Extend function inliner to support schema-defined functions #6931
Conversation
Signed-off-by: gramalingam <grama@microsoft.com>
Signed-off-by: gramalingam <grama@microsoft.com>
Signed-off-by: gramalingam <grama@microsoft.com>
Signed-off-by: gramalingam <grama@microsoft.com>
Signed-off-by: gramalingam <grama@microsoft.com>
Signed-off-by: gramalingam <grama@microsoft.com>
|
I wonder if it would be useful for the inliner to also retain info about the original function in the expanded nodes. For example, this could be through incorporating additional ValueInfo in the nodes or by keeping the new nodes in a model-local function instead of the top-level graph? |
I did think if it would be useful to create a model-local function from the schema-function. I ended up doing the inlining as the simplest. |
| @@ -111,6 +112,22 @@ def test_selective_exclusion(self): | |||
| self.assertEqual(function_nodes[0].op_type, "Add") | |||
| self.assertEqual(function_nodes[1].op_type, "Mul") | |||
|
|
|||
| def test_schema_function_inlining(self): | |||
There was a problem hiding this comment.
Optional: Test when function opset versions do not match?
Signed-off-by: gramalingam <grama@microsoft.com>
Signed-off-by: gramalingam <grama@microsoft.com>
Description
The previous version of inliner supported inlining only for model-local functions. This PR extends the support for schema-defined functions as well.
Motivation and Context
The increasing focus on extending the ONNX standard via new function-ops makes it useful for users to be able to inline these functions.