FLGo-Bench

Produce results of federated algorithms on various benchmarks

Citation

Please cite our paper in your publications if this code helps your research.

@misc{wang2023flgo,
      title={FLGo: A Fully Customizable Federated Learning Platform}, 
      author={Zheng Wang and Xiaoliang Fan and Zhaopeng Peng and Xueheng Li and Ziqi Yang and Mingkuan Feng and Zhicheng Yang and Xiao Liu and Cheng Wang},
      year={2023},
      eprint={2306.12079},
      archivePrefix={arXiv},
      primaryClass={cs.LG}
}

Usage

• Tuning Command

python tune.py --task TASKNAME --algorithm ALGORITHM --config CONFIG_PATH --gpu GPUids 

• Running Command

python run.py --task TASKNAME --algorithm ALGORITHM --config CONFIG_PATH --gpu GPUids 

• Optional Args

Name	Type	Desc.
model	str	the file name in the dictionary model/ that denotes a legal model in FLGo
load_mode	str	be one of ['', 'mmap', 'mem'], which respectively denotes DefaultIO, MemmapIO, and InMemory Dataset Loading
max_pdev	int	the maximum number of processes on each gpu device
available_interval	int	the time interval (s) to check whether a device is available
put_interval	int	the time interval (s) to put one process into device
seq	bool	whether to run each process in sequencial
train_parallel	int	the number of parallel client local training processes, default is 0
test_parallel	bool	whether to use data parallel when evaluating model
use_cache	bool	whether to use the disk to dynamically cache clients' states

• Example

# Tuning FedAvg on MNIST-IID with GPU 0 and 1
python tune.py --task mnist_iid_c100 --algorithm fedavg --config ./config/general.yml --gpu 0 1

# Runing FedAvg on MNIST-IID with GPU 0 and 1
python run.py --task mnist_iid_c100 --algorithm fedavg --config ./config/general.yml --gpu 0 1

Analysis

# Show tuning Result
python get_tune_res.py --task TASK --algorithm ALGORITHM --model MODEL --config CONFIG_PATH 

# Show Running Result
python get_run_res.py --task TASK --algorithm ALGORITHM --model MODEL --config CONFIG_PATH 

Algorithmic Configuration

We search the algorithmic hyper-parameter for each algorihtm according to the table below

Algorithm	Hyper-Parameter
fedavg	-
fedprox	μ ∈ [0.0001, 0.001, 0.01, 0.1, 1.0]
scaffold	η = 1.0
feddyn	α ∈ [0.001, 0.01, 0.03, 0.1]
moon	μ ∈ [0.1, 1.0, 5.0, 10.0], τ=0.5

Remark: To specify the search space of the hyper-parameters, add algo_para: [V1, V2, ...] in the corresponding config file.

Experimental Results

Nevigation

• CIFAR10
• CIFAR100
• TinyImageNet
• MNIST
• FEMNIST
• AgNews
• Office-Caltech10
• DomainNet
• PACS
• Digits
• ProstateMRI
• Camelyon17
• Fundus
• EndoPolyp
• SpeechCommand
• Shakespeare

CIFAR10

100 Clients

iid	dir5.0	dir2.0	dir1.0	dir0.1

Configuration

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
proportion: 0.2
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 2000
num_epochs: [5]
clip_grad: 10
early_stop: 500
train_holdout: 0.2
local_test: True
no_log_console: True

