It would be useful to have a reference implementation of the ONNX ops, to complement the existing spec.

The existing guidelines for adding a new op to the spec already recommends a reference implementation of the op in Python, and this has been followed for a number of recent ops. However, there are still two limitations that would be worth addressing.

• Many of the ops added to the spec earlier do not have reference implementations.
• The reference implementations are not available in a well-organized way that makes it easy for users to call them.

The proposal is to enable users to write code as below:

import onnx.opset15 as op

def my_function(x, y, bias):
   t1 = op.MatMul(x, y)  # calls operation MatMul (of opset 15) with inputs x and y
   t2 = op.Add(t1, bias)
   t3 = op.Elu(t2, alpha=0.4)  # calls operation Elu with input t2 and value 0.4 for attribute alpha.
   return t3

Specifically, this requires ensuring that all reference implementations use a consistent API, and type-representation for values (eg., numpy arrays for tensors), and that they be organized into a module/class like object that allows us to invoke the ops belonging to a specific opset version.