uot-bench

Core concepts

uot-bench is built around four abstractions that map onto the typical optimal-transport workflow.

The four objects

Generator  ──generates──▶  Problem  ──solver_inputs()──▶  BaseSolver
                                                               │
                                              Experiment ◀────┘ (wraps measurement)
                                                  │
                                           run_pipeline
                                                  │
                                           pandas.DataFrame

Class	Role
Problem	Holds one OT problem: two or more marginal BaseMeasures and cost function(s). Computes cost matrices on demand.
Generator	A factory that yields Problem instances. All hyper-parameters live in __init__; generate() takes nothing.
BaseSolver	Accepts marginals and cost arrays, returns a SolverOutput dict.
Experiment	Pairs a measurement function with a solver call. Used to abstract what to measure (time, precision, GPU usage) from how to run it.
run_pipeline	Sweeps an Experiment over every (problem, solver, param_set, fold) combination and returns a pd.DataFrame.

Canonical imports

from uot import (
    Problem, Generator,           # base classes — subclass these
    TwoMarginalProblem,           # concrete problem (2 marginals)
    BarycenterProblem,            # barycenter (N marginals + lambdas)
    BaseSolver, SolverConfig,     # solver infrastructure
    Experiment, run_pipeline,     # experiment runner
    BaseMeasure,                  # measure base class
    PointCloudMeasure,            # scattered points + weights
    GridMeasure,                  # regular grid + weights
)

Where to go next

• Writing a custom Problem
• Writing a custom Generator
• Writing a custom Solver
• Running an Experiment in Python
• Measures and cost functions
• CLI & YAML pipeline

This site is open source. Improve this page.