Algorithm	model	iid	dir5.0	dir2.0	dir1.0	dir0.1
fedavg	CNN	lr=0.1	lr=0.1	lr=0.05	lr=0.1	lr=0.1
fedprox	CNN	lr=0.05, μ=0.001	lr=0.1, μ=0.01	lr=0.1, μ=0.001	lr=0.1, μ=0.01	lr=0.05, μ=0.001
scaffold	CNN	lr=0.1	lr=0.1	lr=0.1	lr=0.1	lr=0.1
moon	CNN	lr=0.1, μ=0.1	lr=0.1, μ=0.1	lr=0.05, μ=0.1	lr=0.1, μ=0.1	lr=0.1, μ=0.1
feddyn	CNN	lr=0.1, α=0.1	lr=0.1, α=0.1	lr=0.05, α=0.1	lr=0.1, α=0.03	lr=0.05, α=0.03
fedavg	ResNet18	lr=0.1	lr=0.1	lr=0.1	lr=0.05	lr=0.1
fedprox	ResNet18	lr=0.05, μ=0.001	lr=0.1, μ=0.1	lr=0.1, μ=0.001	lr=0.05, μ=0.0001	lr=0.05, μ=0.001
scaffold	ResNet18	lr=0.1	lr=0.1	lr=0.1	lr=0.1	lr=0.1
moon	ResNet18	lr=0.1, μ=0.1	lr=0.05, μ=0.1	lr=0.05, μ=0.1	lr=0.05, μ=1.0	lr=0.05, μ=0.1
feddyn	ResNet18	lr=0.1, α=0.1	lr=0.1, α=0.1	lr=0.1, α=0.1	lr=0.1, α=0.1	lr=0.1, α=0.1
fedavg	ResNet18-GN	lr=0.1	lr=0.1	lr=0.1	lr=0.1	lr=0.1

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds
proportion: 0.2

Global Test

Algorithm	model	iid	dir5.0	dir2.0	dir1.0	dir0.1
fedavg	CNN	81.54±0.14	79.95±0.22	78.00±0.24	76.54±0.39	69.87±0.50
fedprox	CNN	80.83±0.20	79.86±0.30	78.39±0.26	76.58±0.50	69.41±0.33
scaffold	CNN	85.08±0.21	83.84±0.23	82.15±0.29	80.42±0.21	65.06±0.51
moon	CNN	80.88±0.27	79.63±0.20	77.21±0.33	75.67±0.26	62.44±1.10
feddyn	CNN	85.09±0.22	83.65±0.09	81.54±0.30	80.26±0.40	70.82±0.50
fedavg	ResNet18	94.07±0.12	93.56±0.18	92.59±0.09	91.53±0.11	78.79±0.56
fedprox	ResNet18	93.84±0.21	93.55±0.05	92.40±0.21	91.46±0.19	77.68±1.08
scaffold	ResNet18	95.09±0.12	94.67±0.11	93.86±0.13	92.90±0.18	79.82±0.51
moon	ResNet18	94.14±0.08	93.63±0.18	92.51±0.11	91.34±0.21	77.95±0.57
feddyn	ResNet18	95.00±0.13	93.87±0.08	92.98±0.23	92.72±0.21	79.76±0.73
fedavg	ResNet18-GN	91.25±0.23	89.93±0.24	88.21±0.40	86.42±0.73	66.39±1.66

Local Test

Algorithm	model	iid	dir5.0	dir2.0	dir1.0	dir0.1
fedavg	CNN	80.98±0.39	79.97±0.12	77.43±0.12	77.03±0.29	70.02±0.70
fedprox	CNN	80.68±0.18	79.81±0.41	77.80±0.13	77.05±0.20	69.72±0.79
scaffold	CNN	85.22±0.33	84.23±0.44	81.90±0.28	80.23±0.19	65.00±0.66
moon	CNN	80.26±0.38	79.29±0.29	76.81±0.58	76.12±0.45	62.26±1.09
feddyn	CNN	85.25±0.26	83.99±0.17	81.76±0.17	80.48±0.42	71.69±0.29
fedavg	ResNet18	94.58±0.08	93.54±0.11	93.12±0.23	91.67±0.25	79.46±0.77
fedprox	ResNet18	94.17±0.30	93.44±0.18	92.85±0.18	91.37±0.13	78.11±1.32
scaffold	ResNet18	95.51±0.13	94.89±0.16	94.28±0.16	93.18±0.15	80.39±0.54
moon	ResNet18	94.73±0.17	93.67±0.23	92.90±0.18	91.63±0.32	78.77±0.78
feddyn	ResNet18	95.16±0.20	94.19±0.13	93.41±0.19	92.83±0.26	80.70±1.02
fedavg	ResNet18-GN	91.69±0.19	90.04±0.26	87.84±0.48	86.52±0.69	66.98±1.59

Impact of Sampling Ratio

Task	Algorithm	model	p=0.1	p=0.2	p=0.5	p=1.0
iid	fedavg	CNN	81.70±0.30	81.54±0.14	81.34±0.23	81.87±0.17

