From WebML WG Teleconference – 18 Nov 2021:
RESOLUTION: Create a light-weight process to guide submitting new operator requests to WebNN
Recording this resolution as an issue so we remember to bootstrap this process.
In our discussions, the following ONNX guidelines were suggested to be adopted for WebNN:
- "Prefer static attributes over dynamic input values"
- "Shape inference logic should be included"
Concretely, we could implement this process as an issue/PR template.
Feel free to share best practices/guidelines here to be used as a basis for the first version.