LiltLayoutEmbeddingLayer<T>

Layers · AiDotNet.NeuralNetworks.Layers

LiLT’s layout-flow embedding (Wang et al., ACL 2022; reference LiltLayoutEmbeddings). It runs these steps:

  • Six coordinate embeddings, each hidden / 6 wide: x0, y0, x1, y1, the height and the width.

The x table serves both horizontal edges, and the y table both vertical edges.

  • The six are concatenated, which gives back hidden whenever it divides by 6, as in every released LiLT;

otherwise it gives the largest multiple of 6 below it, and the linear map sizes itself to that.

  • A linear map to the layout flow’s width hidden / channelShrinkRatio.
  • A learned reading-order position embedding of that width is added.
  • LayerNorm.

Inheritance: LayerBase<T> → LiltLayoutEmbeddingLayer<T>

Implements: IShapeContract

For Beginners

This turns each word’s bounding box into a vector for LiLT’s small layout transformer, which runs beside the text transformer.

How It Works

The input is one box per token, [S, 4] or [B, S, 4], as integer (x0, y0, x1, y1) on the 0-1000 page grid. Coordinates are clamped into the tables, as the reference requires them to be.

Constructors

Constructor Summary
LiltLayoutEmbeddingLayer(int hiddenDim, int maxSequenceLength, int maxPosition2D, int channelShrinkRatio) Creates the embedding.

Properties

Property Summary
IReadOnlyList<LayerPort> InputPorts { get; } Generated from [TensorPort] declarations.
IReadOnlyList<LayerPort> OutputPorts { get; } Generated from [TensorPort] declarations.
bool SupportsTraining { get; }
int LayoutDim { get; } Width of the layout flow this embedding feeds: hidden / channelShrinkRatio.

Methods

Method Summary
IReadOnlyList<ILayer<T>> GetSubLayers()
IReadOnlyList<OutputAxisContract> OutputAxesFor(int inputRank)
LayerInputDomain GetInputDomain(int[] inputShape) Generated from the primary [TensorPort] declaration.
void ResetState()