Numpy API for ONNX

Many users have difficulties to write onnx graphs. Many packages tries to symplify it either by implementing their own api very close to onnx operators (sklearn-onnx, tf2onnx, spox, onnx-script). This contribution tries a different approach by implementing a numpy API for ONNX. It does not cover everything numpy or ONNX can do but it can easily be used to define loss functions for example without knowing too much about ONNX.

Note

control flow

The first version (onnx==1.15) does not support control flow yet (test and loops). There is no easy syntax for that yet and the main challenge is to deal with local context.

Overview

• First examples with onnx-array-api
  • A loss function from numpy to ONNX
  • Eager mode
• First examples with onnxruntime
  • Example
  • Profiling
  • Benchmark
  • Plots