Pyro GP mcmc result missing and Markov Chain thinning

Misc.
1

I am trying to use MCMC function to infer a GP regression problem with the following code:

calculate computational time

from pyro.infer import MCMC, NUTS, Predictive

st = time.time()

pyro.clear_param_store()

#generate a RBF kernel
RBF_kernel = gp.kernels.RBF(input_dim=1)
RBF_kernel.set_prior(“variance”, dist.Uniform(torch.tensor(1.), torch.tensor(500)))
RBF_kernel.set_prior(“lengthscale”, dist.Uniform(torch.tensor(1.), torch.tensor(3000)))

#generate a GP gression model conditioned on RSL data
gpr = PSTHM.GPRegression_V(X, y, RBF_kernel,noise=torch.tensor(y_sigma**2),jitter=1e-5)
gpr = gpr.double()
hmc_kernel = NUTS(gpr.model)
mcmc = MCMC(hmc_kernel, num_samples=2500,warmup_steps=500)
mcmc.run()

The code works well and gives me a nice result. If I print kernel.variance, it is a random variable with the posterior distribution, here an output example:

However, when I ran a forward GP regression for my new X data like:
gpr.kernel.variance
tensor(129.2338)

x_test = torch.arange(-500,2025,5.,requires_grad=True)
y_mean2, y_var2 = gpr(x_test.double(), full_cov=True,noiseless=True)

The kernel variance is now like:
gpr.kernel.variance
tensor(480.2222, dtype=torch.float64, grad_fn=)

I am wondering why is that? And if I want to thining the mcmc samples, let’s say selecting the first point for every five point of Markov Chain, how can I self-define a new regression based on thinned mcmc samples?

2

By default, the forward method will use parameters drawn from approximated inference (SVI). For mcmc, you need to use Predictive as in this tutorial Inferences for Deep Gaussian Process models in Pyro | fehiepsi's blog

3

Hi, thanks for this informaiton! But the Predictive function only return back the mean prediction, how can I find the covariance function for all ensembles?

4

Preditive will return all sample/deterministic sites for you. Just follow the tutorial and add a sample statement for your covariance.