This is a public repository for the following paper:
On learning the geodesic path for incremental learning
CVPR 2021 (ORAL)

Paper: https://arxiv.org/abs/2104.08572

Dataset

The ImageNet-Subset dataset (100 classes) can be downloaded at: https://drive.google.com/drive/folders/1cuqI8yuqc8u1lN_N5CbFBxEdF_VSUNbG

Run

Run on ImageNet-subset with 50 classes as an initialization and 10 tasks:

python3 class_incremental_imagenet.py --dataset imagenet --datadir  {your imagenetsubset dir} --num_classes 100 --nb_cl_fg 50 --nb_cl 5 --nb_protos 20 --rs_ratio 0.0 --imprint_weights --less_forget --resume --lamda 10 --adapt_lamda

incremental_train_and_eval_LF.py is the file containing the distillation loss with our method.

Citation

@inproceedings{Christian2021MGeoCont,
author = {Simon, Christian and Koniusz, Piotr and  Harandi, Mehrtash},
title = {On Learning the Geodesic Path for Incremental Learning},
booktitle = {IEEE Computer Vision and Pattern Recognition},
year = {2021}
}

Thanks to LUCIR codebase https://github.com/hshustc/CVPR19_Incremental_Learning