ShapeInference

Helpers & Utilities · AiDotNet.NeuralNetworks

Computes output shapes from declared axis roles (TensorLayoutAttribute) plus declared axis relations (IShapeContract), without running a forward pass.

How It Works

The join between the two halves of the shape system. The layout attribute names a tensor’s axes so a raw int[] becomes {Batch: 1, Channels: 3, Height: 32, Width: 32}; the contract then says how each OUTPUT axis is sized from those. Neither half is sufficient alone — roles without relations cannot tell a stride-1 convolution from a stride-2 one, and relations without roles have no way to say which input axis a relation refers to.

Everything here declines rather than guesses. A type that is unannotated, whose declared layout does not accept the given rank, or whose contract contains an Unknown axis yields no shape at all. A shape system that returns a plausible answer when it does not know is worse than one that returns nothing, because callers cannot tell the two apart.

Methods

Method Summary
IEnumerable<TensorLayoutAttribute> InputLayouts(Type type) Declared input layouts, read through the open generic definition.
IEnumerable<TensorLayoutAttribute> OutputLayouts(Type type) Declared output layouts, read through the open generic definition.
IReadOnlyDictionary<TensorAxis, int> NameAxes(Type type, IReadOnlyList<int> shape) Names the axes of a concrete shape using a type’s declared INPUT layout.
ShapeRelationKind DeriveRelationKind(IReadOnlyList<OutputAxisContract> contract) Derives the coarse ShapeRelationKind implied by a symbolic contract.
bool ContractMatchesLayout(object instance, int inputRank, string mismatch) Checks that a type’s IShapeContract agrees with its declared output layout.
bool HasDeclaredOutputShapeContract(Type type) Reports whether a type’s effective OutputAxesFor implementation declares a probeable symbolic law rather than an explicit or conditional unavailable marker.
int[] InferOutputShape(object instance, IReadOnlyList<IReadOnlyList<int>> inputShapes) Computes the output shape a MULTI-INPUT type produces for the given input shapes.
int[] InferOutputShape(object instance, IReadOnlyList<int> inputShape) Computes the output shape a MULTI-INPUT type produces for the given input shapes.
int[] InferOutputShape(object instance, IReadOnlyList<int> inputShape, bool isBatched) Computes the output shape a type produces, saying whether inputShape INCLUDES a batch axis.