Federated-Long-tailed-Learning

This repository is the official implementation of "Global Balanced Experts for Federated Long-Tailed Learning"

Installation

conda env create -f ldae.yaml

Run

cd fedml_experiments/clsimb_fedavg/

For ldae on cifar100-lt, run:

python -u ./main_fedavg.py \
--comm_round 2000 --epochs 2 --batch_size 64 --client_optimizer sgd --lr 0.6 --lr_decay 0.05 \
--imb_factor 0.01 --partition_alpha 0.1 --method ldae_train_exp_esti_global --frequency_of_the_test 50 --beta 0.8

Other re-balance strategies are available: focal, ldam, lade, blsm, ride. They can use different class priors like: local re-balance: ldam; global re-balance: ldam_real_global; GPI: ldam_esti_global.