SplitChildLearningRateScales
Models & Types · AiDotNet.Models.Options
Per-attribute learning-rate multipliers applied to a Gaussian’s children immediately after it splits (#1835 excellence goal #4).
For Beginners
When a Gaussian splits into two, the children need different learning rates than their parent — smaller position steps (they’re where they need to be) and larger scale steps (they need to shrink to fit). Reference impls hard-code the paper values; here you can override any of them.
How It Works
Paper (Kerbl et al. 2023 §5.3): position LR decreases (children are now local — small movements matter more), scale LR increases (children are smaller and need faster updates to converge), others unchanged. Values persist on the child Gaussian and compound across second-generation splits.
Constructors
| Constructor | Summary |
|---|---|
SplitChildLearningRateScales() |
Creates a new instance with paper defaults. |
SplitChildLearningRateScales(SplitChildLearningRateScales other) |
Deep-copies every field from other. |
Properties
| Property | Summary |
|---|---|
double Color { get; set; } |
Color LR multiplier for split children. |
double Opacity { get; set; } |
Opacity LR multiplier for split children. |
double Position { get; set; } |
Position LR multiplier for split children. |
double Rotation { get; set; } |
Rotation LR multiplier for split children. |
double Scale { get; set; } |
Scale LR multiplier for split children. |