Experiments: Consistency Models for Posterior Approximation

Setup

We provide an Apptainer definition file cmpe-env-cuda.def, that installs all necessary dependencies in an Apptainer container. If no GPU support is required, the cmpe-env.def creates a container without CUDA support. The container can be built using the following command:

$ sudo apptainer build cmpe-env.sif cmpe-env[-cuda].def

After the container is build, the contained Python can be used in the following way:

apptainer exec --bind /path/to/cmp path/to/cmpe-env.sif python <filename>

Depending on your file system structure, the --bind option might not be necessary.

A JupyterLab server in the environment can be started with:

apptainer exec --bind /path/to/cmp path/to/cmpe-env.sif jupyter lab --no-browser

Experiments 1-3

The low-dimensional experiments are located in experiments/benchmarks. Note that for producing a reference posterior for GMM, a working Stan installation (accessible via cmdstanpy) is required.

Experiment 4

The Bayesian Denoising experiment is located in experiments/bayesian_denoising. See the corresponding README for details on how to run the experiment.

Experiment 5

The tumor model experiment is located in experiments/tumor_model. The experiment builds on the PyABC implementation (Link), which contains details about the simulator and data.