This is the code for paper "Fed-GraB: Federated Long-tailed Learning with Self-Adjusting Gradient Balancer".
| parameters | description |
|---|---|
rounds |
Number of rounds in training process, option:500 |
num_users |
Number of clients, option:40,20 |
local_bs |
Batch size for local training, option:5 |
beta |
Coefficient for local proximal term, option: 0.01,0 |
model |
neural network model, option: resnet18,resnet34,resnet50 |
dataset |
Dataset, option:cifar10,cifar100,imagenet and inat |
iid |
Action iid or non iid, option: store_true |
alpha_dirichlet |
Parameter for Dirichlet distribution, option: 10,1 |
- To train on CIFAR-10 with IID data partition and imbalanced factor 100 over 40 clients:
python fed_grab.py --dataset cifar10 --iid --IF 0.01 --local_bs 5 --rounds 500 --num_users 40 --beta 0 --dataset cifar10 --model resnet18 --gpu 0
- To train on CIFAR-10 with non-IID data partition with imbalanced factor 100 , alpha=1 over 40 clients:
python fed_grab.py --dataset cifar10 --alpha_dirichlet 1 --IF 0.01 --local_bs 5 --rounds 500 --num_users 40 --beta 0 --dataset cifar10 --model resnet18 --gpu 0