Rasmusafj / 02460-NMF-with-GP-priors

Non-negative matrix factorization (NMF) with Gaussian Process priors for simulated Raman spectroscopy data.

02460-NMF-with-GP-priors

This repository contains the final product from DTU course 02460 - advanced machine learning, where we worked on Non-negative matrix factorization with Gaussian Process priors for raman spectroscopy. The group consists of David Frich Hansen, Rasmus Arpe Fogh Jensen, Peter Edsberg Møllgaard and Søren Emil Schmidt.

The main hand-in of the project was the article included in article/NMF-GPP-Raman.pdf

Code

The code for the project is divided into seperate files.

• core.py - Main file containing the code for the implementation of NMF-GPP, LS-NMF and the hamiltonian sampling scheme.

• raman_plots.py - Script used to generate the raman spectra plots of the article

• sampling.py - Generates three monte carlo chains and dumps them into the /chains/ folder.

• sampling_plots.py - Script used to generate the sampling plots of the article based on the generated chains from sampling.py hamiltonian sampling scheme

Setup

Run the following commands to setup the conda environment.

conda env create -f environment.yml
source activate gpp-nmf

While running the sampling.py script, we encountered a "chains not unique" bug. From search on the internet, we found that the current (only) way to solve the problem was to change the source code according to the this comment. This source code manipulation is required to run the sampling script.