KairosMultiSizePatchLayer<T>
Layers · AiDotNet.NeuralNetworks.Layers
Kairos Mixture-of-Size patch embedder: emits N parallel patch-size paths through the SAME transformer backbone shape (numPatches varies per path but hiddenDim is fixed), then combines them via a learned router that weights each path per-input. This is the Mixture-of-Size analog of Mixture-of-Experts: the “experts” are alternative tokenization granularities rather than alternative FFNs.
Inheritance: LayerBase<T> → KairosMultiSizePatchLayer<T>
Implements: IShapeContract
How It Works
Forward:
This layer outputs a single summarized [B, hiddenDim] tensor per input, ready to feed a downstream transformer stack that treats each input as a single token (since multi-size pooling has already collapsed the sequence). For the full per-patch sequence interpretation, the caller should use only one patch size.
Constructors
| Constructor | Summary |
|---|---|
KairosMultiSizePatchLayer(int contextLength, int hiddenDim, int[] patchSizes) |
Initializes a new KairosMultiSizePatchLayer. |
Properties
| Property | Summary |
|---|---|
bool SupportsTraining { get; } |
Methods
| Method | Summary |
|---|---|
IReadOnlyList<ILayer<T>> GetSubLayers() |
|
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank) |
|
void ResetState() |
|
void UpdateParameters(T learningRate) |