AdaptiveDensificationSchedule
Models & Types · AiDotNet.NeuralRadianceFields.Data
Adaptive schedule (excellence goal): fires when the caller-observed loss plateaus below a threshold OR the gradient-norm variance settles — signals that the cloud has stabilized enough that split noise won’t destabilize training, while its remaining representational capacity has been exhausted. Reference impls have no equivalent — the fixed-interval schedule bakes in a fork’s assumptions about when densification will help. This adaptive path fires at the RIGHT time per-scene, not the wrong time on a global clock.
Implements: IDensificationSchedule
Constructors
| Constructor | Summary |
|---|---|
AdaptiveDensificationSchedule() |
Properties
| Property | Summary |
|---|---|
double LossPlateauThreshold { get; set; } |
Loss must improve by less than this fraction across the last window for the plateau trigger to consider firing. |
int MinInterval { get; set; } |
Minimum iterations between two adaptive fires (avoids over-densification during noisy early training when both signals may temporarily satisfy the threshold). |
int PlateauWindow { get; set; } |
Window (number of iterations) over which loss plateau is measured. |
Methods
| Method | Summary |
|---|---|
bool ShouldFire(int iteration, double lastLoss, double lastGradientNorm) |