Contribution
It documents the Python Optimal Transport library as of early 2021 and gives it a citable reference. It states what the library solves, how it is engineered and maintained, and how it compares with three other OT packages. It proves nothing and runs no benchmark.
Key insight
The paper’s own pitch is breadth and reliability, not speed: “generic solvers that can be used for conducting novel fundamental research” (abstract, p. 1), tests on every contribution, and closed-form cases used to check correctness (p. 4). For deep-learning use of Sinkhorn losses the authors point elsewhere: GeomLoss “is currently a better alternative to POT for using Sinkhorn losses in deep learning architectures” (p. 3).
Assumptions
- Discrete measures throughout: histograms a, b and a cost matrix M (eq. 1, p. 3).
- Dependencies NumPy, SciPy, Cython over C++; Pymanopt and Autograd for some sub-modules; CuPy for GPU entropic OT (p. 5).
Key results
- Problems solved (§3, pp. 3–4): the Kantorovich linear program W_M(a, b) = min ⟨T, M⟩ over Π(a, b) (eq. 1); its entropic and squared-L2 regularizations; Wasserstein barycenters, entropic by Bregman projections or exact by LP (eq. 2, Sample 2). “ot.gromov and ot.partial contain respectively solvers for Gromov-Wasserstein and partial OT problems” (p. 4); ot.da for domain adaptation with a scikit-learn API; ot.optim for user-specified regularizers.
- Table 1 (p. 3), POT’s column: EMD, Sinkhorn, stabilized Sinkhorn, unbalanced/partial OT, regularized OT, Wasserstein barycenter, Gromov–Wasserstein, sliced Wasserstein and domain adaptation checked; GPU partial; multiscale EMD not. In the table only POT has Gromov–Wasserstein.
- Engineering (§4, pp. 4–5): PEP8; tests for every contribution “with a line coverage of 92% of the code”; continuous integration; “more than 20 collaborators”; “more than 25 examples”; the exact solver is a C++ one by Bonneel et al. (2011). A PyTorch backend was “currently underway” (p. 5).
- Uptake: “more than 5K downloads each month on PyPI” (p. 2).
- Others (§2, pp. 2–3): OTT is “still considered a preliminary version”; OTJulia wraps POT for exact EMD.
Claims
| id | claim | strength | support |
|---|---|---|---|
| C1 | POT implements the solvers listed in Table 1 | strong | the paper’s description; the code is public |
| C2 | All functions and classes are tested, 92% line coverage | moderate | stated (p. 4); not checked |
| C3 | POT is the broadest of the four packages compared | moderate | Table 1, the authors’ own comparison, as of 2021 |
Concepts
- Wasserstein distance — the optimal value of eq. (1) when M is a metric: “When the ground cost is a metric, the optimal value of the OT problem is also a metric” (p. 3).
- Transport plan — T ≥ 0 with T1 = a and Tᵀ1 = b (p. 3).
Connections
It packages methods the record holds as papers: Gromov–Wasserstein averaging (Peyré, Cuturi and Solomon 2016; LIT-1818) and the GW machinery behind network GW (LIT-1822). Later POT releases add fused GW (LIT-1813) and co-optimal transport (LIT-1804), which are not in this paper. HOT (LIT-1814) and AvAtar (LIT-1807) use their own code.
Bearing on the record
- CASE-185: the LIT overstates what this paper covers. The
exchange named POT as the tool for every transport arm of the case, “GW,
fused and partial GW, and co-optimal transport (
ot.coot)”. The 2021 paper names Gromov–Wasserstein and partial OT (p. 4), not fused GW, not partial GW, not co-optimal transport. The library has them now (checked in POT 0.9.7), so the case’s tooling is right; the citation that stands for it should be this paper for the library and the methods’ own papers, LIT-1813 and LIT-1804, for those arms, with the POT version stated. - No THEORY. The paper reports software and makes no claim about what is true of any system.
- Practice. Its content is tooling for machine-learning practice, a library and how to use it. If the anthology wants it, that is its question; the LIT carries anthology-candidate.
Limitations
- No benchmarks, no timings, no comparison of accuracy with other packages; Table 1 compares features only, as the authors saw them in 2021.
- It describes a moving target (“evolves quickly”, p. 5). Which functions exist depends on the version, which the paper’s citation does not fix.
Open questions
- None about the paper. For the record: which POT version a study used should be stated wherever POT is cited for a method this paper does not describe.