Table of Contents

Class CompressionAnalyzer<T>

Namespace
AiDotNet.ModelCompression
Assembly
AiDotNet.dll

Analyzes model weights to determine optimal compression strategies.

public class CompressionAnalyzer<T>

Type Parameters

T

The numeric type used for calculations (e.g., float, double).

Inheritance
CompressionAnalyzer<T>
Inherited Members

Remarks

CompressionAnalyzer examines model weight distributions to recommend the best compression technique and hyperparameters. It analyzes properties like weight sparsity, magnitude distribution, and redundancy to make informed recommendations.

For Beginners: Before compressing a model, it helps to understand its weights.

This analyzer looks at your model's weights and answers questions like:

  • Are many weights already close to zero? (Good for pruning)
  • Are weights clustered around certain values? (Good for quantization)
  • What's the distribution of weight values? (Affects all techniques)

Based on this analysis, it recommends:

  • Which compression technique to use
  • What settings (hyperparameters) to use
  • What compression ratio to expect

This helps you make informed decisions without trial-and-error.

Constructors

CompressionAnalyzer(double, int)

Initializes a new instance of the CompressionAnalyzer class.

public CompressionAnalyzer(double nearZeroThreshold = 0.01, int histogramBins = 256)

Parameters

nearZeroThreshold double

Threshold for considering a weight as "near zero" (default: 0.01).

histogramBins int

Number of bins for histogram analysis (default: 256).

Methods

Analyze(Vector<T>, bool)

Analyzes model weights and returns compression recommendations.

public WeightAnalysisResult<T> Analyze(Vector<T> weights, bool isConvolutional = false)

Parameters

weights Vector<T>

The model weights to analyze.

isConvolutional bool

Whether the weights are from convolutional layers (default: false).

Returns

WeightAnalysisResult<T>

Analysis results with compression recommendations.

Remarks

For Beginners: Pass your model weights to get compression recommendations.

The isConvolutional parameter helps optimize recommendations:

  • Convolutional layers: typically use lower pruning rates (60-70%)
  • Fully-connected layers: can use higher pruning rates (90-95%)

This is because convolutional layers have more structured, important patterns.

GenerateReport(WeightAnalysisResult<T>)

Generates a detailed analysis report.

public string GenerateReport(WeightAnalysisResult<T> result)

Parameters

result WeightAnalysisResult<T>

The analysis result to report on.

Returns

string

A formatted string containing the analysis report.

Edit this page