ShapeRelationDiscovery

Helpers & Utilities · AiDotNet.NeuralNetworks

Recovers a layer’s symbolic shape relation by RUNNING it on probe shapes and fitting the simplest relation that reproduces every observation.

How It Works

Hand-writing an IShapeContract for every layer type does not scale and is not what “automatic” means: this codebase has ~199 layer types and only 6 had even the coarse relation kind declared, so annotation-only coverage would sit near zero indefinitely and each new layer would start uncovered. Discovery inverts that — every layer is covered the moment it exists, and a layer that changes its shape behaviour is caught without anyone remembering to update a declaration.

PROBES MUST VARY ONE AXIS AT A TIME or the fit is ambiguous rather than wrong. If every probe has Height equal to Width, Same(Height) and Same(Width) reproduce the observations equally well and there is no evidence to choose between them; the fitted relation would then be right on the probes and wrong on the first non-square input. ProbeShapes generates a discriminating set, and an axis that stays ambiguous is reported as such instead of being guessed.

A FIT IS EVIDENCE, NOT PROOF. Simplest-that-explains is an inductive step: a layer could follow the window formula on every probe and something else beyond them. That is exactly why discovery does not replace IShapeContract — where a layer declares one, the declaration is authoritative and discovery becomes an independent check on it. Where none is declared, a discovered relation is strictly better than the nothing that was there before.

Methods

Method Summary
IReadOnlyList<AxisFinding> Fit(IReadOnlyList<TensorAxis> inputAxes, IReadOnlyList<TensorAxis> outputAxes, IReadOnlyList<ValueTuple<int[], int[]>> observations) Fits a relation for every output axis from observed (input shape, output shape) pairs.
IReadOnlyList<int[]> ProbeShapes(IReadOnlyList<int> baseShape) Generates probe shapes that vary each axis independently, so fits are discriminating.