README.md

MFL

Setup

To run this example, you need to download this directory to your system. You can do so by running the following command in the shell or git bash:

git clone --depth=1 https://github.com/QVQZZZ/HeFlwr.git \
&& mv HeFlwr/examples/mfl . \
&& rm -rf HeFlwr \
&& cd mfl

If you have multiple devices installed with HeFlwr and wish to run federated learning across multiple devices, you need to run the above command on the other devices as well.

To run this example, you need to modify the server_address in the client{N}.py file and the client IP in strategy.py.

Running

Choose a device you prefer and run server.py to let it act as the federated learning server.

Then, on any device, run client.py to let it become a federated learning client.

For example, you can run the following commands in multiple terminals on the same machine:

python server.py --dataset cifar10 --num_rounds 3
python client.py --server_address 127.0.0.1:8080 --client_num 4 --cid 1 --dataset cifar10 --partition noniid --alpha 0.5 --batch_size 32
python client.py --server_address 127.0.0.1:8080 --client_num 4 --cid 2 --dataset cifar10 --partition noniid --alpha 0.5 --batch_size 32
python client.py --server_address 127.0.0.1:8080 --client_num 4 --cid 3 --dataset cifar10 --partition noniid --alpha 0.5 --batch_size 32
python client.py --server_address 127.0.0.1:8080 --client_num 4 --cid 4 --dataset cifar10 --partition noniid --alpha 0.5 --batch_size 32

This will perform federated learning on the CIFAR-10 dataset using the ResNet-18 network, with a total of 3 communication rounds, 4 clients, and a Dirichlet distribution with alpha=0.5 for data partitioning.

Alternatively, you can use the standalone run scripts we provide to quickly reproduce the experiments on the HeFlwr homepage:

./mfl.sh --num_rounds 50 --client_num 8 --dataset mnist --partition noniid --alpha 0.5 --batch_size 64  # Unix shell
.\mfl.ps1 -num_rounds 50 -client_num 8 -dataset mnist -partition noniid -alpha 0.5 -batch_size 64  # Windows powershell

Results

You can get the training process loss and acc data on the device terminal running the server.

A mfl_test_log.txt file will be generated in the running directory of each device, which records the device load information during the training process.