Table of Contents

Enum CalibrationMethod

Namespace
AiDotNet.Enums
Assembly
AiDotNet.dll

Calibration methods for quantization - techniques to determine optimal scaling factors when converting high-precision models to low-precision formats.

public enum CalibrationMethod

Fields

Entropy = 3

Entropy-based calibration - uses KL divergence to minimize information loss. Most accurate but slowest. Best for when you need maximum model quality.

Histogram = 2

Histogram-based calibration - analyzes value distribution using percentiles. More robust than MinMax, handles outliers better. Slightly slower.

MSE = 4

Mean Squared Error (MSE) based calibration. Minimizes difference between original and quantized values.

MinMax = 1

Min-Max calibration - simple range-based scaling. Fast and effective for most cases. Good default choice.

None = 0

No calibration - uses default symmetric scaling. Fastest but may reduce accuracy. Use only if calibration data is unavailable.

Percentile = 5

Percentile-based calibration - uses percentiles to handle outliers. Good for data with extreme values that shouldn't affect scaling.

Remarks

For Beginners: Quantization compresses AI models by using smaller numbers (e.g., 8-bit instead of 32-bit). Calibration is the process of figuring out the best way to map large numbers to small numbers without losing too much accuracy.

Think of it like adjusting a thermostat - you need to find the right scale to represent temperatures accurately. Different calibration methods use different strategies:

  • None: No calibration, uses default scaling. Fastest but least accurate.
  • MinMax: Simple approach using the min/max values seen in data. Fast and usually good enough.
  • Histogram: Analyzes distribution of values to find better scaling. More accurate than MinMax.
  • Entropy: Uses information theory (KL divergence) to minimize information loss. Most accurate but slowest.
  • MSE: Minimizes the mean squared error between original and quantized values.
  • Percentile: Ignores outliers by using percentiles instead of absolute min/max. Good for noisy data.

For most cases, MinMax or Histogram provides a good balance of speed and accuracy.

Edit this page