GEMOptions<T>
Options & Configuration · AiDotNet.ContinualLearning.Strategies
Configuration options for Gradient Episodic Memory strategies.
For Beginners
These options control how GEM prevents forgetting by constraining gradients to not harm previous task performance.
Constructors
| Constructor | Summary |
|---|---|
GEMOptions() |
Properties
| Property | Summary |
|---|---|
MemorySamplingStrategy? MemorySamplingStrategy { get; set; } |
Memory sampling strategy for storing examples. |
ReplaySamplingStrategy? ReplaySamplingStrategy { get; set; } |
Replay sampling strategy for gradient computation. |
bool? NormalizeGradients { get; set; } |
Whether to normalize gradients before checking constraints. |
bool? UseAGEM { get; set; } |
Whether to use Averaged GEM (A-GEM) instead of full GEM. |
double? Margin { get; set; } |
Margin/epsilon for constraint violations. |
double? QPTolerance { get; set; } |
Tolerance for QP solver convergence. |
int? GradientBatchSize { get; set; } |
Batch size for gradient computation from memory. |
int? MaxQPIterations { get; set; } |
Maximum number of iterations for the QP solver. |
int? MemorySizePerTask { get; set; } |
Number of examples to store per task. |
int? RandomSeed { get; set; } |
Random seed for reproducibility. |