Implement `OpFromGraph` `__eq__` and `__hash__`

[ricardoV94]

Description

This allows merging duplicated nodes as well as comparing graph equality.

import pytensor
import pytensor.tensor as pt
from pytensor.compile.builders import OpFromGraph
from pytensor.graph.basic import equal_computations

x = pt.scalar("x")
out1 = OpFromGraph([x], [x + 1])(x)
out2 = OpFromGraph([x], [x + 1])(x)

assert equal_computations([out1], [out2])

It should pass the assert. It fails because out1.owner.op == out2.owner.op evaluates to False. We can probably do something very similar to Scan:

pytensor/pytensor/scan/op.py

Lines 1254 to 1320 in 4b41e09

	def __eq__(self, other):
	if type(self) is not type(other):
	return False
	if self.info != other.info:
	return False
	if self.profile != other.profile:
	return False
	if self.truncate_gradient != other.truncate_gradient:
	return False
	if self.name != other.name:
	return False
	if self.allow_gc != other.allow_gc:
	return False
	# Compare inner graphs
	# TODO: Use `self.inner_fgraph == other.inner_fgraph`
	if len(self.inner_inputs) != len(other.inner_inputs):
	return False
	if len(self.inner_outputs) != len(other.inner_outputs):
	return False
	# strict=False because length already compared above
	for self_in, other_in in zip(
	self.inner_inputs, other.inner_inputs, strict=False
	):
	if self_in.type != other_in.type:
	return False
	return equal_computations(
	self.inner_outputs,
	other.inner_outputs,
	self.inner_inputs,
	other.inner_inputs,
	)
	def __str__(self):
	inplace = "none"
	if self.destroy_map:
	# Check if all outputs are inplace
	if sorted(self.destroy_map) == sorted(
	range(self.info.n_mit_mot + self.info.n_mit_sot + self.info.n_sit_sot)
	):
	inplace = "all"
	else:
	inplace = str(list(self.destroy_map))
	return (
	f"Scan{{{self.name}, while_loop={self.info.as_while}, inplace={inplace}}}"
	)
	def __hash__(self):
	return hash(
	(
	type(self),
	self._hash_inner_graph,
	self.info,
	self.profile,
	self.truncate_gradient,
	self.name,
	self.allow_gc,
	)
	)

[zaxtax]

Do we want to support Alpha equivalence where the following two are considered equal?

x = pt.scalar("x")
out1 = OpFromGraph([x], [x + 1])(x)
y = pt.scalar("y")
out2 = OpFromGraph([y], [y + 1])(y)

assert equal_computations([out1], [out2])

[ricardoV94]

Do we want to support Alpha equivalence where the following two are considered equal?

x = pt.scalar("x")
out1 = OpFromGraph([x], [x + 1])(x)
y = pt.scalar("y")
out2 = OpFromGraph([y], [y + 1])(y)

assert equal_computations([out1], [out2])

Those Ops are equivalent, but not the nodes, since one takes x, and the other takes y. So out2.owner.op == out1.owner.op should be True but equal_computations not. However the following should pass the assert:

x = pt.scalar("x")
out1 = OpFromGraph([x], [x + 1])(x)
y = pt.scalar("y")
out2 = OpFromGraph([y], [y + 1])(x)

assert equal_computations([out1], [out2])

The dummy variables used to define the inner graph are irrelevant.

equal_computations will do the correct thing, since it checks for equivalence in the whole node which means same root inputs + equivalent intermediate nodes (which includes inputs and Op)

[zaxtax]

Oh right, my mistake!

…

On Fri, 3 Jan 2025, 19:08 Ricardo Vieira, ***@***.***> wrote: Do we want to support Alpha equivalence <https://en.wikipedia.org/wiki/Lambda_calculus#Alpha_equivalence> where the following two are considered equal? x = pt.scalar("x")out1 = OpFromGraph([x], [x + 1])(x)y = pt.scalar("y")out2 = OpFromGraph([y], [y + 1])(y) assert equal_computations([out1], [out2]) Those Ops are equivalent, but not the nodes, since one takes x, and the other takes y. So out2.owner.op == out1.owner.op should be True but equal_computations not. However the following should be: x = pt.scalar("x")out1 = OpFromGraph([x], [x + 1])(x)y = pt.scalar("y")out2 = OpFromGraph([y], [y + 1])(x) assert equal_computations([out1], [out2]) The dummy variables used to define the inner graph are irrelevant. — Reply to this email directly, view it on GitHub <#1114 (comment)>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/AAACCUMBRSBBXEOAJTDUANT2I3GY3AVCNFSM6AAAAABTI3V3KWVHI2DSMVQWIX3LMV43OSLTON2WKQ3PNVWWK3TUHMZDKNRZGYZDKNBTGY> . You are receiving this because you commented.Message ID: ***@***.***>