DynamicShapeInfo

Models & Types · AiDotNet.Models

Describes which dimensions of a model’s input/output shapes are dynamic (variable at runtime).

How It Works

For Beginners: Most machine learning models have fixed input sizes (e.g., always 784 features). But some models can handle variable-sized inputs. For example, a model might accept any batch size, or a sequence model might handle different sequence lengths. This class describes which dimensions can vary and what their valid ranges are.

This follows the ONNX convention where -1 in a shape dimension means “variable at runtime”. For example, an input shape of [-1, 784] means “any number of samples, each with 784 features”.

Industry standards:

  • ONNX: Uses -1 for dynamic dimensions in shape arrays
  • TensorFlow Serving: Uses None/null for dynamic dimensions
  • TorchServe: Uses -1 for dynamic dimensions

Constructors

Constructor Summary
DynamicShapeInfo()

Properties

Property Summary
bool HasDynamicBatch { get; } Gets whether the batch dimension (index 0) is dynamic.
bool HasDynamicInput { get; } Gets whether any input dimensions are dynamic.
bool HasDynamicOutput { get; } Gets whether any output dimensions are dynamic.
int[] DynamicInputDimensions { get; set; } Gets the indices of input dimensions that are variable at runtime.
int[] DynamicOutputDimensions { get; set; } Gets the indices of output dimensions that are variable at runtime.
int[] MaxInputDimensions { get; set; } Gets the maximum allowed values for dynamic input dimensions.
int[] MinInputDimensions { get; set; } Gets the minimum allowed values for dynamic input dimensions.

Methods

Method Summary
bool IsValidShape(int[] concreteShape, int[] templateShape, int[] dynamicDimIndices, int[] minDimensions, int[] maxDimensions) Validates a concrete shape against a template shape, respecting dynamic dimensions.

Fields

Field Summary
DynamicShapeInfo None A shared instance representing no dynamic dimensions (all dimensions are fixed).