LearningRateSchedulerType
Enums · AiDotNet.LearningRateSchedulers
Enumeration of available learning rate scheduler types.
How It Works
Use this enum with the LearningRateSchedulerFactory to create
schedulers by type without having to reference the concrete classes directly.
Fields
| Field | Summary |
|---|---|
LearningRateSchedulerType Constant |
Constant learning rate (no decay). |
LearningRateSchedulerType CosineAnnealing |
Cosine annealing: smooth cosine-shaped decay. |
LearningRateSchedulerType CosineAnnealingWarmRestarts |
Cosine annealing with warm restarts (SGDR). |
LearningRateSchedulerType Cyclic |
Cyclic learning rate: oscillate between bounds. |
LearningRateSchedulerType Exponential |
Exponential decay: multiply LR by gamma every epoch. |
LearningRateSchedulerType Lambda |
Custom lambda function scheduler. |
LearningRateSchedulerType LinearWarmup |
Linear warmup followed by optional decay. |
LearningRateSchedulerType MultiStep |
Multi-step decay: multiply LR by gamma at specified milestones. |
LearningRateSchedulerType Noam |
The Noam schedule: linear warmup, then inverse-square-root decay. |
LearningRateSchedulerType NoamHoldAnnealing |
Noam with a hold phase: linear warmup, then a hold at the peak, then a power decay. |
LearningRateSchedulerType OneCycle |
One cycle policy: warmup then annealing. |
LearningRateSchedulerType Polynomial |
Polynomial decay: LR follows polynomial curve to end value. |
LearningRateSchedulerType ReduceOnPlateau |
Reduce on plateau: decrease when metric stops improving. |
LearningRateSchedulerType Sequential |
Sequential composition of multiple schedulers. |
LearningRateSchedulerType Step |
Step decay: multiply LR by gamma every step_size epochs. |
LearningRateSchedulerType TriStage |
Linear warmup, then a constant hold, then linear decay. |
int value__ |