POC combining Keras 3 and Spox for ONNX compatibility

[johacks]

I recently found Spox, a repository for easy building of ONNX graphs, and it reminded me quite a bit of Keras in its definition of models.

I was curious about how well it would integrate with Keras, so I made a proof of concept repo wrapping over Keras to support Spox. I was very happy with the result so I wanted to show it. I think it could also work well as a direct modification of Keras. The repo demonstrates main modifications that would be needed:

• Slight adjustment of Keras base Layer to add a ´onnx_symbolic_call´ method with ONNX build scope context manager.
• Adjust ops API to use corresponding Spox implementation when in that scope (avoids having an ONNX backend as such).
• Adjust Keras variables and symbolic tensors to provide a representation as a Spox variable.

I think that the main advantage is that it is much more explicit with operations used, and avoids typical headaches of automatic ONNX exports.