mnist_dataset

HDF5-backed MNIST Dataset for PyTorch.

This implements a lightweight dataset that lazily opens an HDF5 file and returns (image_tensor, label) pairs. Images are returned as uint8 tensors scaled to [0,1] if a transform isn’t provided.

Classes

MnistH5Dataset

Module Contents

class mnist_dataset.MnistH5Dataset(h5_path: str | pathlib.Path, transform: Callable | None = None, target_transform: Callable | None = None)

Bases: torch.utils.data.Dataset

h5_path

transform = None

target_transform = None

_h5 = None

_ensure_open()

__len__()

Return the number of samples in the dataset.

This reads the labels dataset’s first dimension to determine the length and therefore triggers lazy opening of the HDF5 file.

__getitem__(idx)

Return the (image, label) pair for the given index.

Parameters

idxint: Index of the requested sample.

Returns

(torch.Tensor, int): A tuple of (image_tensor, label). The image tensor is a float tensor scaled to [0,1] with shape (C, H, W). Label is an int and may have been transformed by target_transform if provided.

Raises

IndexError: If the index is out of bounds for the dataset.

close()

Close the underlying HDF5 file handle if open.

This is safe to call multiple times. It is also invoked from __del__ to ensure resources are released when the dataset is garbage-collected.

__del__()