DynamicTaskSampler<T, TInput, TOutput>
Models & Types · AiDotNet.MetaLearning.Data
Samples tasks with probability proportional to the loss observed on previous evaluations. Tasks with higher loss are sampled more frequently, focusing training on areas the model finds most difficult (inspired by hard-example mining and curriculum learning research).
Implements: ITaskSampler<T, TInput, TOutput>
Constructors
| Constructor | Summary |
|---|---|
DynamicTaskSampler(IMetaDataset<T, TInput, TOutput> dataset, int numWays, int numShots, int numQueryPerClass, double explorationRate, int? seed) |
Creates a dynamic task sampler that adapts based on observed losses. |
Properties
| Property | Summary |
|---|---|
int NumQueryPerClass { get; } |
|
int NumShots { get; } |
|
int NumWays { get; } |
Methods
| Method | Summary |
|---|---|
IEpisode<T, TInput, TOutput> SampleOne() |
|
TaskBatch<T, TInput, TOutput> SampleBatch(int batchSize) |
|
void SetSeed(int seed) |
|
void UpdateWithFeedback(IReadOnlyList<IEpisode<T, TInput, TOutput>> episodes, IReadOnlyList<double> losses) |