eddardd / wbtransport

This repo contains the implementation of the Wasserstein Barycenter Transport proposed in "Wasserstein Barycenter Transport for Acoustic Adaptation" at ICASSP21 and "Wasserstein Barycenter for Multi-Source Domain Adaptation" in CVPR21

Hi there 👋 I'm Eduardo Montesuma

About Me 🤔

I am currently a PhD student @ Université Paris-Saclay/CEA-List. The subject of my thesis is on the contributions of Optimal Transport Theory for Transfer Learning, and to Machine Learning in general. My research interests are related to applied mathematics, statistics and machine learning. In the past years, I focused on Transfer Learning, Control Theory, Image Processing, Optimal Transport and Statistical Learning Theory.

How to find me

wbtransport's Issues

Questions about the paper notations

Thank you for sharing your work here! I have some questions about several details of the paper:

The notation in Eq.(8) is quiet confusing. Does $y_{s_k}$ for $k=1,...,N$ above Eq.(8) mean that we combine the labels of all the $n_k$ samples in a certain source domain $k$ as a label vector $y_{s_k} \in \mathbb{R}^{n_k}$, and concatenate all the label vector for different domains together. And in Eq.(8), what is the meanings of the subscripts $j$ and $i$? In Eq(8), $s_k$ is on the superscript, but on the subscript in the notation above Eq.(8)? It's quiet confusing. Could you please provide more clarifications about this?

No Line Search Implementation?

Hi,

I notice there are several lines of code for line search in the function of Sinkhorn Barycenter:
https://github.com/eddardd/WBTransport/blob/main/msda/barycenters.py#L117-L119
it seems you didn't append the implementation of the naive_line_search in this codebase. Could you provide this implementation? Did you find alpha = 1 is optimal?

question about eq10

Hi, dear author, thanks for your great job.

I have a big question about the following equation, since this step aims to calculate the source domain aggregation barycenter, why the coupling γ space is (μ_s, μ_t)? I think it should be (μ_s, μ_b) ?