huiserwang / federated-hierarchical-project

Game Theory Project

Experiments are produced on MNIST and CIFAR10.

Requirments

Install all the packages from requirements.txt:

pip -r install requirements.txt

• Python >= 3.6
• Pytorch >= 1.2.0
• Torchvision >= 0.4.0
• Numpy>=1.15.4

Data

• Download train and test datasets on your own, otherwise they will be automatically downloaded from torchvision datasets.
• Experiments are run on MNIST and Cifar.

Running the experiments

Federated experiment involves training a global model using many local models.

• To run the hierarchical federated learning experiment with CIFAR on CNN:

python federated_hierarchical_main.py --model=cnn --dataset=cifar --gpu=0 --local_ep 2 --epochs=10

• To run the hierarchical federated learning experiment with CIFAR on MLP:

python federated_hierarchical_main.py --model=mlp --dataset=cifar --gpu=0 --local_ep 2 --epochs=10

• To run the hierarchical federated learning experiment with MNIST on CNN:

python federated_hierarchical_main.py --model=cnn --dataset=mnist --gpu=0 --local_ep 2 --epochs=10

• To run the hierarchical federated learning experiment with MNIST on MLP:

python federated_hierarchical_main.py --model=mlp --dataset=mnist --gpu=0 --local_ep 2 --epochs=10

You can change the default values of other parameters to simulate different conditions. Refer to the options section.

Options

The default values for various paramters parsed to the experiment are given in options.py. Details are given some of those parameters:

• --dataset: Default: 'mnist'. Options: 'mnist', 'cifar'
• --model: Default: 'mlp'. Options: 'mlp', 'cnn'
• --gpu: Default: None (runs on CPU). Can also be set to the specific gpu id (i.e. 0).
• --epochs: Default: 10. Number of rounds of training.
• --lr: Learning rate set to 0.01 by default.
• --verbose: Detailed log outputs. Activated by default, set to 0 to deactivate.
• --seed: Random Seed. Default set to 1.

Federated Parameters

• --num_fv: Number of FL Vehicles. Default is 6.

• --num_fr: Number of FL RSUs. Default is 5.

• --local_ep: Number of local training epochs in each user. Default is 2.

• --local_bs: Batch size of local updates in each user. Default is 32.