Table 7 .

Thank you for the inspiring paper in FSCIL!

When reading your paper, it seems that the experimental result of session 4 on CUB200 (Table 7 in the appendix, SoftNet, c = 93% ) is a little misleading. Is it a typo? Or it's the actual results?

Can you specify which python version is used please?

Can you specify which python version is used please?

GPU capacity for running code

Is it possible to run this code with a small GPU (RTX, 12Gb)

Code

Hello, Any idea when will the code be available?

Training on custom dataset

can u teach me how to train on custom dataset?

Continual Learning in LLM

Have you tried applying the method to any NLP classification tasks?

Can you provide training time and the GPU spec for running this project?

Trained model

Thank you for your great work.
Can you provide the trained model?
Looking forward to your reply.

Thanks for great work, regarding its generaliability

Thanks for the published code. Does it have the potential to adapt with MLM and MIM which has been demonstrated greatly recently in the generation?

inference time

How much take inference time?