← The glossary

computation graph

To begin with

Training code mixes a lot of things together. To run a model elsewhere, it can be written down like a recipe: a series of connected operations.

The facts

A computation graph describes operations (multiplying matrices, adding, normalising…) and the data flowing between them, with explicit inputs and outputs. ONNX stores a model in this form; the weights become constants or referenced files. A compatible engine, such as ONNX Runtime, reads the graph and runs it on the available hardware.

Key takeaway

The graph is the recipe for the computation. On its own it is not enough: you need an engine that can run each of its operations.

The stories about it

The notebook opens soon.The first stories are on their way.