IDataLoader<TInput, TOutput>

Interfaces · AiDotNet.Interfaces

Data loader parameterized by both input and output tensor types — the type-theoretic fit for domains where the batch shape is richer than the row-scalar (X, y) supervised pair (image-space photometric NeRF training, video / audio / sequence corpora, RL trajectories). Added as part of #1834 without renaming the existing IDataLoader so currently-shipped loaders (matrix-in / vector-out supervised) migrate at their own pace.

Implements: ICountable, IResettable

How It Works

Reference PyTorch DataLoader is generic over its collate output type; this interface expresses the same idea in C# nominal-typing terms. Facade branches on the concrete TInput / TOutput pair via is checks — an image loader is a IDataLoader<ImageView<T>, PixelBatch<T>>, and the model layer sees exactly that shape.

The original IDataLoader continues to be the base for row-scalar supervised loaders and the specialized IGraphDataLoader / IRLDataLoader / IEpisodicDataLoader hierarchies — no breakage, no rename. When a full migration is done, IDataLoader can be re-expressed as IDataLoader<Matrix<T>, Vector<T>>.

Properties

Property Summary
bool IsLoaded { get; } Whether the underlying data has been fetched into memory and is ready to iterate.
string Description { get; } Description of the dataset and its intended use.
string Name { get; } Human-readable name of this loader.

Methods

Method Summary
IEnumerable<ValueTuple<TInput, TOutput>> IterateBatches(int batchSize) Enumerates the loaded dataset one batch at a time.
Task LoadAsync(CancellationToken cancellationToken) Loads the dataset (file read, network fetch, preprocessing) asynchronously.
void Unload() Releases the loaded dataset from memory.