Introduction

This repository is the code corresponding to our project on ADVI (Automatic Differentiation Variational Inference) optimization method made for the MVA class : graphical models, discrete inference and learning

How to reproduce experiments

• download the porto taxi trajectories dataset available here
• replace FILE_NAME in porto_preprocess.py by the train.csv file path of the taxi dataset
• execute porto_preprocess.py (it can take a while as it goes through the entire dataset)
• In the taxi.ipynb notebook replace array_path by the path of the array created by porto_preprocess.py
• You can go through the notebook (we advise not to execute the notebook as it takes quite a long time to run)

Repository's structure

• advi.py : File containing the ADVI optimization object

• ppca.py : File containing probabilistic PCA implementation (actually there are 3 different versions, not all equivalent)

• gmm.py : File containing Gaussian Mixture model implementation

• porto_preprocess.py : File for preprocessing the taxi dataset

• toy_data.ipynb : Notebook for our experimentations on simple 2D data

• taxis.ipynb : Notebook for our experiment on the taxi_dataset

Contributors

• Gabriel Ben Zenou
• Quentin Macé
• Alexandre Selvestrel