FedFA: Federated Feature Augmentation

This is the PyTorch implementation of our ICLR 2023 paper: FedFA: Federated Feature Augmentation by Tianfei Zhou and Ender Konukoglu, from ETH Zurich.

ChangeLog

• [2023-02-24] Code released with reimplementations of experiments

Preparation

Environment

• python/3.10.4
• pytorch/1.11.0
• cuda/11.3.1
• gcc/6.3.0
• numpy/1.22.4
• scipy/1.8.1
• opencv-python/4.5.5
• wandb (for experiment tracking)

Dataset

• Office-Caltech 10: https://drive.google.com/file/d/1gKd0M2Z6gdFwHntgsMTq5ZF5tQIg_bUD/view?usp=share_link
• Prostate MRI: https://drive.google.com/file/d/1GHy3hWAIU288Lr-NF80lydoJQmC7zlmk/view?usp=share_link
• DomainNet: TBD

Training

Bash training scripts for prostate_mri and office-caltech are provided in run_scripts. They are written for our cluster, but can be easily adapted for any kind of training environments. For scripts of other datasets, they are very similar with existing ones; or I will update them later.

Specify data path in Line 11 or Line 12 in config/prostate_mri/base.py:

self.DIR_ROOT = os.environ.get('TMPDIR')
self.DIR_DATA = os.path.join(self.DIR_ROOT, 'ProstateMRI')

Enter the directory of shell scripts, e.g., run_scripts/euler/prostate_mri:

cd run_scripts/euler/prostate_mri

Run training:

bash euler_train_fedfa.sh

Benchmark

Benchmark for ProstateMRI

Algorithm	Round	BIDMC	HK	I2CVB	BMC	RUNMC	UCL	Average	Log	Ckpt
FedAvg	500	82.60	91.59	89.55	82.00	90.44	86.27	87.08	log	ckpt
FedAvgM	500	83.00	91.56	88.27	82.29	90.39	84.82	86.72	log	ckpt
FedProx	500	83.61	88.31	89.45	80.93	88.13	86.36	86.13	log	ckpt
FedSAM	500	82.14	92.49	91.48	84.61	92.96	87.47	88.52	log	ckpt
FedDyn	500	78.09	84.24	81.13	76.61	82.46	75.80	79.72	log	ckpt
FedFA	500	89.18	92.64	90.08	89.16	90.91	87.71	89.95	log	ckpt

Benchmark for Office-Caltech 10

Algorithm	Round	Amazon	Caltech	DSLR	Webcam	Average	Log	Ckpt
FedAvg	400	84.38	64.44	75.00	91.53	78.84	log	ckpt
FedAvgM	400	80.21	65.78	75.00	91.53	78.13	log	ckpt
FedProx	400	83.85	63.56	78.12	94.92	80.11	log	ckpt
FedSAM	400	81.25	61.78	68.75	83.05	73.71	log	ckpt
FedDyn	400	79.17	60.89	65.62	88.14	73.45	log	ckpt
FedFA	400	82.81	67.11	90.62	93.22	83.44	log	ckpt

Citation

If you find this work useful, please consider citing:

@inproceedings{zhou2023fedfa,
  title={Fed{FA}: Federated Feature Augmentation},
  author={Tianfei Zhou and Ender Konukoglu},
  booktitle={The Eleventh International Conference on Learning Representations (ICLR)},
  year={2023}
}