FedGrAINS

Requirements & Data Preparation

To install requirements

conda env create -f fedgdrop.yaml

Then, METIS (for data generation), please follow https://github.com/james77777778/metis_python

Running examples

• OneDS: Distributing one dataset to a number of clients:

sh run_fedgdrop_cora.sh 10 1 64 0 0 gcflplus configs/gcflplus_cora_10_fedgfn_disjoint.txt

Script repeats the experiment three times with different seeds. You can find scripts for each datasets

Parameters

num of clients=$1
experiment number (for logging results)=$2
number of batches=$3
Dataset =$4 (Only for un_fedgdrop_amazon.sh)
Enable overlap=$5
Enable local training of flow network=$6
FL Algorithm =$7
Config File=$8

Running examples

• After running the above command lines, the raw results are stored in ./outputs/raw/.

• Then, to process the raw results:

python exps/aggregateResults.py --dataset 'Cora' --algo fedavg --model GCNv2 --numcli 5 --exp_num 1 

• Finally, the results are stored in ./outputs/processed/.

Options

The default values for parameters parsed to the experiment are given in ./exps/main_oneDS.py. Details about some of those parameters are given here.

• --dataset: The subgraph FL dataset. Default: 'Cora'. Options: 'Citeseer', 'PubMed', 'MS', .
• --num_rounds: The number of rounds for simulation. Default: 200.
• --local_epoch: The number of local epochs. Default: 1.
• --hidden: The number of hidden units. Default: 64.
• --nlayer: The number of GNN layers. Default: 3.

Parameter tuning

Under exps/, fedgdrop_hp_sweep.py tunes GFlowNet parameters over random 100 runs in the given range of Python script.