Impact of Local Epoch

Back

CIFAR100

100 Clients

iid	dir1.0	dir0.1

Configuration

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
proportion: 0.2
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 2000
num_epochs: [5]
clip_grad: 10
early_stop: 500
train_holdout: 0.2
local_test: True
no_log_console: True

Algorithm	model	iid	dir1.0	dir0.1
fedavg	CNN	lr=0.1	lr=0.1	lr=0.1
fedprox	CNN	lr=0.1, μ=0.001	lr=0.1, μ=0.001	lr=0.05, μ=0.0001
scaffold	CNN	lr=0.1	lr=0.1	lr=0.1
feddyn	CNN	lr=0.001, α=0.1	lr=0.1, α=0.1	lr=0.1, α=0.03
moon	CNN	lr=0.1, μ=0.1	lr=0.1, μ=0.1	lr=0.05, μ=0.1
fedavg	ResNet18	lr=0.1	lr=0.05	lr=0.05
fedprox	ResNet18	lr=0.1, μ=0.0001	lr=0.05, μ=0.01	lr=0.05, μ=0.1
scaffold	ResNet18	lr=0.1	lr=0.1	lr=0.1
feddyn	ResNet18	lr=0.1, α=0.1	lr=0.05, α=0.1	lr=0.05, α=0.1
moon	ResNet18	lr=0.1, μ=10.0	lr=0.05, μ=0.1	lr=0.05, μ=0.1
fedavg	ResNet18-GN	lr=0.1	lr=0.1	lr=0.01

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds
proportion: 0.2

Global Test

Algorithm	model	iid	dir1.0	dir0.1
fedavg	CNN	41.33±0.30	37.93±0.64	22.50±0.52
fedprox	CNN	41.27±0.31	37.79±0.27	22.20±0.16
scaffold	CNN	49.85±0.22	41.08±0.39	18.56±0.50
feddyn	CNN	52.64±0.15	40.19±0.36	26.20±0.44
moon	CNN	41.49±0.40	37.28±0.37	21.09±0.32
fedavg	ResNet18	72.91±0.33	49.11±0.56	16.63±0.63
fedprox	ResNet18	73.26±0.29	48.42±1.00	16.45±0.19
scaffold	ResNet18	76.35±0.22	50.24±0.56	16.40±0.87
feddyn	ResNet18	75.43±0.17	50.23±0.61	17.46±0.62
moon	ResNet18	74.95±0.22	48.94±0.55	17.25±0.67
fedavg	ResNet18-GN	52.71±0.68	34.20±0.83	19.45±0.53

Local Test

Algorithm	model	iid	dir1.0	dir0.1
fedavg	CNN	41.04±0.34	36.94±0.52	22.21±0.51
fedprox	CNN	40.62±0.35	37.00±0.50	21.77±0.21
scaffold	CNN	49.94±0.22	39.70±0.31	18.58±0.61
feddyn	CNN	52.48±0.49	38.92±0.37	26.08±0.15
moon	CNN	40.91±0.30	36.22±0.34	20.67±0.26
fedavg	ResNet18	73.72±0.33	48.75±1.08	16.56±0.48
fedprox	ResNet18	73.70±0.27	47.72±1.04	16.79±0.30
scaffold	ResNet18	76.60±0.28	50.31±0.78	16.80±0.89
feddyn	ResNet18	75.53±0.40	50.39±0.61	17.96±0.65
moon	ResNet18	75.20±0.24	49.11±0.37	17.23±1.00
fedavg	ResNet18-GN	51.87±0.62	33.02±1.11	19.20±0.22

Back

TinyImageNet

100 Clients

iid	dir1.0	dir0.1

Configuration

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
proportion: 0.2
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 2000
num_epochs: [5]
clip_grad: 10
early_stop: 500
train_holdout: 0.2
local_test: True
no_log_console: True

Algorithm	model	iid	dir1.0	dir0.1
fedavg	ResNet18	lr=0.1	lr=0.05	lr=0.05
fedprox	ResNet18	lr=0.1, μ=0.01	lr=0.05, μ=1.0	lr=0.05, μ=1.0
scaffold	ResNet18	lr=0.05	lr=0.1	lr=0.1
feddyn	ResNet18	lr=0.1, α=0.1	lr=0.05, α=0.03	lr=0.05, α=0.03
moon	ResNet18	lr=0.1, μ=1.0	lr=0.05, μ=0.1	lr=0.05, μ=0.1

