Struct DataLoader 

pub struct DataLoader<D>
where
    D: Dataset<Item = (Tensor<f32>, Tensor<f32>)>,
{ /* private fields */ }

DataLoader for batched iteration over datasets.

Provides configurable batching, shuffling, and iteration over datasets.

Implementations

impl<D> DataLoader<D>

where D: Dataset<Item = (Tensor<f32>, Tensor<f32>)>,

pub fn new(dataset: D, batch_size: usize) -> Self

Creates a new DataLoader with the specified batch size.

pub fn shuffle(self, shuffle: bool) -> Self

Enables or disables shuffling.

pub fn drop_last(self, drop_last: bool) -> Self

Sets whether to drop the last incomplete batch.

pub fn num_workers(self, num_workers: usize) -> Self

Sets the number of worker threads for parallel data loading.

pub fn batch_size(&self) -> usize

Returns the batch size.

pub fn len(&self) -> usize

Returns the number of batches.

pub fn is_empty(&self) -> bool

Returns true if the DataLoader is empty.

pub fn dataset_len(&self) -> usize

Returns the dataset length.

pub fn iter(&self) -> DataLoaderIter<'_, D>

Creates an iterator over batches.

impl<D> DataLoader<D>

where D: Dataset<Item = (Tensor<f32>, Tensor<f32>)>,

pub fn prefetch_to_gpu(&self, device: Device) -> GpuPrefetchIter

Returns an iterator that prefetches batches onto the target GPU device in a background thread.

This overlaps CPU data loading and collation with GPU computation: while the training loop processes the current GPU-resident batch, the background thread loads and transfers the next batch.

Typically provides 10-30% speedup on GPU training workloads.

Arguments

• device - Target GPU device (e.g., Device::Cuda(0))

Example

let loader = DataLoader::new(dataset, 64).shuffle(true);
for batch in loader.prefetch_to_gpu(Device::Cuda(0)) {
    // batch.data and batch.targets are already on GPU
}