CS230Project_FL

Dependency Installation

(On all instances)

• Install flower, pytorch libraries using the following commands.

sudo apt update
sudo apt install python3-pip
pip3 install torch torchvision --no-cache-dir ## Note that the memory on instances are too small, we need to add --no-cache-dir flag
pip3 install flwr==1.4.0
pip3 install hydra-core
pip3 install ray==1.11.1 # Required only when running simulation

Data Preparation

• Run the command below to finish dataset setup

  cd datasets/SumoSimulation
  python3 data_preprocess.py
  

Framework Configuration and Execution

• Flower uses port 8081 to communicate between nodes. Allow access through port 8081 from client nodes.

• change the config file conf/base.yaml of the server and client parameters. For testing on a real machine, please change the below parameters:

  • client_id: must start from 0 and should be sequantially assigned different number to every client
  • server_address: IP, port of the server identified by a client
  • num_clients: the total number of the client (for training data partitions)
  • uniform_data_distribution: True for uniform data distribution, False for heterogenous data distribution

• On the server, run the server script using the below command:

python3 server.py

• Wait until the initialization of the server script finishes (Should show "Requesting initial parameters from one random client"). On each client, run the server script using the below command:

python3 client.py

• Note that the server only starts when there are equal to or more than 2 client nodes. Wait until the training process finishes.