Design

It replicates the same Python API and the same C++ API to enable a smooth replacement.

ONNX, Concepts, link to protobuf

• Introduction to ONNX
• Uncompromising Objective
• Detailed Differences between onnx and onnx_light
• Protobuf format applied to ONNX
• ORT flatbuffer format: parallelization and alignment
• ModelProto creation and no-copy ownership
• Complex loading and saving scenarios

Library Split

• How the C++ libraries are split
• Linking onnx-light in C++

Kernels and Backend Tests

• C++ Kernels
• Using backend tests to evaluate a runtime
• Backend test-case coverage
• Runtime test coverage (onnxruntime and shape_inference)
• Custom kernels for ReferenceEvaluator

Shape Inference

• Shape inference

Fuzzing

• Fuzzing