ZüNIS: Normalizing flows for neural importance sampling

ZüNIS (Zürich Neural Importance Sampling) a work-in-progress Pytorch-based library for Monte-Carlo integration based on Neural imporance sampling [1], developed at ETH Zürich. In simple terms, we use artificial intelligence to compute integrals faster.

The goal is to provide a flexible library to integrate black-box functions with a level of automation comparable to the VEGAS Library [2], while using state-of-the-art methods that go around the limitations of existing tools.

Installation

Using pip

As the library is not yet fully mature, we have not released it to the Python Package Index (PyPI). You can nevertheless install it with pip from this repository as follows:

 pip install 'git+https://github.com/ndeutschmann/zunis#egg=zunis&subdirectory=zunis_lib'

Setting up a development environment

If you would like to contribute to the library, run the benchmarks or try the examples, the easiest is to clone this repository directly and install the extended requirements:

# Clone the repository
git clone https://github.com/ndeutschmann/zunis.git ./zunis
# Create a virtual environment (recommended)
python3.7 -m venv  zunis_venv
source ./zunis_venv/bin/activate
pip install --upgrade pip
# Install the requirements
cd ./zunis
pip install -r requirements.txt
# Run one benchmark (GPU highly recommended)
cd ./experiments/benchmarks
python benchmark_hypersphere.py

Library usage

For basic uses, a RealNVP-based integrator is provided with default choices and can be created and used as follows:

import torch
from src.integration import Integrator

device = torch.device("cuda")


d = 2

def f(x):
    return x[:,0]**2 + x[:,1]**2

integrator = Integrator(d=d,f=f,device=device)
result, uncertainty, history = integrator.integrate()

The function f is integrated over the d-dimensional unit hypercube and

• takes torch.Tensor batched inputs with shape (N,d) for arbitrary batch size N on device
• returns torch.Tensor batched inputs with shape (N,) for arbitrary batch size N on device

A more systematic documentation is under construction here.