IAdaptiveDistillationStrategy<T>
Interfaces · AiDotNet.KnowledgeDistillation.Strategies
Interface for adaptive distillation strategies that adjust temperature based on student performance.
Known implementations: AdaptiveDistillationStrategyBase<T>
For Beginners
Adaptive strategies dynamically adjust the temperature parameter during training based on how well the student is learning. This allows for more flexible knowledge transfer compared to fixed-temperature distillation.
How It Works
Key Concepts:
- Performance Tracking: Monitor student learning progress
- Temperature Adaptation: Adjust temperature based on sample difficulty or student confidence
- Per-Sample Adjustment: Different temperatures for different training samples
When to Use:
- Training data has varying difficulty levels
- Student performance is uneven across samples
- You want automatic temperature tuning instead of manual selection
Properties
| Property | Summary |
|---|---|
double AdaptationRate { get; } |
Gets the adaptation rate for performance tracking. |
double MaxTemperature { get; } |
Gets the maximum temperature value used for adaptation. |
double MinTemperature { get; } |
Gets the minimum temperature value used for adaptation. |
Methods
| Method | Summary |
|---|---|
double ComputeAdaptiveTemperature(Vector<T> studentOutput, Vector<T> teacherOutput) |
Computes the adaptive temperature for a specific sample. |
double GetPerformance(int sampleIndex) |
Gets the current performance metric for a specific sample. |
void UpdatePerformance(int sampleIndex, Vector<T> studentOutput, Vector<T> trueLabel) |
Updates the performance metric for a specific training sample. |