AxisRelation

Models & Types · AiDotNet.Interfaces

How ONE output axis is sized from the input axes and the declaring type’s own configuration.

How It Works

The symbolic half of the shape system. [TensorLayout] says WHICH axes a layer has and in what order; this says HOW BIG each output axis is. Roles alone cannot distinguish a stride-1 convolution from a stride-2 one — both declare [Batch, Channels, Height, Width] in and out — yet one halves the spatial extent. Without the relation, “automatic shape inference” can only propagate names.

A RELATION, NOT A CONSTANT. Every form below is evaluated against the actual input at resolve time and reads the declaring instance’s real options, so a layer built with stride: 2 reports the halving and one built with stride: 1 does not. That is the whole reason this is an interface rather than more attribute arguments: C# attributes take compile-time constants only, and kernel, stride, padding and scale factor are all constructor arguments. A declaration that cannot see the configuration it describes is a declaration that goes stale the first time someone passes a different stride.

Properties

Property Summary
Form Kind { get; } Which form this relation takes.
IReadOnlyList<AxisRelation> Factors { get; } The factors, for ProductOfRelations; empty for every other form.
IReadOnlyList<TensorAxis> Sources { get; } Input axes this output axis is computed from.
int Denominator { get; } Scale denominator, for Scaled.
int Dilation { get; } Window dilation, for Window.
int Kernel { get; } Window extent, for Window.
int Numerator { get; } Scale numerator, for Scaled.
int Padding { get; } Padding added to BOTH sides, for Window.
int Stride { get; } Window step, for Window.
int Value { get; } The size, for Fixed.
string Reason { get; } Why the size is unknown, for Unknown.

Methods

Method Summary
AxisRelation Affine(TensorAxis source, int numerator, int denominator, int offset) This output axis is an input axis scaled by numerator/denominator and then shifted by offset, which MAY be negative.
AxisRelation Fixed(int size) This output axis has a size the configuration fixes, whatever the input.
AxisRelation FromPort(int port, TensorAxis axis) This output axis copies an axis of a SPECIFIC input port.
AxisRelation Product(TensorAxis[] sources) This output axis is the product of several input axes — a flatten.
AxisRelation ProductOf(AxisRelation[] factors) This output axis is the product of several DERIVED relations — a flatten over axes that are each transformed first.
AxisRelation Same(TensorAxis source) This output axis copies an input axis unchanged.
AxisRelation Scaled(TensorAxis source, int numerator, int denominator) This output axis is an input axis scaled by numerator/denominator.
AxisRelation ScaledBy(AxisRelation source, int numerator, int denominator) This output axis is another RELATION scaled by numerator/denominator.
AxisRelation SumOf(AxisRelation[] terms) This output axis is the SUM of several relations — what a concatenation does to its joined axis.
AxisRelation Unknown(string reason) This output axis genuinely cannot be derived from the input shape.
AxisRelation Window(TensorAxis source, int kernel, int stride, int padding, int dilation) This output axis follows the sliding-window formula floor((in + 2*padding - dilation*(kernel-1) - 1) / stride) + 1.
bool TryResolve(IReadOnlyDictionary<TensorAxis, int> inputAxes, int size) Computes this axis’s size from SEVERAL input ports’ axis sizes.
bool TryResolve(IReadOnlyList<IReadOnlyDictionary<TensorAxis, int>> ports, int size) Computes this axis’s size from SEVERAL input ports’ axis sizes.
string ToString()