Normalizing flows

Bobby's fork of VishakhG's NF repo.

Two goals here:

• get it running for the 4 potentials from Rezendez et al. (originally in this repo)
• git it running fro two seemingly simple potentials. (FAILS to train :/ )

Setup

We'll use pyenv for managing different versions of python and venv for our python virtual envrionment.

Quick background and useful commands appear at the bottom of this readme.

The setup follows reference: https://www.freecodecamp.org/news/manage-multiple-python-versions-and-virtual-environments-venv-pyenv-pyvenv-a29fb00c296f/

Steps for using this repo:

1. Set pyenv envrionment to python 3.9.14 (assuming it is installed using. if not, install it using pyenv install 3.9.14)

    pyenv local 3.9.14
   

2. Initialize virtual envrionment in the .venv folder:

   python3 -m venv .venv
   

   in .venv/bin should be a copy of python3.9

3. Start virtual envrionment:

    source .venv/bin/activate 
   

   At this point, you should see you should see (venv) before your terminal, running which pip and which python should produce a path to the pip and python instances in .venv/bin/.

   Running python --version should produce 3.9.14.

   VS Code and Jupyter users may need to point them to the right interpreter. (.vinv/bin/python)

4. Install required packages from ./requirements.txt file:

    pip install -r requirements.txt
   

   NOTE: if a new packages is needed then use the pip install <package> (which calls .venv/bin/pip) to install it, and

5. Before pushing, if new required packages were installed, these need to be added to the repo and pushed. Run

    pip freeze > requirements.txt
   

   and push the new requirements.txt.

   Note that the .venv folder is gitignored and should not ship with the repo.

From VishakhG's Readme:

Attempting to implement the potential function experiments from:

Danilo Jimenez Rezende and Shakir Mohamed. Variational inference with normalizing
flows. In Proceedings of the 32nd International Conference on Machine Learning, pages
1530–1538, 2015.

Other reference:

Papamakarios, George, et al. Normalizing Flows for Probabilistic Modeling and Inference. Dec. 2019. arxiv.org, https://arxiv.org/abs/1912.02762v1.

To reproduce plots run exp/run_2d_potential_exp.sh or take a look at src/fit_flows.py.

Target densities, corresponding to the 4 potentials from the paper:

Samples from a 2-D diagonal gaussian passed through 32 learned Planar flows: