Accurate Forgetting for Heterogeneous Federated Continual Learning

The implementation of AF-FCL.

Requirements

The needed libraries are in requirements.txt.

Dataset preparation:

All datasets can be automatically downloaded with torchvision.datasets

Experiments

To run on EMNIST-Letters, excute:

  python main.py --dataset EMNIST-Letters --data_split_file data_split/EMNIST_letters_split_cn8_tn6_cet2_cs2_s2571.pkl --num_glob_iters 60 --local_epochs 100 --lr 1e-4 --flow_lr 1e-4 --k_loss_flow 0.5 --k_flow_lastflow 0.4 --flow_explore_theta 0

To run on EMNIST-shuffle, excute:

  python main.py --dataset EMNIST-Letters-shuffle --data_split_file data_split/EMNIST_letters_shuffle_split_cn8_tn6_cet2_cs2_s2571.pkl --num_glob_iters 60 --local_epochs 100 --lr 1e-4 --flow_lr 1e-3 --k_loss_flow 0.05 --k_flow_lastflow 0.02 --flow_explore_theta 0.5

To run on EMNIST-noisy with M noisy clients, excute:

  python main.py --dataset EMNIST-Letters-malicious --data_split_file data_split/EMNIST_letters_split_cn8_tn6_cet2_cs2_s2571.pkl --num_glob_iters 60 --local_epochs 100  --lr 1e-4 --flow_lr 1e-3 --k_loss_flow 0.5 --k_flow_lastflow 0.1 --flow_explore_theta 0.5 --malicious_client_num $M

To run on MNIST-SVHN-FASHION, excute:

  python main.py --dataset MNIST-SVHN-FASHION --data_split_file data_split/EMNIST_letters_split_cn8_tn6_cet2_cs2_s2571.pkl --num_glob_iters 60 --local_epochs 100 --lr 1e-4 --flow_lr 1e-3 --k_loss_flow 0.1 --k_flow_lastflow 0 --flow_explore_theta 0 --fedprox_k 0.001

To run on CIFAR100, excute:

  python main.py --dataset CIFAR100 --data_split_file data_split/CIFAR100_split_cn10_tn4_cet20_s2571.pkl --num_glob_iters 40 --local_epochs 400 --lr 1e-3 --flow_lr 5e-3 --k_loss_flow 0.5 --k_flow_lastflow 0.1 --flow_explore_theta 0.1 --fedprox_k 0.001

Reference

The code structure is based on the code in FedCIL.

The normalizaing flow code refers to nflows.