Rémi Flamary

Hello, You do not compute the same thing. A and B in ot.emd are the weights on each sample and M is the cost matrices (here between samples with positions...

Hello, I have a start of implementation that uses the fast exact solver on cpu but with a seamless cpu/gpu wrapper for torch. The actual solving is not done on...

Should we normalize like that the M matrix inside emd then? This is weird because it seemed to me that when using np.inf in the M matrix the exact solver...

interesting bug, i will try to reproduce it on my machine and get back to you.

yes that's it good catch! basically we cannot solve the problem precisely if there exist two values different values in m such that m_1+_2=m_1 due to numerical precision. I dont'...

POT exact OT solver is a wrapper around the C++ LEMON solver and we modifying it is out of what we can do. In addition, it might require some kind...

you are right good example

Hello @patricieni, Thank you for your detailed answer, I was writing mine when yours appeared but you did a quicker and better job. I think we should add an additional...

Yes a PR would be nice. Basically we just need to add a small ```eps``` value in the divisions around the following lines https://github.com/rflamary/POT/blob/master/ot/bregman.py#L357 this eps=1e-16 by default should be...

This is a good point. Note that the Sinkhorn divergence in the paper below with exactly the same notation but different definition (and is studied theoretically) has the 1/2 scaling...