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds
proportion: 0.2

Global Test

Algorithm	model	iid	dir1.0	dir0.1
fedavg	ResNet18	58.89±0.34	19.10±0.51	6.42±0.17
fedprox	ResNet18	58.61±0.26	27.77±0.65	8.84±0.19
scaffold	ResNet18	60.02±0.33	24.89±0.74	8.54±0.79
feddyn	ResNet18	61.22±0.41	27.52±0.61	8.13±0.62
moon	ResNet18	58.70±0.26	22.55±0.47	6.78±0.22

Local Test

Algorithm	model	iid	dir1.0	dir0.1
fedavg	ResNet18	59.16±0.19	19.06±0.26	5.99±0.34
fedprox	ResNet18	59.33±0.20	28.03±0.14	8.38±0.16
scaffold	ResNet18	60.91±0.31	25.37±0.81	8.74±0.61
feddyn	ResNet18	62.07±0.20	27.25±0.68	7.75±0.59
moon	ResNet18	59.30±0.23	22.80±0.38	6.33±0.28

Back

MNIST

100 Clients

iid	dir5.0	dir2.0	dir1.0	dir0.1

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
proportion: 0.2
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 1000
num_epochs: [5]
clip_grad: 10
early_stop: 250
train_holdout: 0.2
local_test: True
no_log_console: True

Algorithm	model	iid	dir5.0	dir2.0	dir1.0	dir0.1
fedavg	CNN	lr=0.1	lr=0.05	lr=0.05	lr=0.05	lr=0.05
fedprox	CNN	lr=0.1, μ=0.0001	lr=0.05, μ=0.001	lr=0.1, μ=1.0	lr=0.1, μ=1.0	lr=0.05, μ=0.01
scaffold	CNN	lr=0.01	lr=0.01	lr=0.01	lr=0.01	lr=0.01
feddyn	CNN	lr=0.1, α=0.1	lr=0.1, α=0.1	lr=0.1, α=0.1	lr=0.05, α=0.1	lr=0.05, α=0.1
moon	CNN	lr=0.01, μ=1.0	lr=0.05, μ=0.1	lr=0.05, μ=1.0	lr=0.05, μ=0.1	lr=0.1, μ=0.1

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds
proportion: 0.2

Global Test

Algorithm	model	iid	dir5.0	dir2.0	dir1.0	dir0.1
fedavg	CNN	99.20±0.00	99.05±0.01	99.01±0.05	98.87±0.07	98.31±0.06
fedprox	CNN	99.22±0.03	99.05±0.03	99.06±0.04	98.90±0.02	98.36±0.07
scaffold	CNN	99.11±0.02	99.13±0.02	99.15±0.02	99.20±0.03	99.11±0.02
feddyn	CNN	99.37±0.01	99.23±0.02	99.25±0.03	99.20±0.04	98.89±0.04
moon	CNN	99.07±0.04	99.05±0.04	99.10±0.08	99.00±0.05	98.43±0.07

Local Test

Algorithm	model	iid	dir5.0	dir2.0	dir1.0	dir0.1
fedavg	CNN	98.87±0.03	98.97±0.03	99.07±0.07	98.67±0.06	97.96±0.09
fedprox	CNN	98.89±0.05	98.91±0.07	99.02±0.04	98.75±0.03	97.97±0.06
scaffold	CNN	98.97±0.02	98.96±0.02	99.15±0.02	98.87±0.05	98.90±0.05
feddyn	CNN	99.10±0.02	99.11±0.02	99.35±0.03	98.99±0.02	98.66±0.06
moon	CNN	98.79±0.04	98.89±0.02	99.12±0.01	98.73±0.02	98.08±0.13

Impact of Sampling Ratio

Task	Algorithm	model	p=0.1	p=0.2	p=0.5	p=1.0
iid	fedavg	CNN	99.20±0.03	99.20±0.00	99.21±0.02	99.22±0.00
dir5.0	fedavg	CNN	99.00±0.03	99.05±0.01	99.04±0.01	99.05±0.01
dir2.0	fedavg	CNN	99.01±0.04	99.01±0.05	99.04±0.02	99.05±0.01
dir1.0	fedavg	CNN	98.94±0.01	98.87±0.07	98.90±0.03	98.94±0.00
dir0.1	fedavg	CNN	98.27±0.08	98.31±0.06	98.32±0.04	98.33±0.02

