LEOModel<T, TInput, TOutput>

Models & Types · AiDotNet.MetaLearning.Models

LEO model for few-shot classification with latent space optimization.

Implements: IModel<TInput, TOutput, ModelMetadata<T>>

For Beginners

After LEO adapts to a new task by optimizing in latent space, this model stores:

How It Works

The adapted state of LEO for one task: its own copy of the feature encoder, the linear softmax classifier LEO generated and adapted - one weight row per class - and the adapted latent codes. Predict returns class probabilities per example, [rows, NumClasses], for Tensor and Matrix outputs - a class the task did not contain gets probability zero - and the most probable class of each example for a Vector output.

Constructors

Constructor Summary
LEOModel(IFullModel<T, TInput, TOutput> featureEncoder, Vector<T> classifierParams, Vector<T> latentCode, LEOOptions<T, TInput, TOutput> options) Initializes a new instance of the LEOModel with one classifier row per class, class c in row c.

Properties

Property Summary
ModelMetadata<T> Metadata { get; }
Vector<T> ClassifierParams { get; } Gets the adapted classifier parameters, one row of EmbeddingDimension weights per class.
Vector<T> LatentCode { get; } Gets the optimized latent code, one LatentDimension block per class.
int LatentDimension { get; } Gets the latent dimension.
int NumClasses { get; } Gets the number of classes.

Methods

Method Summary
ModelMetadata<T> GetModelMetadata()
TOutput Predict(TInput input)
Vector<T> GetParameters()
void Train(TInput inputs, TOutput targets)
void UpdateParameters(Vector<T> parameters)