PruningConfig
Options & Configuration · AiDotNet.Interfaces
Configuration for pruning operations.
How It Works
Layer-aware pruning: When used with models implementing ILayeredModel,
the CategorySparsityTargets property enables per-category sparsity levels.
For example, attention layers can be pruned less aggressively than dense layers.
Constructors
| Constructor | Summary |
|---|---|
PruningConfig() |
Properties
| Property | Summary |
|---|---|
Dictionary<LayerCategory, double> CategorySparsityTargets { get; set; } |
Per-category sparsity targets (category → sparsity). |
Dictionary<string, double> LayerSparsityTargets { get; set; } |
Per-layer sparsity targets (layer name → sparsity). |
SparseFormat OutputFormat { get; set; } |
Output sparse format for storage. |
SparsityPattern Pattern { get; set; } |
Sparsity pattern to use. |
bool FineTuneAfterPruning { get; set; } |
Whether to fine-tune after pruning. |
bool GradualPruning { get; set; } |
Whether to use gradual pruning (multiple iterations). |
bool LayerWiseSparsity { get; set; } |
Whether to apply different sparsity per layer (sensitivity-based). |
double InitialSparsity { get; set; } |
Initial sparsity for gradual pruning. |
double TargetSparsity { get; set; } |
Target sparsity level (0.0 to 1.0). |
int FineTuningEpochs { get; set; } |
Number of fine-tuning epochs after pruning. |
int PruningIterations { get; set; } |
Number of pruning iterations for gradual pruning. |
int SparsityM { get; set; } |
M value for N:M sparsity (group size). |
int SparsityN { get; set; } |
N value for N:M sparsity (zeros per group). |