LayerShape
Structs · AiDotNet.NeuralNetworks.Layers
A layer’s shape, in which individual axes may be dynamic and may be named.
Inheritance: ValueType → LayerShape
Implements: IEquatable<LayerShape>
For Beginners
Some parts of a tensor’s shape are known in advance (how many features a layer outputs) and some are not (how many words are in this particular sentence). This type keeps track of which is which, gives the unknown ones names so two of them can be declared equal, and refuses to hand out a number for an axis nobody knows yet.
How It Works
A plain int[] cannot say “this axis is fixed and that one is not”, so every layer whose
output length depends on its runtime input had to either invent a value or refuse to resolve.
Inventing is what went wrong repeatedly: PositionalEncodingLayer declared its encoding
table size, CrossAttentionLayer declared a defaulted 64, and
TransformerDecoderLayer declared whatever its first input happened to be and then went
stale. Each fabricated number was then read as fact by parameter-vector slicing, chain
resolution and ONNX export.
Where this sits relative to other frameworks. Keras represents an unknown axis as
None and ONNX gives it a name; PyTorch has no static shape contract at all. Two things
here go past that:
Constructors
| Constructor | Summary |
|---|---|
LayerShape(int[] axes) |
Creates a shape from axis sizes, using Dynamic for unknown axes. |
LayerShape(int[] axes, string[] names) |
Creates a shape whose dynamic axes carry names. |
Properties
| Property | Summary |
|---|---|
bool HasDynamicAxis { get; } |
true when at least one axis is dynamic. |
bool IsFullyConcrete { get; } |
true when every axis has a known size. |
int Item { get; } |
|
int Rank { get; } |
Number of axes. |
Methods
| Method | Summary |
|---|---|
bool Equals(LayerShape other) |
|
bool Equals(object obj) |
|
bool TryGetConcrete(int[] concrete) |
Gets the concrete axis sizes, or returns false if any axis is dynamic. |
int GetHashCode() |
|
int[] RequireConcrete(string reason) |
Gets the concrete axis sizes, throwing when any axis is dynamic. |
int[] ToArray() |
The axis sizes as a plain array, with Dynamic for unknown axes. |
string NameOf(int axis) |
The name of an axis, or null when it is unnamed. |
string ToString() |
Fields
| Field | Summary |
|---|---|
int Dynamic |
Marks an axis whose size is not known until a real input arrives. |