IInitializationStrategy<T>

Interfaces · AiDotNet.Initialization

Defines a strategy for initializing neural network layer parameters.

Known implementations: InitializationStrategyBase<T>

For Beginners

This controls how the network sets up its initial weights.

Different strategies have different trade-offs:

  • Lazy initialization makes network construction fast (good for tests)
  • Eager initialization is the traditional approach (slightly slower construction)
  • FromFile loads pre-trained weights (for transfer learning)

How It Works

This interface allows control over when and how layer weights are initialized. Different strategies can be used for different use cases:

  • Lazy: Defer initialization until first Forward() call (fast construction)
  • Eager: Initialize immediately on construction (current behavior)
  • FromFile: Load weights from a file instead of random initialization
  • Zero: Initialize all weights to zero (useful for testing)

Properties

Property Summary
bool IsLazy { get; } Gets a value indicating whether this strategy defers initialization until first use.
bool LoadFromExternal { get; } Gets a value indicating whether weights should be loaded from an external source.

Methods

Method Summary
void InitializeBiases(Tensor<T> biases) Initializes the biases tensor with appropriate values.
void InitializeWeights(Tensor<T> weights, int inputSize, int outputSize) Initializes the weights tensor with appropriate values.