This is a PyTorch implementation of the SGM algorithm from our TMLR-2024 paper. An arXiv version of our paper is available.
The conda environment that we used for SGM has been shared in the GitHub repository.
The yml file sgmenv.yml includes all the libraries. We have tested the code with the packages and versions
specified in the yml file. We recommend setting up a conda environment using the sgmenv.yml file:
conda env create -f sgmenv.yml
Python code to reproduce results presented in the paper. The folder imagenet_files contains the data ordering files. To use the code, simply run the following:
-
To train our proposed SGM model:
train_sgm.sh -
To train the vanilla model:
train_vanilla.sh
To get results for other configures, change the relevant arguments.
In our main experiments, we use ImageNet-1K pretrained ConvNeXt V2 Femto model (i.e.,convnextv2_femto_1k_224_ema.pt file in the repository).
You can also find pre-trained weights here.
If using this code, please cite our paper.
@article{
harun2024overcoming,
title={Overcoming the Stability Gap in Continual Learning},
author={Md Yousuf Harun and Christopher Kanan},
journal={Transactions on Machine Learning Research},
issn={2835-8856},
year={2024},
url={https://openreview.net/forum?id=o2wEfwUOma},
note={}
}