REMatch Kernel overflow about dscribe

[lauri-codes]

Hi @Abdelrahmansameh,

With the REMatch kernel, you will have to be careful about large numbers ending up in the Sinkhorn algorithm. Depending on your chosen metric, you may need to for example normalize the input in some way. This is also mentioned in the REMatch tutorial.

With the SOAP descriptor and a linear metric in particular, just normalize the input vectors to unit norm, so add this step between the descriptor creation and the kernel calculation:

from sklearn.preprocessing import normalize

t = normalize(t)

This normalization step is not done by default by DScribe, because the need and best way to to it really depends on your kernel. E.g. with a Gaussian metric, this normalization is not necessary.

[Abdelrahmansameh]

Thank you @lauri-codes for the quick response! I will go ahead and do that.