Impact of Local Epoch

Back

FEMNIST

3597 Clients

client-id

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 2000
num_epochs: 5
clip_grad: 10
proportion: 0.2
early_stop: 400
train_holdout: 0.2
local_test: True
no_log_console: True
log_file: True

Algorithm	model	client-id
fedavg	CNN	lr=0.1
fedprox	CNN	lr=0.1, μ=0.0001
scaffold	CNN	lr=0.05
feddyn	CNN	lr=0.05, α=0.03
moon	CNN	lr=0.1, μ=0.1

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds

Global Test

Algorithm	model	client-id
fedavg	CNN	86.25±0.04
fedprox	CNN	86.24±0.02
scaffold	CNN	86.92±0.07
feddyn	CNN	86.90±0.05
moon	CNN	86.29±0.01

Local Test

Algorithm	model	client-id
fedavg	CNN	84.91±0.05
fedprox	CNN	84.76±0.06
scaffold	CNN	87.39±0.08
feddyn	CNN	87.30±0.08
moon	CNN	86.63±0.05

Back

AGNEWS

100 Clients

iid	dir1.0	dir0.1

learning_rate: [0.1, 0.5, 1.0, 5.0, 10.0]
batch_size: 50
weight_decay: 1e-4
momentum: 0.9
lr_scheduler: 0
learning_rate_decay: 0.998
num_rounds: 600
num_epochs: 1
clip_grad: 10
proportion: 0.2
early_stop: 125
train_holdout: 0.2
local_test: True
no_log_console: True
log_file: True

Algorithm	model	iid	dir1.0	dir0.1
fedavg	EmbeddingBag+Linear	lr=1.0	lr=1.0	lr=0.5
fedprox	EmbeddingBag+Linear	lr=1.0, μ=0.01	lr=1.0, μ=0.01	lr=0.5, μ=0.0001
scaffold	EmbeddingBag+Linear	lr=1.0	lr=1.0	lr=0.5
feddyn	EmbeddingBag+Linear	lr=0.1, α=0.01	lr=0.5, α=0.01	lr=0.5, α=0.01
moon	EmbeddingBag+Linear	lr=1.0, μ=10.0	lr=0.5, μ=10.0	lr=0.5, μ=0.1

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds
proportion: 0.2

Global Test

Algorithm	model	iid	dir1.0	dir0.1
fedavg	EmbeddingBag+Linear	90.93±0.06	89.37±0.14	87.87±0.15
fedprox	EmbeddingBag+Linear	90.94±0.11	89.39±0.12	87.97±0.11
scaffold	EmbeddingBag+Linear	90.28±0.28	87.98±0.65	85.96±0.45
feddyn	EmbeddingBag+Linear	91.04±0.06	91.02±0.02	91.14±0.04
moon	EmbeddingBag+Linear	91.54±0.11	90.28±0.07	87.61±0.13

Local Test

Algorithm	model	iid	dir1.0	dir0.1
fedavg	EmbeddingBag+Linear	91.53±0.03	89.84±0.08	88.35±0.16
fedprox	EmbeddingBag+Linear	91.50±0.01	89.87±0.06	88.30±0.20
scaffold	EmbeddingBag+Linear	90.58±0.50	88.31±0.69	86.50±0.49
feddyn	EmbeddingBag+Linear	91.59±0.04	91.11±0.01	91.04±0.01
moon	EmbeddingBag+Linear	92.03±0.05	90.64±0.04	88.08±0.19

Back

Office-Caltech10

4 Clients

domain

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 500
num_epochs: 1
clip_grad: 10
sample: full
proportion: 1.0
early_stop: 100
train_holdout: 0.2
local_test: True
no_log_console: True
log_file: True

Algorithm	model	domain
standalone	AlexNet	lr=0.1
fedavg	AlexNet	lr=0.1
fedprox	AlexNet	lr=0.1, μ=0.1
scaffold	AlexNet	lr=0.1
feddyn	AlexNet	lr=0.01, α=0.1
moon	AlexNet	lr=0.05, μ=1.0

