Fix dim lengths created from coords or ConstantData

[michaelosthege]

In this comment we proposed to make dimensions created from coords non-resizeable, because it would prevent users from running into issues like #5181.

What needs to be done

In

pymc/pymc/model.py

Lines 1070 to 1076 in dd23c90

	def add_coord(
	self,
	name: str,
	values: Optional[Sequence] = None,
	*,
	length: Optional[Variable] = None,
	):

the set_coord method could take another kwarg resizable: bool=False or fixed: bool=True.

Then we can actively turn off re-sizing of dimensions that are created from ConstantData or coords.

[ricardoV94]

Let's remember to add #5181 as a test

[LukeLB]

Hi guys just been looking at how this can be implemented. Can I check my understanding first... So this line:

self._dim_lengths[name] = length or aesara.shared(len(values))

Is causing the problem due to the aesara.shared(len(values)) being resizeable. So the sugestion here is to create a new kwarg which we can use in an if statement, something like:

if fixed:
    self._dim_lengths[name] = length or aesara.tensor.constant(len(values))
else: 
    self._dim_lengths[name] = length or aesara.shared(len(values))

[michaelosthege]

@LukeLB yes, that's the idea

[canyon289]

@LukeLB Did you want to take a shot at fixing this, and if so when were you planning on doing so? I ask as this is holding up the Bambi migration to V4 so we want to get fixed soon here!

[LukeLB]

@canyon289 I've got time today so I'll have a crack at it.

[canyon289]

Sounds great! Looking forward to the PR and thank you. Let us know if you need any help.

[LukeLB]

@canyon289 just put in a WIP-PR as I'm heading out and was having some problems with locally failing tests.