zaocan666/af-fcl

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.

af-fcl's Issues

同学你好，我在复现EMNIST-Letters数据集时出现了一些问题

运行的是：python main.py --dataset EMNIST-Letters --data_split_file ./data_split/EMNIST_letters_split_cn8_tn6_cet2_cs2_s2571.pkl --num_glob_iters 10 --local_epochs 15 --lr 1e-4 --flow_lr 1e-4 --k_loss_flow 0.5 --k_flow_lastflow 0.4 --flow_explore_theta 0
报错为：
Traceback (most recent call last):
File "main.py", line 99, in
main(args)
File "main.py", line 30, in main
run_job(args, args.seed)
File "main.py", line 25, in run_job
server = create_server_n_user(args, seed)
File "main.py", line 16, in create_server_n_user
server=FedPrecise(args, model, i)
File "/home/aiswjtu/yzy/AF-FCL/FLAlgorithms/servers/serverPreciseFCL.py", line 19, in init
self.data = get_dataset(args, args.dataset, args.datadir, args.data_split_file)
File "/home/aiswjtu/yzy/AF-FCL/utils/dataset.py", line 121, in get_dataset
testify_client_y_list(train_y_list, split_data['train_inds'], split_data['client_y_list'])
File "/home/aiswjtu/yzy/AF-FCL/utils/dataset.py", line 16, in testify_client_y_list
assert y_c_t_set==set(client_y_list[c_i][t_i])
AssertionError

同学你好，我在复现CIFAR100数据集出现了一些问题

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
跑出来的结果出现很多nan
pickle.csv

AF-FCL复现的bug：不存在.pkl文件

同学你好，我在AF-FCL中运行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命令时，出现了报错

Traceback (most recent call last): File "main.py", line 98, in <module> main(args) File "main.py", line 30, in main run_job(args, args.seed) File "main.py", line 25, in run_job server = create_server_n_user(args, seed) File "main.py", line 16, in create_server_n_user server=FedPrecise(args, model, i) File "/home/wyc/AF-FCL/FLAlgorithms/servers/serverPreciseFCL.py", line 19, in __init__ self.data = get_dataset(args, args.dataset, args.datadir, args.data_split_file) File "/home/wyc/AF-FCL/utils/dataset.py", line 118, in get_dataset with open(os.path.join(datadir, data_split_file), 'rb') as f: FileNotFoundError: [Errno 2] No such file or directory: 'datasets/PreciseFCL/data_split/EMNIST_letters_split_cn8_tn6_cet2_cs2_s2571.pkl'

在get_dataset函数读取.pkl文件时，找不到文件，请问这个文件时怎么生成的呢