`OpFromGraph` difficult to subclass due to not allowing shared variables to be positioned explicitly as an input

[ricardoV94]

Description

The changes introduced in aesara-devs/aesara#902 make it hard to safely subclass OpFromGraph with initialization kwargs.

Here is an example:

import pytensor
import pytensor.tensor as pt
from pytensor.compile.builders import OpFromGraph

class OpFromGraphSub(OpFromGraph):
    
    def __init__(self, *args, kwarg, **kwargs):
        self.kwarg = kwarg
        super().__init__(*args, **kwargs)


x = pytensor.shared(2)
y = pytensor.shared(3)

out = OpFromGraphSub([], [x + 1], kwarg="test")()
out.owner.op(y) 

# TypeError: OpFromGraphSub.__init__() missing 1 required keyword-only argument: 'kwarg'

One would have to subclass made_node for this to work correctly, but that's highly non-trivial.

[ricardoV94]

PS: Shared variables are a mistake

[ricardoV94]

The problem here is that the initialization of OpFromGraph prohibits explicit shared variables as inputs, but internally it still builds an Apply that has them as explicit inputs (much like Scan does, although scan allows shared variables as explicit inputs).

The change in #473 allows one to recreate an existing OpFromGraph node where the shared variable is changed, sometimes by non-shared variables (desirable), but that requires that the Op itself be recreated because OpFromGraph forces shared variables to be the rightmost input (undesirable) of the respective Apply nodes.

pytensor/pytensor/compile/builders.py

Lines 823 to 858 in 9df55cc

	if not all(inp_s == inn_s for inn_s, inp_s in inner_and_input_shareds):
	# The shared variables are not equal to the original shared
	# variables, so we construct a new `Op` that uses the new shared
	# variables instead.
	replace = dict(
	zip(self.inner_inputs[num_expected_inps:], new_shared_inputs)
	)
	# If the new shared variables are inconsistent with the inner-graph,
	# such errors should arise in this step
	new_inner_outputs = clone_replace(
	self.inner_outputs, replace=replace, copy_inputs_over=True
	)
	# It's possible that the new shared variable inputs aren't actually
	# shared variables. When they aren't we need to add them as new
	# inputs.
	unshared_inputs = [
	inp for inp in new_shared_inputs if not isinstance(inp, SharedVariable)
	]
	new_inner_inputs = self.inner_inputs[:num_expected_inps] + unshared_inputs
	new_op = type(self)(
	inputs=new_inner_inputs,
	outputs=new_inner_outputs,
	inline=self.is_inline,
	lop_overrides=self.lop_overrides,
	grad_overrides=self.grad_overrides,
	rop_overrides=self.rop_overrides,
	connection_pattern=self._connection_pattern,
	name=self.name,
	**self.kwargs,
	)
	new_inputs = (
	list(non_shared_inputs) + unshared_inputs + new_op.shared_inputs
	)

Imagine we had an OpFromGraph whose node inputs where [non_shared_1, shared_1, shared_2], and we want to replace shared_2 by a non shared variable. The new signature would have to artificially be [non_shared_1, non_shared_2, shared_1] which requires building the internal fgraph.