gpeyre/SinkhornAutoDiff

SinkhornAutoDiff - Python toolbox to integrate optimal transport loss functions using automatic differentiation and Sinkhorn's algorithm

Overview

Python toolbox to compute and differentiate Optimal Transport (OT) distances. It computes the cost using (generalization of) Sinkhorn's algorithm [1], which can in turn be applied:

• To optimize barycenters and their weights [2].
• To perform shape registration [9].
• As a loss between machine learning features [1].
• To perform density fitting for generative model [8] (see also image bellow).

Organization

• code/ contains the core routine to compute OT losses and their differentials.
• notebooks/ contains a collection of Jupyter notebooks that showcase simple usage of the toolbox.

Planed Features

• Classical Sinkhorn using matrix/vector multiplications [1].
• Acceleration for separable kernels (e.g. Gaussian kernels for images) [5].
• Log-domain stabilized Sinkhorn [7].
• Unbalanced transport (for several f-divergence fidelity) [7].
• Heavy-ball acceleration [10].
• Barycenters [7].
• Gromov-Wasserstein [4].
• Dictionary learning [3].

Installation

Current implementations are available using the following automatic-differentiation toolboxes:

• Theano
• TensorFlow
• Chainer

References:

[1] Marco Cuturi, Sinkhorn Distances: Lightspeed Computation of Optimal Transport, NIPS 2013

[2] N. Bonneel, G. Peyré, M. Cuturi. Wasserstein Barycentric Coordinates: Histogram Regression Using Optimal Transport. ACM Transactions on Graphics (Proc. SIGGRAPH 2016), 35(4), pp. 71:1–71:10, 2016.

[3] A. Rolet, M. Cuturi, G. Peyré. Fast Dictionary Learning with a Smoothed Wasserstein Loss. In Proc. AISTATS'16, pp. 630–638, 2016.

[4] G. Peyré, M. Cuturi, J. Solomon. Gromov-Wasserstein Averaging of Kernel and Distance Matrices. In Proc. ICML'16, pp. 2664–2672, 2016.

[5] J. Solomon, F. de Goes, G. Peyré, M. Cuturi, A. Butscher, A. Nguyen, T. Du, L. Guibas. Convolutional Wasserstein Distances: Efficient Optimal Transportation on Geometric Domains. ACM Transactions on Graphics (Proc. SIGGRAPH 2015), 34(4), pp. 66:1–66:11, 2015.

[6] M. Cuturi, M. Blondel, Soft-DTW: A Differentiable Loss Function for Time Series, ICML 2017.

[7] L. Chizat, G. Peyré, B. Schmitzer, F-X. Vialard. Scaling Algorithms for Unbalanced Transport Problems, Preprint Arxiv:1607.05816, 2016.

[8] Aude Genevay, Gabriel Peyré, Marco Cuturi, Sinkhorn-AutoDiff: Tractable Wasserstein Learning of Generative Models, Preprint Arxiv:1706.00292, 2017

[9] Francois-Xavier Vialard, Gabriel Peyré, Optimal Transport for Diffeomorphic Registration, Jean Feydy, Benjamin Charlier, MICCAI 2017.

[10] G. Peyré, L. Chizat, F-X. Vialard, J. Solomon. Quantum Optimal Transport for Tensor Field Processing. Preprint Arxiv:1612.08731, 2016.

Collaborators

• Gwendoline de Bie (ENS)
• Marco Cuturi (ENSAE)
• Jean Feydy (ENS)
• Aude Genevay (ENS)
• Gabriel Peyré (CNRS and ENS)
• Morgan Schmitz (CEA)
• Francois-Xavier Vialard (Paris-Dauphine)
• ...

Links

• Python Optimal Transport library
• Mathematical coffees: Optimal Transport Meets Machine Learning
• Soft Dynamic time warping [6].

Copyright (c) 2017 Noria's team