pulsatrix::DataLoader Class Reference

Orchestrates sampling and collation into batches – pulsatrix's DataLoader (PyTorch DataLoader / torch::data::DataLoader analogue). More...

#include <data_loader.hpp>

Public Member Functions
	DataLoader (std::shared_ptr< Dataset > dataset, DeviceBackend *backend, DataLoaderOptions options={})
	Constructs a DataLoader over a random-access Dataset.
	DataLoader (std::shared_ptr< IterableDataset > dataset, DeviceBackend *backend, DataLoaderOptions options={})
	Constructs a DataLoader over a streaming IterableDataset.
void	reset_epoch ()
	Resets to the start of a new epoch (re-seeds/reshuffles the sampler, or resets the stream).
std::optional< Batch >	next_batch ()
	Fetches the next batch.
int64_t	num_batches () const
	Number of batches per epoch.

Detailed Description

Orchestrates sampling and collation into batches – pulsatrix's DataLoader (PyTorch DataLoader / torch::data::DataLoader analogue).

Note

Phase 1 ships a fully synchronous implementation (num_workers=0 is the only exercised path): DataThreadPool/BoundedQueue exist (see those headers) but are not yet wired into this class's fetch/collate path – wiring num_workers > 0 through a real staged pipeline is a dedicated, separately-tested later mission (Decision Point 7 in campaign_exai_dl_library_data_pipeline).

Constructor & Destructor Documentation

DataLoader() [1/2]

pulsatrix::DataLoader::DataLoader	(	std::shared_ptr< Dataset >	dataset,
		DeviceBackend *	backend,
		DataLoaderOptions	options = {}
	)

Constructs a DataLoader over a random-access Dataset.

Exceptions

std::invalid_argument

if dataset is null or options.batch_size <= 0.

DataLoader() [2/2]

pulsatrix::DataLoader::DataLoader	(	std::shared_ptr< IterableDataset >	dataset,
		DeviceBackend *	backend,
		DataLoaderOptions	options = {}
	)

Constructs a DataLoader over a streaming IterableDataset.

Exceptions

std::invalid_argument

if dataset is null, options.batch_size <= 0, or options.num_workers > 1 – sharded multi-worker streaming is out of scope for this phase (a single worker can safely consume the whole stream).

Member Function Documentation

next_batch()

std::optional< Batch > pulsatrix::DataLoader::next_batch

(

)

Fetches the next batch.

Returns

The next Batch, or std::nullopt once the epoch is exhausted (or, with drop_last=true, once fewer than batch_size samples remain).

num_batches()

int64_t pulsatrix::DataLoader::num_batches

(

)

const

Number of batches per epoch.

Exceptions

std::logic_error

if this DataLoader was constructed from an IterableDataset (its length is unknown ahead of iteration).

reset_epoch()

void pulsatrix::DataLoader::reset_epoch

(

)

Resets to the start of a new epoch (re-seeds/reshuffles the sampler, or resets the stream).

---

The documentation for this class was generated from the following file:

• include/pulsatrix/data_loader.hpp