Pytorch Federated Learning

This framework enables to use a FL system with Pytorch.

It is being carried as my final thesis @ King's College London.

It currently works with the Human grasp dataset. Repo here.

Scheme

Orchestrator process:

Results

You can see the results of the experiments done here.

How to run

Step 1. Copy the dataset locally

• Run the following commands:

  wget http://stag.csail.mit.edu/datasets/classification_lite.zip -O temp.zip
  unzip temp.zip -d data/classification/
  rm temp.zip
  

Step 2. There are 3 main ways to run the program

1. Spawning Python with the shell script (Recommended).

   • Note: screen must be installed.
   • Change the PYTHON_PATH in the initialize.sh file to point to your python where all the requirements are installed.
   • Add or remove as many clients as desired in the initialize.sh file.
   • Attention! The number and ports of clients in the initialize.sh file should match the hosts.yml file.
   • Run the command: DELETE_OLD_LOGS=1 RESTART_SCREEN=1 SPLIT_TYPE=<split_type> ./initialize.sh
     • There are 3 env variables which can be set:
       • DELETE_OLD_LOGS: if set to 1, removes all logs in logs/ folder. Useful for new run.
       • RESTART_SCREEN: if set to 1, kills all current screens. Useful for new run.
       • N_FRAMES: Number of input frames.
       • SPLIT_TYPE: Can be set to 3 values (more information in dissemination):
         • no_split: No split in the dataset is performed.
         • iid: IID split type is performed.
         • non-iid-a: non-IID split type is performed.
   • If you want to attach to a running screen, run the command screen -S <screen_name>.
   • If you want to dettach from a screen, when attached, run the commands ctrl+A ctrl+D.
   • If you want to kill to an exiting screen, when attached, run the commands ctrl+A ctrl+K Y.
   • All connections and training status will be visible here.

2. Spawning Python manually (If screen not installed or more control desired).

   • Induce the functionality from the initialize.sh file.

3. Using Docker (Currently not compatible with GPU acceleration):

   • Create the image by running: make build.
   • Run the image: make run. This will spawn the main server, secure aggregator and two clients. In order to increase the number of clients, modify the docker-compose.yml and hosts.yml files.

Frontend

To start the frontend just run the command: python frontend/app.py. It is recommended to attach it to a service file and deploy it to a server in order to be always running. To restart the frontend just perform the request: http://95.179.192.253:8002/restart.

Here is a screenshot of the frontend currently running on http://95.179.192.253:8002/frontend.

TODO

• Create an 'orchestrator' with a frontend to manage all the connections and processes.
• Make proper documentation explaining how to select no split, IID or NON-IID dataset distributions.
• Repeat experiments with 5 clients.
• RAM optimizations.
• Perform aggregation even if one client is down for a certain period of time.
• RSA Encryption.
• Documentation of the frontend.
• Repeat experiments with multiple machines
• Limit resources of each client in order to be able to perform comparisons.
• Split dataset in multiple clients doesn't work for iid and non-iid