This repo contains solutions to homework assignments for the course about Generative Neural Networks for the Sciences at the University of Heidelberg.
Co-authored with Agata Kaczmarek
Below are short descriptions of what each assignment was about.
Implement and use various models to learn and generate two-moons dataset from sklearn and digits dataset from sklearn:
- 2D histogram
- a single gaussian distribution
- a Gaussian mixture model (GMM)
- a kernel density estimator (KDE)
Maximum mean discrepancy (MMD) is used as a metric to compare the generated data with the original dataset.
Implement a simple autoencoder. Experiment with hyperparameters such as bottleneck size. Add MMD loss to make the codes dimension similar to gaussian distribution and sample from it to generate new data.
Use two-moons dataset from sklearn and digits dataset from sklearn.
Implement an Invertible Neural Network, called RealNVP. Experiment with different numbers of Coupling Blocks and other hyperparameters. Implement a conditional version of RealNVP and a version with an artificial bottleneck. Compare behaviour of these models.
Use two-moons dataset from sklearn and digits dataset from sklearn, but also digits from torch MNIST dataset.
Generate and work with simulated epidemiology data from a basic SIR model. Train conditional normalizing flow with summary network and analyze model quality: posterior calibration, confidence intervals and posterior predictive checks, but also sensitivity analysis and model misspecification.