This python package aims to provide an easy interface for integrating datasets using optimal transport (OT)- and Gromov-Wasserstein (GW)-based methods. We plan to extend the package beyond data integration, with additional OT-related unsupervised and semi-supervised methods. Warning: This package is still in an early stage. Feel free to open an issue in case of unexpected behvior.
Install from source (latest version, may be unstable)
git clone https://github.com/Risitop/transmorph
pip install ./transmorphInstall from PyPi (stable version)
pip install transmorphSee three example notebooks in examples/ directory.
This package offers four main integration techniques, two based on OT and two based on GW. Both OT and GW comes in two variants, balanced (similar to SCOT) and unbalanced, using a quadratic program in order to estimate data points weights. These weights are chosen so that the weighted Gaussian mixture distribution is close to be uniform over the dataset.
Assuming two numpy arrays X and Y representing source and target datasets, WOTi can be used in the following way. First, create a Woti object. The scale parameter adjusts kernel bandwidth, and needs some tuning according to cloud sparsity.
import transmorph as tr
X, Y = ... # datasets, np.ndarray
t = tr.Transmorph(method='ot')Then, simply apply the integration method to project X onto Y.
X_integrated = t.fit_transform(X, Y)https://www.biorxiv.org/content/10.1101/2021.05.12.443561v1
Work in progress.