Remove the automatic fix_float64=True functionality in ModelWrapper.transform which replaces float64 initializers with float32 ones - it will be an optional argument (default False) passed to the ModelWrapper contstructor instead.
This was a workaround for ONNX models that had invalid nodes according to previous versions of the standard, using two different datatypes for their inputs. This has been relaxed in recent versions of the standard and is no longer an issue for the most part.
Remove the automatic
fix_float64=Truefunctionality inModelWrapper.transformwhich replaces float64 initializers with float32 ones - it will be an optional argument (default False) passed to the ModelWrapper contstructor instead.This was a workaround for ONNX models that had invalid nodes according to previous versions of the standard, using two different datatypes for their inputs. This has been relaxed in recent versions of the standard and is no longer an issue for the most part.