This repo consists of a FEDn client code for KSO. To set up FEDn which includes a reducer, combiner, and database-related base services, clone the FEDn FEDn repository and follow the instructions in its readme. The code has been tested with V0.3.1 tag of FEDn using quay.io/minio/minio:latest for the minio base service.
Fedkso is a POC developed as part of the federated learning project funded by EOSC-Nordic. The POC involves training an object detection model in a federated learning setting using the baltic subsea movies dataset collected Koster Seafloor Observatory. More details about the dataset and experiment setup are in the repository.
The test dataset used for test and development consists of 3 hours of subsea video footage.
The annotations provide the bounding boxes of the sea life present on the movies classifying them into species. Quality of the dataset: testing, small, real-world, challenging, single class, class-imbalanced dataset.
For this POC we have used YOLOv5−tiny as our baseline model for object detection. The choice was dictated by considering conditions such as the architecture of the model. The architecture supports a relatively small available dataset, small number of classes, and facilitates fast training speeds on limited computational resources during test and development.
We use an open-source framework called FEDn to communicate and coordinate between the local model training clients and the central model averaging server. The FEDn architecture can be seen from here.
The easiest way to start clients for quick testing is by using Docker. We provide a docker-compose template for convenience.
docker-compose -f docker-compose.yaml -f private-network.yaml up