I've just read the source code and it seems that the differentiation of Cholesky decomposition and solving triangular systems is directly supported and more efficient than I thought. Also, I didn't realize that solvers could solve for many b at once, so no inversion needed.
On Monday, November 7, 2016 at 12:36:01 PM UTC+1, Kiuhnm Mnhuik wrote: > > Hi, > I need to compute > > y' inverse(XX') y > > for many y (say 10000), where X may be large but XX' is small (say 50x50 > or 100x100). > I need to do this symbolically because I need the gradient of that > expression wrt a parameter X depends on. > Usually, I'd get L from the Cholesky decomposition, invert L, etc... I'd > invert L because I have many y, as I said. > How should I do this in theano for maximum efficiency? > -- --- You received this message because you are subscribed to the Google Groups "theano-users" group. To unsubscribe from this group and stop receiving emails from it, send an email to [email protected]. For more options, visit https://groups.google.com/d/optout.
