DOC: HSGP.prior_linearized example raises error

[wd60622]

Issue with current documentation:

I'm running into issue with the example of the HSGP.prior_linearized
https://www.pymc.io/projects/docs/en/stable/api/gp/generated/classmethods/pymc.gp.HSGP.prior_linearized.html#pymc.gp.HSGP.prior_linearized

There is a pymc.exceptions.NotConstantValueError raised on the following line

phi, sqrt_psd = gp.prior_linearized(Xs=Xs)

Traceback (most recent call last):
  File "~/.../pymc/gp/hsgp_approx.py", line 290, in prior_linearized
    Xs, _ = self.cov_func._slice(Xs)
  File "~/.../pymc/gp/cov.py", line 189, in _slice
    [xdims] = constant_fold([xdims])
  File "~/.../pymc/pytensorf.py", line 1225, in constant_fold
    raise NotConstantValueError
pymc.exceptions.NotConstantValueError

I am running it with the latest main branch.

Idea or request for content:

Maybe I am missing some context in the example that could be added within the page or the functionality has changed and the example should be removed

[twiecki]

CC @bwengals

[bwengals]

Looks like recent update to PyMC here broke this. We can't use constant_fold there when X is passed as a mutable variable, the use case for that example.

Previously I just had .eval() there, since the values of X can mutate but not the dimensions (and if they do, the warning will show). @ricardoV94, is the constant_fold needed here or is there another way?

Also, x_dims is only actually used here to maybe give a warning. This warning has come in handy for me from time, but would prioritize the HSGP example working over the warning. So another option is to get rid of this bit of code.

[ricardoV94]

You can put it in a try/except an skip the warning if it didn't manage to constant_fold. But it seems like it causes more trouble than it helps?

[ricardoV94]

Alternatively you could be more strict and add a PyTensor assert, and don't support the "partial use of columns" case at all? If you think that's very likely to be a bug in most cases?

[wmbelk]

FYI; When I changed cov.py line 188, from:
def _slice(self, X, Xs=None):
xdims = X.shape[-1]
if isinstance(xdims, Variable):
[xdims] = constant_fold([xdims])

To be:
def _slice(self, X, Xs=None):
xdims = X.shape[-1]
if isinstance(xdims, Variable):
[xdims] = constant_fold([xdims], raise_not_constant=False)

Overriding the default True on raise_not_constant... It seems to work.