Awesome

Preserving Linear Separability in Continual Learning by Backward Feature Projection

This repo provides the official implementation for the CVPR 2023 paper Preserving Linear Separability in Continual Learning by Backward Feature Projection (BFP).

Overview

To achieve a better stability-plasticity trade-off in continual learning, we propose Backward Feature Projection (BFP), a method for continual learning that allows the new features to change up to a learnable linear transformation of the old features. BFP preserves the linear separability of the old classes while allowing the emergence of new feature directions to accommodate new classes.

Overview image for BFP

Setup

To run the code, please first install PyTorch according to the setup of your system.

Then install the following dependencies using pip.

pip install numpy setproctitle six wandb onedrivedownloader av tqdm matplotlib scipy quadprog

Experiments

This repo use wandb to log experiments by default. Please log in your account by wandb login [KEY].

At the root folder of this repo, run the following commands to reproduce results for DER++ w/ BFP/FD/BFP-2 in the paper.

# Based on DER++
## CIFAR10
### Baseline DER++
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar10 --buffer_size 200 --alpha_bfp 0 --exp_suffix derpp_bfp0
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar10 --buffer_size 500 --alpha_bfp 0 --exp_suffix derpp_bfp0

### DER++ w/ different type of projection layer
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar10 --buffer_size 200 --alpha_bfp 1 --proj_type 0 --final_feat --pool_dim hw --exp_suffix derpp_fd
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar10 --buffer_size 200 --alpha_bfp 1 --proj_type 1 --final_feat --pool_dim hw --exp_suffix derpp_bfp
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar10 --buffer_size 200 --alpha_bfp 1 --proj_type 2 --final_feat --pool_dim hw --exp_suffix derpp_bfp2

python ./utils/main.py --project bfp --model bfp --dataset seq-cifar10 --buffer_size 500 --alpha_bfp 1 --proj_type 0 --final_feat --pool_dim hw --exp_suffix derpp_fd
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar10 --buffer_size 500 --alpha_bfp 1 --proj_type 1 --final_feat --pool_dim hw --exp_suffix derpp_bfp
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar10 --buffer_size 500 --alpha_bfp 1 --proj_type 2 --final_feat --pool_dim hw --exp_suffix derpp_bfp2

## CIFAR100
### Baseline DER++
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar100 --buffer_size 500 --alpha_bfp 0 --exp_suffix derpp_bfp0
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar100 --buffer_size 2000 --alpha_bfp 0 --exp_suffix derpp_bfp0

### DER++ w/ different type of projection layer
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar100 --buffer_size 500 --alpha_bfp 1 --proj_type 0 --final_feat --pool_dim hw --exp_suffix derpp_fd
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar100 --buffer_size 500 --alpha_bfp 1 --proj_type 1 --final_feat --pool_dim hw --exp_suffix derpp_bfp
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar100 --buffer_size 500 --alpha_bfp 1 --proj_type 2 --final_feat --pool_dim hw --exp_suffix derpp_bfp2

python ./utils/main.py --project bfp --model bfp --dataset seq-cifar100 --buffer_size 2000 --alpha_bfp 1 --proj_type 0 --final_feat --pool_dim hw --exp_suffix derpp_fd
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar100 --buffer_size 2000 --alpha_bfp 1 --proj_type 1 --final_feat --pool_dim hw --exp_suffix derpp_bfp
python ./utils/main.py --project bfp --model bfp --dataset seq-cifar100 --buffer_size 2000 --alpha_bfp 1 --proj_type 2 --final_feat --pool_dim hw --exp_suffix derpp_bfp2

## TinyImageNet
### Baseline DER++
python ./utils/main.py --project bfp --model bfp --dataset seq-tinyimg --buffer_size 4000 --alpha_bfp 0 --exp_suffix derpp_bfp0

### DER++ w/ different type of projection layer
python ./utils/main.py --project bfp --model bfp --dataset seq-tinyimg --buffer_size 4000 --alpha_bfp 1 --proj_type 0 --final_feat --pool_dim hw --exp_suffix derpp_fd
python ./utils/main.py --project bfp --model bfp --dataset seq-tinyimg --buffer_size 4000 --alpha_bfp 1 --proj_type 1 --final_feat --pool_dim hw --exp_suffix derpp_bfp
python ./utils/main.py --project bfp --model bfp --dataset seq-tinyimg --buffer_size 4000 --alpha_bfp 1 --proj_type 2 --final_feat --pool_dim hw --exp_suffix derpp_bfp2

Add --n_runs 5 to each command above to get results of 5 runs.
Add --base_method er to each command above to get results of ER w/ BFP/FD/BFP2.

Citation

If you find this repo useful, please cite our paper.

@article{gu2023bfp,
  title={Preserving Linear Separability in Continual Learning by Backward Feature Projection},
  author={Gu, Qiao and Shim, Dongsub and Shkurti, Florian},
  journal={arXiv preprint arXiv:2303.14595},
  year={2023}
}

Acknowledgements

This repo is based on Mammoth codebase.