fastmachinelearning / qonnx

QONNX: Arbitrary-Precision Quantized Neural Networks in ONNX
https://qonnx.readthedocs.io/
Apache License 2.0
124 stars 39 forks source link

Add support for multiple default configurations #124

Open klassen9 opened 4 months ago

klassen9 commented 4 months ago

This PR addresses the problem of not being able to assign a default configurations of the same type for different node types. For example, one could not assign the default preferred_impl_style for node type "MVAU" as "hls" and "rtl" for node type "FMPadding" in FINN.

An example of a specialze layer config for FINN which uses the updated parser:

{
    "Defaults": {
        "preferred_impl_style": ["hls", ["MVAU"], "rtl", ["FMPadding"]]    
    }
}

The changes made to the parser do not break old configurations.