ITaskSampler<T, TInput, TOutput>

Interfaces ยท AiDotNet.Interfaces

Controls the strategy for sampling meta-learning tasks (episodes) from a dataset. Implementations can provide uniform, balanced, curriculum-based, or dynamic sampling.

Known implementations: BalancedTaskSampler<T, TInput, TOutput>, CurriculumTaskSampler<T, TInput, TOutput>, DynamicTaskSampler<T, TInput, TOutput>, EpisodicDataLoaderTaskSamplerAdapter<T, TInput, TOutput>, ModelPredictiveTaskSampler<T, TInput, TOutput>, UniformTaskSampler<T, TInput, TOutput>

For Beginners

A task sampler decides which tasks (episodes) to give the meta-learner during training. Different strategies can make training faster or more robust:

  • Uniform: Picks tasks completely at random โ€” simple and effective.
  • Balanced: Ensures every class appears equally often across tasks.
  • Dynamic: Focuses on tasks the model struggles with the most.
  • Curriculum: Starts with easy tasks, gradually increases difficulty.

Properties

Property Summary
int NumQueryPerClass { get; } Gets the number of query examples per class used by this sampler.
int NumShots { get; } Gets the K-shot configuration used by this sampler.
int NumWays { get; } Gets the N-way configuration used by this sampler.

Methods

Method Summary
IEpisode<T, TInput, TOutput> SampleOne() Samples a single episode from the underlying meta-dataset.
TaskBatch<T, TInput, TOutput> SampleBatch(int batchSize) Samples a batch of tasks from the underlying meta-dataset.
void SetSeed(int seed) Sets the random seed for reproducible sampling.
void UpdateWithFeedback(IReadOnlyList<IEpisode<T, TInput, TOutput>> episodes, IReadOnlyList<double> losses) Updates the sampler state after observing a loss for a batch of tasks.