Split-Mix Federated Learning for Model Customization

Official PyTorch Code for Paper: "Efficient Split-Mix Federated Learning for On-Demand and In-Situ Customization" Junyuan Hong, Haotao Wang, Zhangyang Wang and Jiayu Zhou, ICLR 2022.

paper / code / slides / blog

TL;DR Split-Mix is an efficient and flexible Federated Learning algorithm allowing customizing model sizes and robustness during both training and testing time.

Overview

Federated learning (FL) provides a distributed learning framework for multiple participants to collaborate learning without sharing raw data. In many practical FL scenarios, participants have heterogeneous resources due to disparities in hardware and inference dynamics that require quickly loading models of different sizes and levels of robustness. The heterogeneity and dynamics together impose significant challenges to existing FL approaches and thus greatly limit FL's applicability. In this paper, we propose a novel Split-Mix FL strategy for heterogeneous participants that, once training is done, provides in-situ customization of model sizes and robustness. Specifically, we achieve customization by learning a set of base sub-networks of different sizes and robustness levels, which are later aggregated on-demand according to inference requirements. This split-mix strategy achieves customization with high efficiency in communication, storage, and inference. Extensive experiments demonstrate that our method provides better in-situ customization than the existing heterogeneous-architecture FL methods.

Usage

Preparation:

Package dependencies: Use conda env create -f environment.yml to create a conda env and activate by conda activate splitmix. Major dependencies include pytorch, torchvision, wandb, numpy, thop for model size customization, and advertorch for adversarial training.
Data: Set up your paths to data in utils/config.py. Refer to FedBN for details of Digits and DomainNet datasets.
- Cifar10: Auto download by python -m utils.data_utils --download=Cifar10.
- Digits: Download the FedBN Digits zip file to DATA_PATHS['Digits']' defined in utils.config. Unzip all files.
- DomainNet: Download the FedBN DomainNet split file to DATA_PATHS['DomainNetPathList'] defined in utils.config. Download Clipart, Infograph, Painting, Quickdraw, Real, Sketch, put under DATA_PATH['DomainNet'] directory. Unzip all files.

Train and test:

Customize model sizes: Set --data to be one of Digits, DomainNet, Cifar10.

# SplitMix
python fed_splitmix.py --data Digits --no_track_stat # train
python fed_splitmix.py --data Digits --no_track_stat --test --test_slim_ratio=0.25 # test
# FedAvg
python fedavg.py --data Digits --width_scale=0.125 --no_track_stat # train
python fedavg.py --data Digits --width_scale=0.125 --no_track_stat --test # test
# SHeteroFL
python fed_hfl.py --data Digits --no_track_stat # train
python fed_hfl.py --data Digits --no_track_stat --test --test_slim_ratio=0.25 # test

Customize robustness and model sizes (during training and testing)

# SplitMix + DAT
python fed_splitmix.py --adv_lmbd=0.5
python fed_splitmix.py --adv_lmbd=0.5 --test --test_noise=LinfPGD --test_adv_lmbd=0.1  # robust test
python fed_splitmix.py --adv_lmbd=0.5 --test --test_noise=none --test_adv_lmbd=0.1  # standard test
# individual FedAvg + AT
python fedavg.py --adv_lmbd=0.5
python fed_splitmix.py --adv_lmbd=0.5 --test --test_noise=LinfPGD  # robust test
python fed_splitmix.py --adv_lmbd=0.5 --test --test_noise=none  # standard test

For fed_splitmix, you may use --verbose=1 to print more information and --val_ens_only to speed up evaluation by only evaluating widest models.

Reproduce Results

We provide detailed parameter settings in sweeps. Check sweeps/Slimmable.md for experiments of customizing model sizes. Check sweeps/AT.md for customizing robustness and joint customization of robustness and model sizes. Example use of sweep:

~> wandb sweep sweeps/fed_niid/digits.yaml
wandb: Creating sweep from: sweeps/fed_niid/digits.yaml
wandb: Created sweep with ID: <ID>
wandb: View sweep at: https://wandb.ai/<unique ID>
wandb: Run sweep agent with: wandb agent <unique ID>
~> export CUDA_VISIBLE_DEVICES=0  # choose desired GPU
~> wandb agent <unique ID>

Pre-trained models will be shared upon request.

Citation

@inproceedings{hong2022efficient,
  title={Efficient Split-Mix Federated Learning for On-Demand and In-Situ Customization},
  author={Hong, Junyuan and Wang, Haotao and Wang, Zhangyang and Zhou, Jiayu},
  booktitle={ICLR},
  year={2022}
}

Acknowledgement

This material is based in part upon work supported by the National Institute of Aging 1RF1AG072449, Office of Naval Research N00014-20-1-2382, National Science Foundation under Grant IIS-1749940. Z. W. is supported by the U.S. Army Research Laboratory Cooperative Research Agreement W911NF17-2-0196 (IOBT REIGN).

illidanlab/SplitMix