namespace

`SushiAI::Data`

Declared in: include/SushiAI/data/dataset.hpp

Contains

Typedefs

using SushiAI::Data::DatasetFactory = std::unique_ptr<Dataset> (*)(SushiBLAS::Engine& engine, const Config::Value& attributes, std::string_view context)

Builds one dataset from its configuration attributes.

Parameters

engine: The engine that allocates and fills the tensors.
attributes: The dataset's own keys; the factory rejects any it does not know.
context: What to call this dataset in a diagnostic.

Returns

The dataset.

Variables

constexpr std::array< DatasetEntry, 2 > DATASET_REGISTRY = { DatasetEntry{"rings", &make_rings}, DatasetEntry{"mnist", &make_mnist}, }

Every dataset type a configuration file may name.

A file-scope constant table; nothing self-registers. See config/registry.hpp for why that is forced rather than preferred.

Functions

bool mnist_available(const MnistOptions &options)

Whether the IDX files for options are present and readable.

Parameters

options: Where to look and which split.

Returns

True when both the image and label files exist.

std::string mnist_expected_paths(const MnistOptions &options)

The path pair mnist_available checks, for a diagnostic message.

Parameters

options: Where to look and which split.

Returns

"images: <path>, labels: <path>".

std::unique_ptr< Dataset > try_load_mnist(SushiBLAS::Engine &engine, const MnistOptions &options={})

Loads MNIST from IDX files, returning nullptr if absent.

Parameters

engine: Engine allocating device tensors.
options: Directory, split, and sample count limits.

Returns

Loaded dataset or nullptr when files are missing.

Exceptions

Error: If an existing IDX file is malformed.

std::unique_ptr< Dataset > make_rings(SushiBLAS::Engine &engine, const Config::Value &attributes, std::string_view context)

Builds a Rings dataset from configuration attributes.

Parameters

attributes: Dataset parameters dictionary.
context: Diagnostic context string.

Returns

Constructed dataset instance.

Exceptions

Error: If an attribute key is unknown or value is invalid.

std::unique_ptr< Dataset > make_mnist(SushiBLAS::Engine &engine, const Config::Value &attributes, std::string_view context)

Builds the MNIST dataset from configuration attributes.

Reads "directory" (required), "split" ("train" or "test"), "batch_size" and "max_samples". A relative directory is used as given, so it resolves against the working directory of the process.

Returns

The loaded dataset; never null.

std::unique_ptr< Dataset > make_dataset(SushiBLAS::Engine &engine, std::string_view type, const Config::Value &attributes, std::string_view context)

Builds the dataset named type from the registry.

Parameters

type: Registry key identifying dataset implementation.
attributes: Dataset configuration attributes.

Returns

Constructed dataset instance.

Exceptions

Error: If type is not registered.