This is implementation of PAGE: Equilibrate Personalization and Generalization in Federated Learning.

Please install the required packages. The code is compiled with Python 3.8 dependencies in a virtual environment via

pip install -r requirements.txt We construct our own FL enviroments ./my_envs for PAGE and Dap-FL to implement reinforcement learning argorithms. In order to use a standard API of OpenAI Gym, please remove ./my_env to gym/envs, append the contents of gym/envs/my_env/__init__.py to the end of gym/envs/__init__.py and delete gym/envs/my_env/__init__.py

Codes to run PAGE, Dap-FL as well as other baseline methods (FedAvg, FedProx, SCAFFOLD, FedDyn, pFedMe, Ditto, FedALA, Fed-ROD and FedRECON) with the Synthetic dataset, CIFAR100, Tiny-imagenet-200 and Shakespeare are given in ./PAGE&baseline(Dap-FL)/PAGE/, ./PAGE&baseline(Dap-FL)/Dap and ./Baselines/main, respectively.

Before running codes of PAGE and baselines with tiny-imagenet-200 and shakespeare, it is needed to run ./Data/run.sh to generate tiny-imagenet-200 dataset and run ./LEAF/shakespeare/preprocess.sh to generate shakespeare dataset, respeactively.

Please set the parameters in my_env/FL/synthetic_FL.py, my_env/FL/cifar100_FL.py, my_env/FL/imagenet200_FL.py, and my_env/FL/shakespeare_FL.py to control the FL setting in PAGE and Dap-FL with the Synthetic dataset, CIFAR100, Tiny-imagenet-200 and Shakespeare, respectively. Please set the parameters in ./Baselines/main/main_synthetic.py, ./Baselines/main/main_cifar100.py, ./Baselines/main/main_tiny-imagenet200.py, and ./Baselines/main/main_shakespeare.py to control other baselines' FL setting with the Synthetic dataset, CIFAR100, Tiny-imagenet-200 and Shakespeare, respectively.

Generate Shakespeare with non-iid used in PAGE to run

./LEAF/shakespeare/preprocess.sh -s niid --sf 1.0 -k 8000 -t sample -tf 0.8 --smplseed 1685439952 --spltseed 1685439964

Please reference ./LEAF/README.md to know more about the details of generating shakespeare dataset.

CIFAR100 IID, 100 partitions, balanced data, local training and testing sets on a 7:3 scale.

data_obj = ImagenetObjectCrop_noniid(dataset='CIFAR100', n_client=100, rule='homo', unbalanced_sgm=0, split_ratio=0.7, test_client_number=0)

CIFAR100 Dirichlet (0.3), 100 partitions, balanced data, local training and testing sets on a 7:3 scale.

data_obj = ImagenetObjectCrop_noniid(dataset='CIFAR100', n_client=100, rule='hetero', rule_arg=0.3, unlalanced_sgm=0, split_ratio=0.7, test_client_number=0)

CIFAR100 Dirichlet (0.3), 100 partitions, unbalanced (0.1) data, local training and testing sets on a 7:3 scale.

data_obj = ImagenetObjectCrop_noniid(dataset='CIFAR100', n_client=100, rule='hetero', ule_arg=0.3, unlalanced_sgm=0.1, split_ratio=0.7, test_client_number=0)

Generate Tiny-imagenet-200 to run

./Data/run.sh

Tiny-imagenet-200 IID, 100 partitions, balanced data, local training and testing sets on a 7:3 scale.

data_obj = ImagenetObjectCrop_noniid(dataset='Imagenet200', n_client=100, rule='homo', unbalanced_sgm=0, split_ratio=0.7, test_client_number=0)

Tiny-imagenet-200 Dirichlet (0.3), 100 partitions, balanced data, local training and testing sets on a 7:3 scale.

data_obj = ImagenetObjectCrop_noniid(dataset='Imagenet200', n_client=100, rule='hetero', rule_arg=0.3, unlalanced_sgm=0, split_ratio=0.7, test_client_number=0)

Tiny-imagenet-200 Dirichlet (0.3), 100 partitions, unbalanced (0.1) data, local training and testing sets on a 7:3 scale.

data_obj = ImagenetObjectCrop_noniid(dataset='Imagenet200', n_client=100, rule='hetero', ule_arg=0.3, unlalanced_sgm=0.1, split_ratio=0.7, test_client_number=0)

Synthetic dataset Non-IID, 100 partitions, used in PAGE.

data_obj = DatasetSynthetic(alpha=0, beta=1, theta=0, iid_sol=True, iid_data=False, n_dim=30, n_clnt=100, n_cls=30, avg_data=375, split_ratio=0.7, split_ratio_global=0.8,name_prefix='syn_alpha-0_beta-1_theta0')