Closed Giodiro closed 2 years ago
Hello @Giodiro , Sorry for the delay in answering. I just fixed the issue in the main branch now, it was an error in the formula for the gradient of the divide operation, in the case of vector variables. If you can use the main branch in your framework, please let us know if it works ok on your side as well.
Thanks @joanglaunes , I'm testing it now and it seems to be working perfectly! Thanks a lot, Giacomo
Hi,
Congratulations on releasing v2, the compilation speedups look great!
I've been working on updating falkon, and I think I can get rid of my custom branch since you've added inplace support which made me very happy:)
I've been encountering an issue with the backward pass for an rbf kernel with a different lengthscale for each dimension. The error only occurs when
d
is different thant
, but I think it should not matter. Code to reproduce and the error are belowwhich gives the following (truncated) exception:
Thanks again for your work!