glasflow is a Python library containing a collection of Normalizing flows using PyTorch. It builds upon nflows.

glasflow requires Python 3.11 or newer.

glasflow is available to install via pip:

pip install glasflow

or via conda:

conda install glasflow -c conda-forge

Important: glasflow supports using CUDA devices but it is not a requirement and in most uses cases it provides little to no benefit.

By default the version of PyTorch installed by pip or conda will not necessarily match the drivers on your system, to install a different version with the correct CUDA support see the PyTorch homepage for instructions: https://pytorch.org/.

To define a RealNVP flow:

from glasflow import RealNVP

# define RealNVP flow. Change hyperparameters as necessary.
flow = RealNVP(
    n_inputs=2,
    n_transforms=5,
    n_neurons=32,
    batch_norm_between_transforms=True
)

Please see glasflow/examples for a typical training regime example.

glasflow uses a fork of nflows which is included as submodule in glasflow and can used imported as follows:

import glasflow.nflows as nflows

It contains various bugfixes which, as of writing this, are not included in a current release of nflows.

There is also the option to use an independent install of nflows (if installed) by setting an environment variable.

export  GLASFLOW_USE_NFLOWS=True

After setting this variable glasflow.nflows will point to the version of nflows installed in the current python environment.

Note: this must be set prior to importing glasflow.

Pull requests are welcome. You can review the contribution guidelines here. For major changes, please open an issue first to discuss what you would like to change.

If you use glasflow in your work please cite our DOI. We also recommend you also cite nflows following the guidelines in the nflows readme.