Summary of onnxruntime and onnxruntime-training API#

Module onnxcustom leverages onnxruntime-training to train models. Next sections exposes frequent functions uses to run inference and training with onnxruntime and onnxruntime-training.

Most of the code in onnxruntime is written in C++ and exposed in Python using pybind11. For inference, the main class is InferenceSession. It wraps C class C Class InferenceSession. The python class is easier to use. Both have the same name. It adds some short overhead but significant on small models such as a linear regression. It is recommended to use C classes in that case (inference, ortvalue, device).