Module data 

Python bindings for torsh-data — Dataset and DataLoader APIs

Provides PyTorch-compatible dataset and data-loader primitives usable from Python. The bindings are deliberately concrete (no type-parameter leakage into Python) while still routing through the real torsh-data types wherever the API permits it.

Design choices

• PyDataset stores samples as Vec<Vec<f32>> (flat row-per-sample) so that it can implement torsh_data::Dataset and be passed to the real torsh_data::DataLoader::builder(). Each sample is exposed to Python as a Vec<f32>.

• PyDataLoader owns a concrete SimpleDataLoader<PyDataset> or SimpleRandomDataLoader<PyDataset> depending on shuffle. Because these are different types we erase them behind a PyDataLoaderState enum so that a single #[pyclass] struct suffices.

• Iteration is implemented on the Rust side via PyDataLoaderIter — a separate #[pyclass] that satisfies the __iter__/__next__ protocol.

Structs

PyDataLoader

DataLoader wrapping a PyDataset with configurable batching and shuffling.

PyDataLoaderIter

Python iterator that steps through pre-materialised batches.

PyDataset

In-memory dataset of f32 sample rows exposed to Python.

Functions

register_data_module

Register the data sub-module into the parent module m.