BrandonC8310 / Federated-Learning

A federated learning using multi-class logistic regression on MNIST

COMP3221 ASM2

Before running the program, please firstly run the following command (for MacOS) to change the maximum udp transporting data size:

sudo sysctl -w net.inet.udp.maxdgram=65535

Run the server firstly.

There are two options to run the server:

Option 1:

python3 COMP3221_FLServer.py 6000 0 (use port number 6000 for the server and aggregates all the client models)

Option 2:

python3 COMP3221_FLServer.py 6000 1 (use port number 6000 for the server and aggregates random 2 client models from the five models)

Run the five clients Secondly.

There are two options to run a client:

Option 1:

python3 COMP3221_FLClient.py client1 6001 0 (use port number 6001 for the client with id "client1" , and use GD)

Option2:

python3 COMP3221_FLClient.py client1 6001 1 (use port number 6001 for the client with id "client1" , and use Mini-Batch GD)

An example to run all the five clients:

python3 COMP3221_FLClient.py client1 6001 0

python3 COMP3221_FLClient.py client2 6002 0

python3 COMP3221_FLClient.py client3 6003 0

python3 COMP3221_FLClient.py client4 6004 1

python3 COMP3221_FLClient.py client5 6005 1

This indicates that we use GD for client1, client2, and client3. And we use Mini-Batch GD for client 4 and client 5.

After the server and the clients finished executing, run the evaluation program:

python3 COMP3221_FLEvaluation.py

This will output diagrams of the overall performance of the global model (loss and accuracy).

What we've done

Everything including addressing the client failiure issue.