Torchflows: normalizing flows in PyTorch

Torchflows is a library for generative modeling and density estimation using normalizing flows. It implements many normalizing flow architectures and their building blocks for:

• Easy use of normalizing flows as trainable distributions.
• Easy implementation of new normalizing flows.

Example use:

import torch
from torchflows.flows import Flow
from torchflows.architectures import RealNVP

torch.manual_seed(0)

n_data = 1000
n_dim = 3

x = torch.randn(n_data, n_dim)  # Generate some training data
bijection = RealNVP(n_dim)  # Create the bijection
flow = Flow(bijection)  # Create the normalizing flow

flow.fit(x)  # Fit the normalizing flow to training data
log_prob = flow.log_prob(x)  # Compute the log probability of training data
x_new = flow.sample(50)  # Sample 50 new data points

print(log_prob.shape)  # (100,)
print(x_new.shape)  # (50, 3)

Check out examples and the documentation, including the list of supported architectures.

Installing

We support Python versions 3.7 and upwards.

Install Torchflows via pip:

pip install torchflows

Install Torchflows directly from Github:

pip install git+https://github.com/davidnabergoj/torchflows.git

Setup for development:

git clone https://github.com/davidnabergoj/torchflows.git
cd torchflows
pip install -r requirements.txt

Citation

If you use this code in your work, we kindly ask that you cite the accompanying paper:

Nabergoj and Štrumbelj: Empirical evaluation of normalizing flows in Markov Chain Monte Carlo, 2024. arxiv:2412.17136.

BibTex entry:

@misc{nabergoj_nf_mcmc_evaluation_2024,
    author = {Nabergoj, David and \v{S}trumbelj, Erik},
    title = {Empirical evaluation of normalizing flows in {Markov} {Chain} {Monte} {Carlo}},
    publisher = {arXiv},
    month = dec,
    year = {2024},
    note = {arxiv:2412.17136}
}

Contributions

We warmly welcome all contributions and comments. Please do not hesitate to submit issues and pull requests.

Some options to start contributing include:

• Adding references to the documentation page for architecture presets.
• Implementing new normalizing flow architectures (see the developer guide).
• Adding more automated tests for numerical stability and optimization.
• Adding docstrings to undocumented classes.