Normalizing Flows Tutorial

This is a tutorial on Planar Flows and their generalization to Sylvester Flows for AI Sweden's ctrl-allt-dela learning series.

Author: Bobby robert.bridges@ai.se

A few slides are here: https://docs.google.com/presentation/d/11r8P696WUBZubXXUJ8eGihV5Ji42xbBt/edit?usp=sharing&ouid=103663031827858753966&rtpof=true&sd=true

Repo Contents

├── README.md ## this file! ├── Ctrl-Allt-Dela-NF-talk-notes.pdf # remarkable tablet notes ├── learn_sylvester_flows.ipynb # this notebook implements a sylvester flow for sampling from an improper density. do it after the planar flows ├── maf_nf.ipynb # this notebook implements a masked autoregressive flow for synthesizing data ├── normalizing_flows.py # these are PyTorch classes of flows that are imported by the planer and sylvester flows notebooks ├── planar_nf.ipynb # this notebook implements a planar flow for sampling from an improper density ├── plotting_utils.py # thes are functions imported by the notebooks to plot stuff └── talk_info.md # talk abstract and references

Repository Virtual Environment Setup

We used 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:

1. Install python 3.12.0 via pyenv using pyenv install 3.12.0)

2. Set pyenv envrionment to python 3.12.0 using pyenv local 3.12.0

3. Initialize virtual envrionment in the .venv folder python3 -m venv .venv. Now, in the new folder .venv/bin should be a copy of python 3.12.0

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

4. Start virtual envrionment: source .venv/bin/activate

   At this point running which pip and which python should produce a path to the pip and python instances in .venv/bin/.

   Running python --version should produce 3.12.0.

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

5. We shipped our pip freeze output as the requirements_*.txt files file, so one can (try to) install all packages with pip install -r requirements.txt. If that is problematic, one can try cat requirements.txt | xargs -n 1 pip install following https://stackoverflow.com/questions/22250483/stop-pip-from-failing-on-single-package-when-installing-with-requirements-txt (For some reason to install scikit requires: python3 -m pip install scikit-learn)