Hello everyone
I’m trying the forum out! I’ve been enjoying using PyMC but have run into an issue which I haven’t been able to resolve.
The following code minimally replicates my error:
from itertools import product
# columns are times, rows are wavelengths, entries are intensities
intensity_data = np.arange(10).reshape(2, 5)
y = intensity_data.ravel()
image_wavelengths = [700, 800] # / nm
timestep_range = [0, 1, 2, 3, 4]
wavelengths, steps = np.meshgrid(image_wavelengths, timestep_range)
x = np.asarray(list(product(timestep_range, image_wavelengths)))
with pm.Model() as model:
# Specify the covariance function.
size = pm.HalfCauchy('size', 1)
length = pm.Gamma('length', 2, 0.5)
cov_func = size ** 2 * pm.gp.cov.ExpQuad(2, ls=length)
# Specify the GP. The default mean function is `Zero`.
gp = pm.gp.MarginalApprox(cov_func=cov_func, approx='FITC')
Xu = pm.gp.util.kmeans_inducing_points(len(x) // 2, x)
sigma = pm.Exponential('sigma', 1)
likelihood = gp.marginal_likelihood('likelihood', X=x, Xu=Xu, y=y, noise=sigma)
And the error (upon running something like pm.find_MAP() or pm.sample()) is:
ValueError: Random variables detected in the logp graph: [size, length].
This can happen when DensityDist logp or Interval transform functions reference nonlocal variables.
Inference appears to be performed correctly when the covariance function is specified with constant values however this is not really what I need. Any ideas or suggestions would be really appreciated.
Cheers