LayoutGraph<T>
Models & Types · AiDotNet.Document.GraphBased
LayoutGraph: graph-based document layout analysis after GLAM (Wang et al. 2023).
Inheritance: NeuralNetworkBase<T> → DocumentNeuralNetworkBase<T> → LayoutGraph<T>
Implements: ILayoutDetector<T>, IReadingOrderDetector<T>
For Beginners
LayoutGraph analyzes how document parts relate:
- Builds a graph from document structure
- Models reading order and containment
- Learns hierarchical relationships
- Predicts document element types and groupings
Key features:
- Hierarchical graph construction
- Spatial relationship modeling
- Reading order prediction
- Multi-level layout understanding
Example usage:
How It Works
LayoutGraph constructs and analyzes graphs from document layouts, where nodes
represent document elements and edges encode spatial relationships: as in GLAM, a node-feature input
[N, F] whose first four columns are each element’s normalized box is joined into a k-nearest-centre
page graph, graph convolutions propagate over it, and each node is classified with a softmax. It excels
at understanding hierarchical document structures.
Constructors
| Constructor | Summary |
|---|---|
LayoutGraph(NeuralNetworkArchitecture<T> architecture, LayoutGraphOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) |
Creates a LayoutGraph model using native layers for training and inference. |
LayoutGraph(NeuralNetworkArchitecture<T> architecture, string onnxModelPath, LayoutGraphOptions options, IGradientBasedOptimizer<T, Tensor<T>, Tensor<T>> optimizer, ILossFunction<T> lossFunction) |
Creates a LayoutGraph model using a pre-trained ONNX model for inference. |
Properties
| Property | Summary |
|---|---|
DocumentType SupportedDocumentTypes { get; } |
|
IReadOnlyList<LayoutElementType> SupportedElementTypes { get; } |
|
bool RequiresOCR { get; } |
|
int ExpectedImageSize { get; } |
|
int NodeDim { get; } |
Gets the node dimension. |
Methods
| Method | Summary |
|---|---|
Dictionary<string, Tensor<T>> GetNamedLayerActivations(Tensor<T> input) |
|
DocumentLayoutResult<T> DetectLayout(Tensor<T> documentImage) |
|
DocumentLayoutResult<T> DetectLayout(Tensor<T> documentImage, double confidenceThreshold) |
|
ModelMetadata<T> GetModelMetadata() |
|
ReadingOrderResult<T> DetectReadingOrder(DocumentLayoutResult<T> layoutResult) |
|
ReadingOrderResult<T> DetectReadingOrder(Tensor<T> documentImage) |
|
Tensor<T> EncodeDocument(Tensor<T> documentImage) |
|
Tensor<T> ForwardForTraining(Tensor<T> input) |
Training forward: runs the base training forward (which wires layer seeds and applies gradient checkpointing / weight streaming over the graph layers), then mean-pools every axis but the class axis so the per-node logits [numNodes, numClas… |
string GetModelSummary() |
|
void ValidateInputShape(Tensor<T> documentImage) |