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds

Local Test

Algorithm	model	$\textbf{Client}_{Caltech}$	$\textbf{Client}_{Amazon}$	$\textbf{Client}_{Dslr}$	$\textbf{Client}_{Webcam}$	Mean	Weighted-Mean
standalone	AlexNet	47.32±2.82	72.42±3.73	50.67±11.62	62.76±25.00	58.29±7.92	58.82±5.18
fedavg	AlexNet	71.25±1.99	88.00±0.84	80.00±0.00	84.14±4.14	80.85±1.37	79.63±1.11
fedprox	AlexNet	72.68±2.37	72.68±2.37	82.67±3.27	93.10±3.08	83.90±1.66	81.15±1.56
scaffold	AlexNet	70.54±2.33	86.32±1.76	88.00±4.99	93.79±3.38	84.66±2.23	80.30±1.79
feddyn	AlexNet	73.75±2.91	84.00±1.68	88.00±4.99	95.17±1.69	85.23±1.96	81.00±1.79
moon	AlexNet	71.61±2.55	87.16±1.23	78.67±2.67	88.97±3.38	81.60±1.06	79.95±1.27

Back

DomainNet

6 Clients

domain

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 500
num_epochs: 1
clip_grad: 10
sample: full
proportion: 1.0
early_stop: 100
train_holdout: 0.2
local_test: True
no_log_console: True
log_file: True

Algorithm	model	domain
standalone	AlexNet	lr=0.05
fedavg	AlexNet	lr=0.1
fedprox	AlexNet	lr=0.1, μ=0.01
scaffold	AlexNet	lr=0.1
feddyn	AlexNet	lr=0.05, α=0.1
moon	AlexNet	lr=0.1, μ=0.1

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds

Local Test

Algorithm	model	Clipart	Infograph	Painting	Quickdraw	Real	Sketch	Mean	Weighted-Mean
standalone	AlexNet	52.84±0.61	24.43±0.49	44.99±0.94	81.23±0.47	65.69±0.48	48.68±1.24	52.98±0.33	57.12±0.26
fedavg	AlexNet	74.54±0.96	38.42±0.63	64.56±0.60	78.44±0.92	74.67±0.32	72.76±0.92	67.23±0.20	68.90±0.20
fedprox	AlexNet	75.46±1.34	38.51±0.88	64.56±1.26	77.39±1.82	74.04±0.57	73.28±1.16	67.20±0.63	68.67±0.66
scaffold	AlexNet	78.25±0.36	40.77±0.41	66.70±1.39	78.77±0.90	75.12±0.92	77.09±0.69	69.45±0.41	70.65±0.44
feddyn	AlexNet	78.55±1.04	39.82±0.71	67.18±1.20	78.23±0.30	74.34±0.67	75.40±1.32	68.92±0.22	70.09±0.17
moon	AlexNet	72.77±0.58	38.46±1.04	63.38±0.74	79.75±1.40	73.83±0.48	70.66±0.45	66.48±0.54	68.35±0.58
fedbn	AlexNet	77.08±0.74	40.22±1.67	69.26±0.40	88.83±0.32	82.58±0.18	75.89±0.75	72.31±0.28	74.88±0.21

Back

PACS

4 Clients

domain

learning_rate: [0.001, 0.01, 0.05, 0.1, 0.5]
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 500
num_epochs: 5
clip_grad: 10
proportion: 1.0
early_stop: 100
train_holdout: 0.2
local_test: True
no_log_console: True
log_file: True

Algorithm	model	domain
standalone	AlexNet	lr=0.1
fedavg	AlexNet	lr=0.05
fedprox	AlexNet	lr=0.05, μ=0.0001
scaffold	AlexNet	lr=0.1
feddyn	AlexNet	lr=0.1, α=0.03
moon	AlexNet	lr=0.05, μ=1.0

