I’m trying to see if pyro would be good for solving inverse models degraded by noise. Specifically problems of the form
F = Hg + n
Where f is a recorded signal, g is the true object n is noise and H is is some corrupting measurement matrix.
As I think I understand it, using pyro I can build a forward model of HG + n and it’ll use pytorch backpropogation magic to iterate through to maximize likihood say. My question is will pyro handle this properly if H is a large matrix with only pseudo inverses available? Is large matrix inversion something that pytorch itself handles nicely?