Extend gp.Latent to multi-outputs
#7226
Conversation
|
|
hchen19
changed the title
gp.Latent to multi-outputsgp.Latent to multi-outputs
pymc/gp/gp.py
Outdated
| else: | ||
| f = pm.MvNormal(name, mu=mu, cov=cov, **kwargs) | ||
| f_single = pm.MvNormal(name, mu=mu, cov=cov, **kwargs) | ||
| f = pt.stack([f_single] * num_outputs, axis=0) if num_outputs > 1 else f_single |
There was a problem hiding this comment.
This approach won't wont, because what you're doing here is repeating a single GP, so each output is the same. We want each output to be as if they were independent draws from the same GP.
| def _build_prior(self, name, X, reparameterize=True, jitter=JITTER_DEFAULT, **kwargs): | ||
| def _build_prior( | ||
| self, name, X, num_outputs=1, reparameterize=True, jitter=JITTER_DEFAULT, **kwargs | ||
| ): | ||
| mu = self.mean_func(X) | ||
| cov = stabilize(self.cov_func(X), jitter) | ||
| if reparameterize: | ||
| size = np.shape(X)[0] | ||
| v = pm.Normal(name + "_rotated_", mu=0.0, sigma=1.0, size=size, **kwargs) |
There was a problem hiding this comment.
As a starting point, you'll need to make v here have size=(n_outputs, X.shape[0]) (or transpose of that). To check, you can draw some samples from the prior predictive.
import numpy as np
import pymc as pm
import arviz as az
x = np.linspace(0, 10, 100)
with pm.Model() as model:
cov = pm.gp.cov.ExpQuad(1, ls=1)
gp = pm.gp.Latent(cov_func=cov)
f = gp.prior("f", X=x[:, None], num_outputs=3)
idata = pm.sample_prior_predictive(samples=1)
f = az.extract(idata.prior, var_names="f")
# plot and you should see 3 different draws from the same GPYou'd have to do the same for the case reparameterize=False. But honestly, I think that is so rarely used it'd be just fine to just raise a NotImplementedError if n_outputs > 1.
There was a problem hiding this comment.
Thanks for your comments and suggestions, I will work on it according to your suggestions.
There was a problem hiding this comment.
glad I could help, looking forward to it
|
thanks for picking this up! |
| ): | ||
| mu = self.mean_func(X) | ||
| cov = stabilize(self.cov_func(X), jitter) | ||
| if reparameterize: | ||
| size = np.shape(X)[0] | ||
| size = (X.shape[0], n_outputs) if n_outputs > 1 else X.shape[0] |
There was a problem hiding this comment.
I updated the _build_prior() method by changing the size of v
| v = solve_lower(L, rxx) | ||
| mu = self.mean_func(Xnew) + pt.dot(pt.transpose(A), v) | ||
| v = solve_lower(L, rxx.T) | ||
| mu = self.mean_func(Xnew) + pt.dot(pt.transpose(A), v).T |
There was a problem hiding this comment.
Extend _build_conditional() to multiple outputs when marginal_likelihood() returns (n_outputs, X.shape[0]) size tensor
|
This is looking good, though I still need to check out your branch and make sure. It looks like pre-commit isn't happy though. This page should show you how to fix it. Once that's green we can trigger the tests. |
I fixed the pre-commit issue |
Description
add an argument
n_outputstogp.Latnetclass with default value = 1Related Issue
gp.Latentto allow for multiple outputs聽#7152Checklist
Type of change
馃摎 Documentation preview 馃摎: https://pymc--7226.org.readthedocs.build/en/7226/