FedNH

This repo provides an implementation of FedNH proposed in for Tackling Data Heterogeneity in Federated Learning with Class Prototypes, which is accepted by AAAI2023. In companion, we also provide our implementation of benchmark algorithms.

Prepare Dataset

Please create a folder data under the root directory.

mkdir ~/data

• Cifar10, Cifar100: No extra steps are required.

• TinyImageNet

• Download the dataset cd ~/data && wget http://cs231n.stanford.edu/tiny-imagenet-200.zip

• Unzip the file unzip tiny-imagenet-200.zip

Run scripts

We prepared a python file /experiments/gen_script.py to generate bash commands to run experiments.

To reproduce the results for Cifar10/Cifar100, just set the variable purpose to Cifar in the gen_script.py file. Similarly, set purpose to TinyImageNet to run experiments for TinyImageNet.

gen_script.py will create a set of bash files named as [method]_dir.sh. Then use, for example, bash FedAvg.sh to run experiments.

We include a set of bash files to run experiments on Cifar in this submission.

Organization of the code

The core code can be found at src/flbase/. Our framework builds upon three abstrac classes server, clients, and model. And their concrete implementations can be found in models directory and the startegies directory, respectively.

• src/flbase/models: We implemented or borrowed the implementation of (1) Convolution Neural Network and (2) Resnet18.
• src/flbase/strategies: We implement CReFF, Ditto, FedAvg, FedBABU, FedNH, FedPer, FedProto, FedRep, FedROD. Each file provides the concrete implementation of the corresponding server class and client class.

Helper functions, for example, generating non-iid data partition, can be found in src/utils.py.

Credits

The code base is developed with extensive references to the following GitHub repos. Some code snippets are directly taken from the original implementation.

1. FedBABU: https://github.com/jhoon-oh/FedBABU
2. CReFF: https://github.com/shangxinyi/CReFF-FL
3. FedROD: https://openreview.net/revisions?id=I1hQbx10Kxn
4. Personalized Federated Learning Platform: https://github.com/TsingZ0/PFL-Non-IID
5. FedProxL: https://github.com/litian96/FedProx
6. NIID-Bench: https://github.com/Xtra-Computing/NIID-Bench
7. FedProto: https://github.com/yuetan031/fedproto