tdye24 / Fed101

Fed101

[e-mail]

cd Fed101/algorithm/FedAVG &
python fedavg-main.py
-dataset
femnist
-model
femnist
--lr
0.03
--lr-decay
0.99
--decay-step
1
--batch-size
10
--clients-per-round
10
--num-rounds
1000
--seed
12
--epoch
5
--eval-interval
1
--note
run_1_seed_12

python fedavg-main.py 
-dataset femnist 
-model femnist 
--lr 0.03 
--lr-decay 0.99 
--batch-size 10 
--clients-per-round 10 
--num-rounds 1000 
--seed 24 
--epoch 5 
--eval-interval 1 
--note run_2_seed_24

python fedavg-main.py -dataset femnist -model femnist --lr 0.03 --lr-decay 0.99 --batch-size 10 --clients-per-round 10 --num-rounds 1000 --seed 24 --epoch 5 --eval-interval 1 --note run_2_seed_24

Dataset Overview

dataset	task	metric	client	training set	mean|std|skewness	test set	mean|std|skewness	partition	link
MNIST	10 clf	acc	1000	61664	61.664|144.63|24751822	7371	7.371|16.0772|34058.3	power law
FEMNIST	62 clf	acc	3400	671585	197.53|76.681|391488.3	77483	22.7891|8.5105|533.892	realistic partition
CIFAR10	10 clf	acc	100	50000	500|147.22|-286980	10000	NA|NA|NA	LDA

MNIST

Description

​ 1000 clients, refer to fedprox

Model

​ CNN+FCNN

Algorithm & Result

FedAVG

#\T	50	60	65	70	75	80	85

FedProx

#\T	50	60	65	70	75	80	85

FedSP

#\T	50	60	65	70	75	80	85

FedMC

#\T	50	60	65	70	75	80	85

CIFAR10

Description

​ 100 clients (10 groups), for certain group, the clients belong to it share the 90% of the specified class.

Model

​ CNN+FCNN

Algorithm & Result

FedAVG

#\T	30	35	40	45	50	55	60	61	62	63	64	65	66	67	68	69	70	O|R

FedProx

u = 0.1

#\T	30	35	40	45	50	55	60	61	62	63	64	65	66	67	68	69	70	O|R

FedSP

#\T	30	35	40	45	50	55	60	61	62	63	64	65	66	67	68	69	70	O|R
1

FedMC

gradient penalty = 0
critic = 20
with sigmoid

#\T	30	35	40	45	50	55	60	61	62	63	64	65	66	67	68	69	70	O|R

FEMNIST

Overview

Description

​ Partition dataset based on the writer of the digit/character.

​ Sample clients based on the number of samples it has(>=350).

​ original dataset: 3500 clients and 785697 samples.

​ sampled subset: 503 clients, 193081 samples

Model

​ CNN+FCNN

Hyper-parameters

• clients/round: 10/503
• epoch: 5
• batch-size: 10
• lr: 0.03
• lr-decay: 0.99
• decay-step: 1
• rounds: 1000

Algorithm & Result

FedAVG

#\T	50	60	65	70	75	80	81	82	83	84	85	R|O	Note
1	9|51.42	13|60.43	17|65.46	23|70.41	38|75.09	89|80.21	104|81.08	138|82.26	204|83.07	316|84.08		600|84.53	seed=12

FedProx

u = 0.1

#\T	50	60	65	70	75	80	81	82	83	84	85	O|R	Note
1

FedSP

#\T	50	60	65	70	75	80	81	82	83	84	85	O|R	Note
1

FedMC

gradient penalty = 0
critic = 20
with sigmoid

#\T	50	60	65	70	75	80	81	82	83	84	85	O|R	Note
1