Using custom com.microsoft ops from python

[brnorris03]

I would like to use the "com.microsoft" domain ops whose schemas are available in onnxruntime, e.g., "SoftmaxGrad", but cannot figure out how after looking at docs and the codebase. I am configuring with --enable_training --enable_pybind --enable_external_custom_op_schemas. The onnxruntime.training python package builds successfully.

I am new to this, apologies for the possibly very basic questions. Here is a simple test I tried first:

from onnx import checker, helper
opset_import = [
    helper.make_opsetid(domain="com.microsoft", version=1),
    helper.make_opsetid(domain="", version=17),
]

node = helper.make_node(
    op_type="SoftmaxGrad",
    name="/SoftmaxGrad_0",
    inputs=["dY", "Y"],
    outputs=["dX"],
    axis=-1,
    domain="com.microsoft",
    opset_imports=opset_import,
)
checker.check_node(node)

which fails with

Traceback (most recent call last):
  File "/proj/work/bnorris/mlir-compiler/ex.py", line 10, in <module>
    node = helper.make_node(
  File "/.../venv/lib/python3.10/site-packages/onnx/helper.py", line 173, in make_node
    node.attribute.extend(
  File "/.../venv/lib/python3.10/site-packages/onnx/helper.py", line 174, in <genexpr>
    make_attribute(key, value)
  File "/.../venv/lib/python3.10/site-packages/onnx/helper.py", line 960, in make_attribute
    raise ValueError(
ValueError: Could not infer the attribute type from the elements of the passed Iterable value.

I did check onnx.defs.get_all_schemas_with_history(), which does not include the custom ops.

I also attempted to use the custom op documentation generator, which relies on the schemadef package, but that did not appear to have been compiled into the python bindings. It seems to me that none of the python functions in onnxruntime/python/onnxruntime_pybind_schema.cc are being compiled, and I can't figure out what controls whether they are built or not.

The problem (taken from the documentation generation script):

from onnxruntime.capi.onnxruntime_pybind11_state import schemadef

ImportError: cannot import name 'schemadef' from 'onnxruntime.capi.onnxruntime_pybind11_state' (/.../venv/lib/python3.10/site-packages/onnxruntime/capi/onnxruntime_pybind11_state.so)

[xadupre]

Your model is invalid. You need to create a full model with make_model, make_graph, make_node... See https://onnx.ai/onnx/intro/python.html.

[brnorris03]

I don't have a model for this test, just a single node (but creating a model does not work since it fails at node creation). Can you elaborate why exactly I need a model to invoke check_node? I don't see this requirement mentioned anywhere in code or documentation.

The above test is based on https://github.com/onnx/onnx/blob/67c456ba4747412afb44158a1a889c0fc3349641/onnx/test/checker_test.py, which suggests I should be able to create a node and check it without inserting it into a full model.

Thanks!

[Craigacp]

Isn't the opset_imports a property of the model rather than a property of a node? So it's probably trying to convert the opset_imports argument into an ONNX attribute and failing.