GPUExplainerHelper<T>

Helpers & Utilities · AiDotNet.Interpretability.Helpers

Provides GPU acceleration for interpretability explainers.

Implements: IDisposable

For Beginners

This helper accelerates interpretability computations using GPU hardware. Many explanation methods (SHAP, Integrated Gradients, etc.) require computing many predictions or gradients - operations that are perfect for GPU parallelization.

The helper provides:

  1. Batch Prediction: Process many inputs simultaneously on GPU
  2. Parallel Coalition Processing: For SHAP-style algorithms
  3. GPU Matrix Operations: Fast linear algebra for solving attribution problems
  4. Automatic Fallback: Falls back to CPU if no GPU is available

Benefits of GPU acceleration for explainers:

  • 10-100x speedup for batch predictions
  • Enables real-time explanations for complex models
  • Makes ensemble/sample-based methods practical for large models

Usage:

Constructors

Constructor Summary
GPUExplainerHelper(int? maxParallelism) Initializes a new GPU explainer helper using CPU parallel processing.

Properties

Property Summary
bool IsGPUEnabled { get; } Gets whether GPU acceleration is available and enabled.
int MaxParallelism { get; } Gets the maximum parallelism level for CPU fallback operations.
object DeviceInfo { get; } Gets information about the GPU device.

Methods

Method Summary
GPUExplainerHelper<T> CreateCPUOnly(int? maxParallelism) Creates a CPU-only helper (no GPU acceleration).
GPUExplainerHelper<T> CreateWithAutoDetect() Creates a GPU explainer helper with automatic GPU detection.
List<Vector<T>> ComputePermutedPredictions(Func<Matrix<T>, Vector<T>> predictFunction, Matrix<T> data, int featureIndex, int numPermutations, int? seed) Processes feature permutations in parallel for permutation importance.
Matrix<T> BatchComputeGradients(Func<Vector<T>, int, Vector<T>> gradientFunction, Matrix<T> inputs, int outputIndex) Computes gradients for multiple inputs in parallel using backpropagation.
Vector<T> BatchPredict(Func<Matrix<T>, Vector<T>> predictFunction, Matrix<T> inputs) Processes predictions for multiple inputs in parallel.
Vector<T> ComputeCoalitionPredictions(Func<Matrix<T>, Vector<T>> predictFunction, Vector<T> instance, List<bool[]> coalitions, Matrix<T> backgroundData, int nBackgroundSamples) Computes predictions for multiple coalitions in parallel (for SHAP-style algorithms).
Vector<T> ComputeIntegratedGradientsParallel(Func<Vector<T>, int, Vector<T>> gradientFunction, Vector<T> input, Vector<T> baseline, int numSteps, int outputIndex) Computes Integrated Gradients using parallel path integration.
Vector<T> SolveWeightedLeastSquares(Matrix<T> X, Vector<T> y, Vector<T> weights) Solves weighted least squares using GPU-accelerated matrix operations.
void Dispose() Disposes GPU resources.