PML-Project-24-25

Authors: Gaetano Tedesco, Patrick Jensen, Carsten Jørgensen

General

This repository contains the code for the project of the course Probabilistic Machine Learning at UCPH.

The project is divided in two parts:

• Exercise A: Implementation of the DDPM model and comparison of the models
• Exercise B: Gaussian Process

Installation

Before running the code, make sure you have the following packages installed:

• torch==2.3.0
• torchvision==0.18.0
• tocheval==0.0.7
• tqdm==4.66.5
• numpy==1.26.4
• matplotlib==3.9.2
• scipy==1.14.1
• sklearn==1.5.1
• pyro==1.9.1

In order to install the packages, you can download the requirements.txt file form the repo and use the following command: pip install -r requirements.txt

Exercise A

The files relative to exercise A are:

• ddpms.py: file containing all the implementation of the DDPM models produced
• utils.py: file containing all the utility functions used in the exercise (backbone networks definition, training loop, evaluation functions and support functions)
• model_training.ipynb: notebook executing the training loops for all the models
• model_comparison.ipynb: notebook executing the evaluation of the models (in our case generating the data for the models and computing FID)
• visual_comparison.ipynb: notebook executing the visual comparison of the models
• model_checkpoints: folder containing the checkpoints of the models trained. The checkpoints are meant as shortcut to run the model comparison (both numerical and visual) without retraining the models.
• visual: folder containing the visual comparison of the models.

Exercise B

There are 3 notebooks covering exercise B:

• fitting_MAP.final.ipynb
• fitting_MCMC_final.ipynb
• learning.final.ipynb

In fitting_MAP.final.ipynb we use gradient descent to find the hyperparameters of the kernel. In one of the last cells we output an array of test log-likelihood values.

All the code for using MCMC/NUTS to compute hyperparameters are in fitting_MCMC_final.ipynb.

The test log-likelihoods outputed in one of the last cells in fitting_MAP.final.ipynb needs to be pasted into fitting_MCMC_final.ipynb so we can create the boxplot comparing the two method.

fitting_MAP.final.ipynb also outputs kernel parameters. These are needed in one of the first cells in learning.final.ipynb for the integral constraint part.