This repository contains material developed as part of a signal processing master project at Aalborg University. The project focuses on accelerating Ensemble Density Propagation for training deep neural networks.
This repository contains material developed as part of a signal processing master project at Aalborg University.
The purpose of the project is to investigate methods for making Ensemble Density Propagation (EnDP) feasible for large DNNs, as this is not discussed in the original article.
Through an analysis of the EnDP framework, it was found that distributed computing might be a good solution to make the framework a viable option in DNNs with large input and many layers.
A distributed implementation of a convolution layer from the VGG-16 network is used to test how well the framwork distributes to a large network.
The repository has three main folders Function_files, Tests and Strato.
The Function_files folder contains various functions created as part of the project.
The Tests folder contains test scripts used in the project.
The Strato folder contains a guide to how the strato cloud is used with DASK and a ping script for testing proper connection to workers.