Conv3D performance degradation after ONNX conversion #2303
Labels
bug
An unexpected problem or unintended behavior
pending on user response
Waiting for more information or validation from user
Describe the bug
A simple tensorflow model with Conv3D and pooling is 3.6x slower on CPU after converting to ONNX.
The same model with Conv3D replaced by Conv2D is 10x faster on CPU after converting to ONNX.
Urgency
If not resolved in the next 4-6 months, this bug will block the planned release of a TF to ONNX-converted model.
System information
To Reproduce
Test the 3D tensorflow model:
Convert the 3D model to ONNX:
Test the 3D ONNX model:
Test the 2D tensorflow model:
Convert the 2D model to ONNX:
Test the 2D ONNX model:
The text was updated successfully, but these errors were encountered: