Gradient#

onnx-light implements reverse-mode automatic differentiation as a graph transformation. Gradient builders registered for individual operators produce the nodes required to propagate output gradients back to graph inputs while preserving the original forward graph.

The generic gradient interfaces live in lib_onnx_core and the ONNX operator implementations in lib_onnx_gradient. Applications select the inputs and outputs to differentiate, then receive an ONNX graph that can be optimized and executed through the same pipeline as any other model.

See the gradient training example for an end-to-end use case.

API reference#