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.
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 https://cs231n.stanford.edu/tiny-imagenet-200.zip
-
Unzip the file
unzip tiny-imagenet-200.zip
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.
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 implementCReFF
,Ditto
,FedAvg
,FedBABU
,FedNH
,FedPer
,FedProto
,FedRep
,FedROD
. Each file provides the concrete implementation of the correspondingserver
class andclient
class.
Helper functions, for example, generating non-iid data partition, can be found in src/utils.py
.
The code base is developed with extensive references to the following GitHub repos. Some code snippets are directly taken from the original implementation.
- FedBABU: https://github.com/jhoon-oh/FedBABU
- CReFF: https://github.com/shangxinyi/CReFF-FL
- FedROD: https://openreview.net/revisions?id=I1hQbx10Kxn
- Personalized Federated Learning Platform: https://github.com/TsingZ0/PFL-Non-IID
- FedProxL: https://github.com/litian96/FedProx
- NIID-Bench: https://github.com/Xtra-Computing/NIID-Bench
- FedProto: https://github.com/yuetan031/fedproto