ParallelBatchLoader<TBatch>
Models & Types · AiDotNet.Data.Loaders
Provides parallel batch loading with multiple workers for improved throughput.
Implements: IDisposable
For Beginners
Training neural networks often involves:
- Loading data from disk
- Preprocessing (augmentation, normalization)
- GPU training
With single-threaded loading, the GPU waits while data is prepared. With parallel loading, multiple workers prepare batches simultaneously, keeping the GPU constantly fed with data.
Example:
How It Works
ParallelBatchLoader uses multiple worker threads to prepare batches in parallel, similar to PyTorch’s DataLoader with num_workers > 0. This can significantly improve training throughput when batch preparation is CPU-bound.
Constructors
| Constructor | Summary |
|---|---|
ParallelBatchLoader(Func<IEnumerable<int>> indexProvider, Func<int[], TBatch> batchFactory, int batchSize, int? numWorkers, int? prefetchCount) |
Initializes a new instance of the ParallelBatchLoader class. |
Properties
| Property | Summary |
|---|---|
int NumWorkers { get; } |
Gets the number of parallel workers. |
int PrefetchCount { get; } |
Gets the prefetch count. |
Methods
| Method | Summary |
|---|---|
IAsyncEnumerable<TBatch> GetBatchesAsync(CancellationToken cancellationToken) |
Iterates through batches using parallel workers. |
void Dispose() |
Disposes the parallel batch loader. |