IPaperOptimizerVariant
Interfaces · AiDotNet.Interfaces
Implemented by models whose paper specifies different hyperparameters per size or configuration
variant, so the right PaperOptimizerAttribute can be selected.
Known implementations: FishSpeechBase<T>, HuBERT<T>, MusicGenModel<T>, SVTR<T>, Wav2Vec2<T>, WavLM<T>, WhisperModel<T>
For Beginners
Many models come in sizes — tiny, base, large, huge. Papers often train the big ones at a smaller learning rate, because a rate that suits a small model makes a large one diverge. This property tells the library which size this instance is, so it picks the matching row from the paper.
How It Works
A model with a single set of paper hyperparameters does not need this: one unkeyed
[PaperOptimizer] applies to every instance. Implement it only where the paper genuinely
differs by variant, for example InternImage-T versus InternImage-H.
Returning the variant as a string rather than the enum itself keeps this interface free of a
per-model generic parameter, and matches the attribute’s Variant, which is written as
nameof(SomeSizeEnum.Huge) so a rename stays a compile error.
Properties
| Property | Summary |
|---|---|
string PaperOptimizerVariant { get; } |
The variant key for this instance, matching a [PaperOptimizer(Variant = ...)] declaration. |