Handle constant op variants (#3985) #4194
Conversation
| @@ -287,7 +287,7 @@ class ShapeInferenceImplBase { | |||
| has_experimental_op = true; | |||
| } else if (n.op_type() == "Constant" && n.output().size() == 1) { | |||
| for (const auto& attr : n.attribute()) { | |||
| if (attr.name() == "value") { | |||
| if (attr.name() == "value" || attr.name() == "value_ints") { | |||
There was a problem hiding this comment.
Thanks for creating the PR. Unfortunately, I don't think this will be enough to fix the issue. When the name is "value_ints", the attribute type is a list of ints, and not a tensor. So, the lines below also need to be extended to handle it correctly.
There was a problem hiding this comment.
Thanks for the suggestions! If i am correct, I need to add similar else if block which handles list of ints right?
There was a problem hiding this comment.
Yes. The proto fields are as defined here:
Line 168 in 5b1346e
In addition, there is a need to convert the value into a TensorProto.
There was a problem hiding this comment.
Here is how onnxruntime does it: we need something similar: https://github.com/microsoft/onnxruntime/blob/4c37f15c1ba3e24e0c7e2110f34ab410fadb2340/onnxruntime/core/framework/tensorprotoutils.cc#L821
|
This has been completed by a different PR. |
Fixes #3985