pat-rig/csvae4counterfactuals

Seminar paper which presents the CS-VAE model (Klys et al., 2018) in the context of Counterfactual Recourse. Provides a walk-through and rationale for deriving the rather complicated loss function. Code for training a vanilla VAE on MNIST.

1) How to Run the Code?

You need to set up a particular environment before running the notebooks:

cd repo_path

conda env create -f environment.yml

source activate pDL

The model code as well as evaluation is stored in the jupyter notebooks

vae.ipynbcontains a vanilla VAE based on https://github.com/Atcold/pytorch-Deep-Learning/blob/master/11-VAE.ipynb

conv_vae.ipynb contains the superior extension of vae.ipynb which produces better generations through convolving z.

Hyper parameters, such as beta, number of layers and convolutional parameters need to be manipulated in-line.

2) Repo Structure

2.1) Executing either of the notebooks for the first time will create a data directory with the MNIST /data automatically. Do not use other preprocessed versions of MNIST.

2.2) /models contains checkpoints of small training runs on a local machine. Parameter settings can be found in the file name.

2.3) /fig_src contains samples and reconstructions for different models which are specified by the filename of the respective images.