The described error can be reproduced with the following example:
import spox.opset.ai.onnx.v19 as op19
import spox.opset.ai.onnx.v18 as op18
from spox import Tensor, Var, argument, build
import onnx
import numpy
import pytest
def test_if_adaptation_const():
sq = op19.const(1.1453, dtype=numpy.float32)
b = argument(Tensor(numpy.float32, ("N",)))
cond = op18.equal(sq, b)
out = op18.if_(cond,
then_branch=lambda: [sq],
else_branch=lambda: [sq])
build({"b": b}, {"out": out[0]})Spox (correctly) attempts to update the opset of the if node from 18 to 19, but in creating a singleton model fails to add the output of the const as an input of the singleton since this is not an "input" of the If node.
tests/test_adapt.py:131:
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
src/spox/_public.py:117: in build
return graph.to_onnx_model()
src/spox/_graph.py:412: in to_onnx_model
self.to_onnx(concrete=concrete),
src/spox/_graph.py:343: in to_onnx
node_protos = itertools.chain.from_iterable(self.get_adapted_nodes().values())
src/spox/_graph.py:281: in get_adapted_nodes
best_effort = adapt_best_effort(
src/spox/_adapt.py:142: in adapt_best_effort
adapted = adapt_node(
src/spox/_adapt.py:60: in adapt_node
onnx.checker.check_model(source_model, full_check=True)
_ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _ _
model = ir_version: 9
graph {
node {
input: "Equal_0_C"
output: "If_0_outputs_0"
name: "If_0"
op_type: "If"
..._type {
elem_type: 1
shape {
}
}
}
}
}
opset_import {
domain: ""
version: 16
}
full_check = True, skip_opset_compatibility_check = False
def check_model(
model: ModelProto | str | bytes | os.PathLike,
full_check: bool = False,
skip_opset_compatibility_check: bool = False,
) -> None:
"""Check the consistency of a model. An exception is raised if the test fails.
Args:
model: Model to check.
full_check: If True, the function also checks for shapes that can be inferred.
skip_opset_compatibility_check: If True, the function skips the check for
opset compatibility.
"""
# If model is a path instead of ModelProto
if isinstance(model, (str, os.PathLike)):
C.check_model_path(os.fspath(model), full_check, skip_opset_compatibility_check)
else:
protobuf_string = (
model if isinstance(model, bytes) else model.SerializeToString()
)
# If the protobuf is larger than 2GB,
# remind users should use the model path to check
if sys.getsizeof(protobuf_string) > MAXIMUM_PROTOBUF:
raise ValueError(
"This protobuf of onnx model is too large (>2GB). Call check_model with model path instead."
)
> C.check_model(protobuf_string, full_check, skip_opset_compatibility_check)
E RuntimeError: Nodes in a graph must be topologically sorted, however input 'Constant_0_output' of node:
E name: If_0_else_branch__Introduce_0_id0 OpType: Identity
E is not output of any previous nodes.
E
E ==> Context: Bad node spec for node. Name: If_0 OpType: If
../../micromamba/envs/spox/lib/python3.12/site-packages/onnx/checker.py:148: RuntimeError
Spox should either bail gracefully when it cannot adapt a node (this is a best effort thing anyway) or should support this behaviour.
The described error can be reproduced with the following example:
Spox (correctly) attempts to update the opset of the if node from 18 to 19, but in creating a singleton model fails to add the output of the const as an input of the singleton since this is not an "input" of the If node.
Spox should either bail gracefully when it cannot adapt a node (this is a best effort thing anyway) or should support this behaviour.