Local Test

Algorithm	model	$\textbf{Client}_{ArtPainting}$	$\textbf{Client}_{Cartoon}$	$\textbf{Client}_{Photo}$	$\textbf{Client}_{Sketch}$	Mean	Weighted-Mean
standalone	AlexNet	33.04±2.93	59.49±1.90	65.75±2.25	76.94±1.39	58.80±1.40	61.97±1.38
fedavg	AlexNet	61.47±1.47	84.53±2.07	71.14±3.54	82.65±1.56	74.95±0.73	76.83±0.78
fedprox	AlexNet	65.78±2.55	81.88±0.96	76.05±2.18	80.97±1.03	76.17±0.78	77.25±0.46
scaffold	AlexNet	63.82±2.07	82.91±0.85	76.53±2.47	82.81±1.71	76.52±0.48	77.89±0.30
feddyn	AlexNet	63.53±2.31	83.08±0.44	76.65±1.37	81.12±1.29	76.09±1.10	77.23±1.07
moon	AlexNet	62.16±2.04	83.33±3.21	72.93±2.38	83.52±1.67	75.49±0.67	77.33±0.82

Back

Digits

5 Clients

domain

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 200
num_epochs: 5
clip_grad: 10
sample: full
proportion: 1.0
early_stop: 50
train_holdout: 0.2
local_test: True
no_log_console: True
log_file: True

Algorithm	model	dataset
standalone	AlexNet	lr=0.1
fedavg	AlexNet	lr=0.05
fedprox	AlexNet	lr=0.05, μ=0.1
scaffold	AlexNet	lr=0.05
feddyn	AlexNet	lr=0.1, α=0.1
moon	AlexNet	lr=0.05, μ=0.1

Local Test

Algorithm	model	$\textbf{Client}_{USPS}$	$\textbf{Client}_{SVHN}$	$\textbf{Client}_{MNIST}$	$\textbf{Client}_{Synthetic}$	$\textbf{Client}_{MNISTM}$	Mean	Weighted-Mean
standalone	AlexNet	99.59±0.06	18.03±0.00	99.48±0.04	98.77±0.04	98.39±0.06	82.85±0.02	82.85±0.02
fedavg	AlexNet	99.54±0.10	96.50±0.08	99.75±0.03	99.24±0.08	97.88±0.20	98.58±0.03	98.58±0.03
fedprox	AlexNet	99.61±0.05	96.67±0.16	99.79±0.02	99.30±0.04	98.83±0.04	98.84±0.04	98.84±0.04
scaffold	AlexNet	99.62±0.03	97.51±0.10	99.84±0.02	99.44±0.04	98.86±0.07	99.05±0.04	99.05±0.04
feddyn	AlexNet	99.47±0.22	96.82±0.46	99.77±0.03	99.39±0.04	98.89±0.09	98.87±0.07	98.87±0.07
moon	AlexNet	99.57±0.08	96.54±0.20	99.77±0.03	99.27±0.03	97.94±0.25	98.62±0.07	98.62±0.07

Back

ProstateMRI

6 Clients

domain

learning_rate: [0.00005, 0.0001, 0.0005, 0.001, 0.005]
batch_size: 16
weight_decay: 1e-4
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 500
num_epochs: 1
clip_grad: 10
proportion: 1.0
early_stop: 100
train_holdout: 0.2
local_test: True
optimizer: Adam
no_log_console: True
log_file: True

Algorithm	model	domain
standalone	UNet	lr=0.0005
fedavg	UNet	lr=0.0001
fedprox	UNet	lr=0.0001, μ=0.0001
scaffold	UNet	lr=0.0001
feddyn	UNet	lr=0.00005, α=0.1
moon	UNet	lr=0.0001, μ=0.1

Back

Algorithm	model	BMC	UCL	BIDMC	RUNMC	HK	I2CVB	Mean	Weighted-Mean
standalone	UNet	91.54±0.24	79.77±1.65	91.29±0.69	91.63±0.83	92.78±0.60	94.02±0.65	90.17±0.47	91.15±0.41
fedavg	UNet	91.09±0.69	90.74±0.99	93.02±0.69	94.32±0.39	94.84±0.54	95.98±0.10	93.33±0.12	93.59±0.13
fedprox	UNet	91.94±0.67	90.83±0.70	93.27±0.29	94.87±0.19	94.84±0.29	95.67±0.32	93.57±0.23	93.86±0.24
scaffold	UNet	56.16±0.09	49.23±0.03	51.56±0.02	54.69±0.06	53.50±0.05	46.92±0.07	52.01±0.02	51.98±0.02
feddyn	UNet	91.41±1.14	90.96±1.49	91.89±1.49	94.28±0.38	93.93±0.85	93.83±1.00	92.72±0.28	92.90±0.25
moon	UNet	91.93±0.36	89.85±1.10	92.39±0.63	94.30±0.43	94.06±0.29	96.17±0.28	93.12±0.23	93.57±0.20
fedbn	UNet	95.47±0.23	91.13±0.81	93.41±0.30	93.89±0.54	94.56±0.62	96.65±0.11	94.18±0.23	94.64±0.19

