POT is an optimal transport library providing a collection of solvers for computing Wasserstein, Gromov-Wasserstein, and Fused Gromov-Wasserstein distances between probability distributions. It functions as a differentiable tensor framework that integrates with various tensor libraries to enable automatic differentiation and GPU acceleration.
The project is distinguished by its ability to align data distributions across different metric spaces by comparing internal relational structures rather than coordinates. It implements mathematical optimization algorithms as differentiable layers, allowing for gradient-based updates within neural network workflows.
The toolkit covers a broad range of capabilities, including domain adaptation for aligning source and target distributions, the computation of various barycenters for distributions and graphs, and the estimation of transport mappings. It also provides tools for graph analysis, such as subgraph matching and dictionary learning, as well as dimensionality reduction techniques that preserve Wasserstein distance structures.
Solvers are implemented through a backend-agnostic tensor interface to support high-performance hardware acceleration across multiple tensor frameworks.