Finish HSGP #61
Finish HSGP #61
Conversation
|
In case it is useful: https://colab.research.google.com/drive/10FcpmOm6awyVcR-1DK-xzHTqneD3xfS6 |
|
yes thank you @junpenglao! Birthdays would make a great example |
| omega, phi, m_star = self._eigendecomposition(X, self.L, self.M, self.D) | ||
| psd = self.cov_func.psd(omega) | ||
| self.beta = pm.Normal(f"{name}_coeffs_", size=m_star) | ||
| self.f = pm.Deterministic( |
There was a problem hiding this comment.
Is it possible to pass dims for the variable?
There was a problem hiding this comment.
yes! thanks for catching this
pymc_experimental/gp/hsgp.py
Outdated
|
|
||
| def conditional(self, name, Xnew): | ||
| fnew = self._build_conditional(name, Xnew) | ||
| return pm.Deterministic(name, fnew) |
There was a problem hiding this comment.
Is it possible to pass dims to the variable?
|
I think I'll move HSGP over into to pymc-devs/pymc#6036. HSGP needs that spectral covariance stuff to work. Having a prereq PR in pymc for something in pymc-experimental is just weird anyway. This split I think was a bad idea in hindsight. I don't know how to make the reverse option work, moving the spectral covariance support into experimental, since you need to mess with the Covariance classes quite a bit. And I think that the HSGP method is on firm footing with the main GP stuff. It has a couple papers, and having used it some it really does work super well. |
| "dimension." | ||
| ) | ||
| try: | ||
| if len(m) != cov_func.D: |
There was a problem hiding this comment.
Do covariance functions have a D attribute?
There was a problem hiding this comment.
Is this supposed to be len(active_dims)?
| """A helper function which gives the approximate kernel or covariance matrix K. This can be | ||
| helpful when trying to see how well an approximation may work. | ||
| """ | ||
| X, _ = self.cov_func._slice(X) |
There was a problem hiding this comment.
_slice takes two positional arguments.
|
@bwengals I think you can ignore my comments -- looks like this depends on your PyMC PR being merged as well! |
The goal of this PR is to finish the implementation of the HSGP and make it generally usable, so tests and docs and all that. This is a function-space GP approximation that is interesting because it uses a fixed set of basis vectors -- but allows the kernel hyperparameters to be learned. The restriction is that the kernel must be stationary and you need it's power spectral density.
TODO list:
priormethodconditionalmethodgp.Latentandgp.Marginaldo