Codebase for "A deep cut into Split Federated Self-supervised Learning" This repository implements Momentum-Aligned contrastive Split Federated Learning (MonAcoSFL).

It is an extension of MocoSFL, which keeps the online/momentum client models aligned during parameter synchronizations. This stabilizes the training and yields vastly improved results in deeper cut-layers that are more communication-efficient.

Python > 3.7 with Pytorch, Torchvision

/run_monacosfl.py -- Source files for MonAcoSFL

/scripts -- Evaluation scripts used on a single-GPU machine

This repository is based on the original MocoSFL codebase. We would like to thank the authors for open-sourcing it.