Open sarahboufelja opened 2 years ago
sorry for long time reply. Yes, regularization values should be chosen with care, as it depends closely to the nature of data(and cost) at hand. Yet, the behavior should be consistent between several runs with exactly the same data and regularization parameters. If it is not the case, could you set up a small running example, that does not require extra steps or data, we could work upon ?
Describe the bug
I am using the Sinkhorn and Sinkhorn with Group Lasso implementations in OT package to reproduce the results in this paper: "Optimal Transport for Domain Adaptation", by Nicolas Courty et al. However, if I run the same following code for a few times, I get inconsistent convergence results:
sometimes the same code on the same data, converges with no errors and sometimes the algorithm fails to converge. I did try different Reg rates but I don't want to increase the rate significantly, as ths would obviously lead to a uniform mapping. Is this a known convergence issue with the Sinkhorn implementation ? How to choose the right the reg. rate in your opinion? With Cross-validation?
To Reproduce
Steps to reproduce the behavior:
Screenshots
Code sample
Expected behavior
I ma expecting the Sinkhorn algorithm to consistently converge to the optimal coupling.
Environment (please complete the following information):
pip
,conda
): pip3 installOutput of the following code snippet:
Additional context