ModelHyperparameterOptions
Options & Configuration · AiDotNet.Models.Options
Base class for the per-model-family hyperparameter options: the sizes and shapes that come from the paper a model was published in.
Inheritance: ModelOptions → NeuralNetworkOptions → ModelHyperparameterOptions
Known derived types: AdversarialImageEvaluatorOptions, AudioHyperparameterOptions, AudioNeuralNetworkOptions, AutoencoderOptions, BayesianNeuralNetworkOptions, ConvolutionalNeuralNetworkOptions, DeepBeliefNetworkOptions, DeepBoltzmannMachineOptions, DeepQNetworkOptions, DenseNetOptions, DetectionBackboneOptions, DocumentNeuralNetworkOptions, EfficientNetOptions, EmbeddingModelOptions, FeedForwardNeuralNetworkOptions, FinancialNeuralNetworkOptions, GRUOptions, GanOptions, GraphEncoderOptions, GraphGenerationModelOptions, GraphModelOptions, HTMNetworkOptions, HopeNetworkOptions, HyperbolicNeuralNetworkOptions, MCDropoutNeuralNetworkOptions, MixtureOfExpertsOptions<T>, MobileNetV2Options, MobileNetV3Options, NEATOptions, OccupancyNeuralNetworkOptions, OctonionNeuralNetworkOptions, PanopticSegmentationOptions, PhysicsInformedOptions, RecurrentNeuralNetworkOptions, ResNetOptions, ReservoirOptions, ResidualNeuralNetworkOptions, RestrictedBoltzmannMachineOptions, SegmentationModelOptions, SelfOrganizingMapNNOptions, SequenceModelOptions, SiameseNeuralNetworkOptions, SparseNeuralNetworkOptions, SpeakerRecognitionOptions, SpeechEmotionRecognizerOptions, SpikingNeuralNetworkOptions, TextConditionerOptions, VGGOptions, VariationalAutoencoderOptions, VideoHyperparameterOptions, VisionLanguageInputOptions, VisionTransformerOptions, VoxelModelOptions
For Beginners
Every model in this library ships with the settings from the research paper that introduced it, so you can use a model without configuring anything. This class holds the settings that apply to nearly every model, and the family classes that derive from it add the ones specific to a kind of model.
How It Works
Pattern. Properties here are non-nullable, and each concrete options class assigns
its published defaults in its parameterless constructor. This differs deliberately from
the nullable + GetEffectiveX() pattern used for infrastructure configuration
(telemetry, profiling, AutoML): null there means “decide at runtime from the data
or the hardware”, which is meaningful for a batch size and meaningless for a layer count.
Every property here must be read by the model that owns it. A property nobody reads is worse than no property, because it advertises configurability that does not exist. The options-surface ratchet test enforces this.
Properties
| Property | Summary |
|---|---|
double MaxGradNorm { get; set; } |
Gets or sets the maximum global gradient norm, above which gradients are rescaled during training. |