Awesome

Class-Incremental Learning by Knowledge Distillation with Adaptive Feature Consolidation

1. Code Dependencies

Install required packages:

conda env create --file environment.yaml

Swith to new environment:

conda activate afc

2. Experiments

To reproduce Table1 with 50 steps on CIFAR100 with three different class orders LSC with CNN:

python3 -minclearn --options options/AFC/AFC_cnn_cifar100.yaml options/data/cifar100_3orders.yaml \
    --initial-increment 50 --increment 1 --fixed-memory \
    --device <GPU_ID> --label AFC_cnn_cifar100_50steps \
    --data-path <PATH/TO/DATA>

LSC with NME:

python3 -minclearn --options options/AFC/AFC_nme_cifar100.yaml options/data/cifar100_3orders.yaml \
    --initial-increment 50 --increment 1 --fixed-memory \
    --device <GPU_ID> --label AFC_nme_cifar100_50steps \
    --data-path <PATH/TO/DATA>

Likewise, for ImageNet100 (Table 2):

python3 -minclearn --options options/AFC/AFC_cnn_imagenet100.yaml options/data/imagenet100_1order.yaml \
    --initial-increment 50 --increment 1 --fixed-memory \
    --device <GPU_ID> --label AFC_cnn_imagenet100_50steps \
    --data-path <PATH/TO/DATA>

And ImageNet1000 (Table 2):

python3 -minclearn --options options/AFC/AFC_cnn_imagenet1000.yaml options/data/imagenet1000_1order.yaml \
    --initial-increment 500 --increment 50 --fixed-memory --memory-size 20000 \
    --device <GPU_ID> --label AFC_cnn_imagenet1000_10steps \
    --data-path <PATH/TO/DATA>

Citation

If you use this code for your research, please cite our paper.

@inproceedings{Kang2022afc,
  author = {Kang, Minsoo and Park, Jaeyoo and Han, Bohyung},
  booktitle = {CVPR},
  title = "{Class-Incremental Learning by Knowledge Distillation with Adaptive Feature Consolidation}",
  year = {2022}
  }

Acknowledgements

This repository is developed based on PODNet