Federated Learning GMM

An implementation of the Gaussian Mixture Model according to federated learning paradigm

The aim of this project is demonstrating an effective implementation of the Gaussian Mixture Model (GMM) with Expectation-Maximization (EM) algorithm according to the vanilla federated learning paradigm as decribed in the paper Communication-Efficient Learning of Deep Networks from Decentralized Data.

The Gaussian Mixture Model is employed in unsupervised learning problems, especially in clustering tasks. This repository allows to execute alternatively a baseline local version of GMM and a federated distributed implementation of the same model, in order to compare their performance.

Parameters

Name	Description	Default	Baseline	Federated
--dataset	Name of the dataset.	blob	X	X
--components	Number of Gaussians to fit.	3	X	X
--init	Model initialization method: random or kmeans (over a 0.5% fraction of the dataset).	random	X	X
--seed	Number to have random consistent results across executions.	None	X	X
--samples	Number of samples to generate.	10000	X	X
--features	Number of features for each generated sample.	2	X	X
--soft	Specifies if cluster bounds are soft or hard.	True	X	X
--plots_3d	Specifies if plots are to be done in 3D or 2D.	False	X	X
--plots_step	Specifies the number of rounds or epochs after which saving a plot.	1	X	X
--epochs	Number of epochs of training.	100	X
--rounds	Number of rounds of training.	100		X
--local_epochs	Number of local epochs for each client at every round.	10		X
--K	Total number of clients.	100		X
--C	Fraction of clients to employ in each round. From 0 to 1.	0.1		X
--S	Number of shards for each client. If None data are assumed to be IID, otherwise are non-IID.	None		X

Datasets

Commands

Requirements

References

License

MIT