Neardata Fl For Transcriptomics

---

Table of Contents

• Environment setup
• Running flower tutorial example
• License

Environment setup

Compile the requirements.in file to get requirements.txt

pip-compile requirements.in

Create and activate your virtual environment.

python -m venv venv
source venv/bin/activate

Now install packages from the requirements.txt file.

pip install -r requirements.txt

Install also neardata-fl-for-transcriptomics from source to run scripts from the root dir.

pip install .

We use NeptuneAI as the MLOps tool. The api token should be stored in .env file, so that python-dotenv package can load it as an env variable. The .env file should have the following entry, where 'xxx' is the api token:

NEPTUNE_API_TOKEN=xxx

Running flower tutorial example

Running locally

First, run the data split service, which is a http server that assigns samples to each client (example for a dataset with 60000 samples):

python3 fl/flower_tutorial/scripts/run_data_split_service.py --service-ip localhost --n-samples=60000 --n-splits 3 --manual-seed 1

You can instantiate the FL server by running:

python3 fl/flower_tutorial/scripts/run_server.py --num-clients 3 --server-ip localhost --num-rounds 50 --num-local-epochs 1

The --num-clients argument is the minimum number of clients the federated learning round can start with. The --server-ip argument is the address of the server. The client can be instantiated by running:

python3 fl/flower_tutorial/scripts/run_client.py --server-ip localhost --data-split-service-ip localhost 

Clients and server are using port 8081.

Running through SLURM

Setting up the environment

This script will create a virtual environment that can be used to run the tutorial example.

SBATCH run_configure_venv.sh

To run a FL workflow, you can use:

sbatch -n 4 run_flower_tutorial.sh 3

This script will run a parallel job a server + 3 clients on 4 nodes.

Running through Docker Compose

You can run the workflow using Docker Compose. First, the data split service:

docker compose -f ./docker/flower_tutorial/docker-compose.yml up -d data-split-service

Second, instantiate the server container:

docker compose -f ./docker/flower_tutorial/docker-compose.yml up --build -d server

Then, you can build and start client containers:

docker compose -f ./docker/flower_tutorial/docker-compose.yml up --build -d --scale client=4

The --scale flag will allow you to instantiate a number of client containers, in this case, 4.

Running genotypes use case

Running locally

First, run the data split service, which is a http server that assigns samples to each client (example for a dataset with 60000 samples):

python3 fl/genotypes/scripts/run_data_split_service.py --service-ip localhost --n-samples=400 --n-splits 2 --manual-seed 1

You can instantiate the FL server by running:

python3 fl/genotypes/scripts/run_server.py --num-clients 2 --server-ip localhost --num-rounds 2 --num-local-epochs 1

The --num-clients argument is the minimum number of clients the federated learning round can start with. The --server-ip argument is the address of the server. The client can be instantiated by running:

python3 fl/genotypes/scripts/run_client.py --server-ip localhost --data-split-service-ip localhost 

Clients and server are using port 8081.

Running through Docker Compose

You can run the workflow using Docker Compose. First, the data split service:

docker compose -f ./docker/genotypes/docker-compose.yml up -d data-split-service

Second, instantiate the server container:

docker compose -f ./docker/genotypes/docker-compose.yml up -d server

Then, you can build and start client containers:

docker compose -f ./docker/genotypes/docker-compose.yml up -d --scale client=2

The --scale flag will allow you to instantiate a number of client containers, in this case, 4.

License

This project is licensed under the MIT License