A Federated Learning Framework using Apache SINGA.
This framework provides a codebase for developing federated learning algorithm with Apache SINGA Deep Learning Framework. Fed-SINGA simulates the real world setting by a server-client structure, where each can be placed in a docker container.
The baseline method implements Fed-Avg algorithm and Secure Aggregation.
Install dependencies
pip install -r requirements.txtPrepare the data
# mnist
python -m src.client.data.download_mnist
# bank
# 1. download the data manually from https://archive.ics.uci.edu/ml/datasets/bank+marketing
# 2. put it under the /data folder
# 3. run the following command which:
# (1) splits the dataset into N subsets
# (2) splits each subsets into train set and test set (8:2)
python -m src.client.data.bank NTraining the model
# Server and clients in host machine
bash run.sh
# Server and clients in docker containers
docker-compose -f scripts/docker-compose.train_10.yml upGenerate the docker training scripts
python scripts/compose_setup.py -c 10 --mode trainUnit tests and smoke tests are provided to ensure the correctness of the implementation. Run following commands to check.
pytest --cov=src tests/