Wesley Maddox
Wesley Maddox
I'm still working on it - the variational approach is basically done, but I was waiting on the Gpytorch changes to finalize that bit before working on Pyro + NUTS....
I can start taking a look then. One issue that I remember seeing is related to both versions - how exactly does Pyro deal with the stochasticity in the log...
At least for FKL, it makes sense to use an MCMC approach to sample the latent GPs (as in the experiments of http://proceedings.mlr.press/v9/murray10a/murray10a.pdf) as we may not have a great...
Sorry, I lost track of this in the final set of NeurIPS updates. I was pretty close to finishing up but didn't have a chance to completely update to the...
Sorry, I lost track on this before fixing the documentation. Because there's a ton of old wacky attempts to get it working before the last couple of commits, should I...
what versions of botorch and gpytorch are you using?
not sure i follow but yes that's what i'd expect the code to be doing
Because gpytorch has lazy computation (and it looks like thop isn't even finding the computations), we'd expect minimal computation in the forwards pass. What does something like `likelihood(model(train_x)).log_prob(train_y)` cost?