Camelyon17

5 Clients

hospital

Back

Fundus

4 Clients

hospital

learning_rate: [0.00005, 0.0001, 0.0005, 0.001, 0.005]
batch_size: 16
weight_decay: 1e-4
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 500
num_epochs: 1
clip_grad: 10
proportion: 1.0
early_stop: 50
train_holdout: 0.2
local_test: True
optimizer: Adam
no_log_console: True
log_file: True

Algorithm	model	domain
standalone	UNet	lr=0.001
fedavg	UNet	lr=0.001
fedprox	UNet	lr=0.001, μ=0.0001
scaffold	UNet	lr=0.001
feddyn	UNet	lr=0.0005, α=0.01
moon	UNet	lr=0.0005, μ=0.1

Local Test

Algorithm	model	$\textbf{Client}_{1}$	$\textbf{Client}_{2}$	$\textbf{Client}_{3}$	$\textbf{Client}_{4}$	Mean	Weighted-Mean
standalone	AlexNet	81.44±1.92	74.51±1.03	82.92±2.03	74.42±0.94	78.32±1.08	78.32±1.09
fedavg	AlexNet	76.60±4.68	60.69±4.27	81.42±5.98	77.87±2.23	74.15±2.91	77.05±3.32
fedprox	AlexNet	82.64±1.02	66.90±2.62	78.45±4.51	74.38±3.77	75.59±2.44	75.61±3.02
scaffold	AlexNet	54.24±3.17	49.25±0.65	72.92±0.80	65.51±1.85	60.48±0.88	65.74±0.58
feddyn	AlexNet	82.58±1.90	67.23±2.81	85.21±5.53	78.61±4.88	78.41±3.28	80.09±4.41
moon	AlexNet	80.61±3.66	66.55±3.48	87.23±1.48	79.72±3.03	78.53±1.35	81.15±1.36

Back

EndoPolyp

5 Clients

hospital

Back

SpeechCommand

2112 Clients

client-id

learning_rate: [0.001, 0.005, 0.01, 0.05, 0.1]
batch_size: 50
weight_decay: 1e-3
lr_scheduler: 0
learning_rate_decay: 0.998
num_rounds: 2000
num_epochs: 1
clip_grad: 10
proportion: 0.05
early_stop: 250
train_holdout: 0.0
no_log_console: True
log_file: True

Algorithm	model	client-id
fedavg	M5	lr=1.0
fedprox	M5	lr=1.0, mu=0.01
scaffold	M5	lr=1.0
feddyn	M5	lr=0.1, α=0.001
moon	M5	lr=1.0, mu=0.1

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds

Global Test

Algorithm	model	client-id
fedavg	M5	69.11±0.91
fedprox	M5	69.30±0.87
scaffold	M5	64.40±0.42
feddyn	M5	60.65±0.76
moon	M5	69.08±0.86

Back

Shakespeare

1012 Clients

client-id

learning_rate: [0.1, 0.5, 1.0, 5.0, 10.0]
batch_size: 50
weight_decay: 5e-4
lr_scheduler: 0
learning_rate_decay: 0.9998
num_rounds: 200
num_epochs: 5
clip_grad: 10
proportion: 0.1
early_stop: 100
train_holdout: 0.2
local_test: True
no_log_console: True
log_file: True

Algorithm	model	client-id
fedavg	LSTM	lr=0.1
fedprox	LSTM	lr=0.1, μ=0.0001
scaffold	LSTM	lr=0.5

Main Results

seed: [2,4388,15,333,967] # results are averaged over five random seeds

Global Test

Algorithm	model	client-id
fedavg	LSTM	52.85±0.06
fedprox	LSTM	53.09±0.06
scaffold	LSTM	49.93±0.09

Local Test

Algorithm	model	client-id
fedavg	LSTM	52.76±0.17
fedprox	LSTM	53.31±0.04
scaffold	LSTM	50.01±0.14

Back