Datasets

Dataset utilities provide convenient access to commonly used datasets for experimentation and benchmarking.

All dataset loaders documented here are part of KeyDNN’s public presentation API.

---

load_mnist

Load the MNIST dataset with sensible defaults.

This is a thin convenience wrapper around :class:MNIST that: - Uses a default dataset root (data/) - Automatically downloads the dataset if it is not already present - Exposes only the commonly tuned parameters in the public API

Parameters:

Name	Type	Description	Default
train	bool	Whether to load the training split. If False, loads the test split.	True
transform	callable	Optional transform applied to each image.	None
target_transform	callable	Optional transform applied to each label.	None
normalize	bool	Whether to apply standard MNIST mean/std normalization.	True
return_numpy	bool	Whether to return NumPy arrays. If False, callers may wrap outputs into tensors downstream.	False
dtype	str	Floating-point dtype used for image conversion.	"float32"
root_path	str or Path	Base directory for dataset storage. The MNIST files are stored under <root_path>/mnist/raw.	"data"

Returns:

Type	Description
MNIST	An initialized MNIST dataset instance.

---

load_cifar10

Load the CIFAR-10 dataset with sensible defaults.

This is a thin convenience wrapper around :class:CIFAR10 that: - Uses a default dataset root (data/) - Automatically downloads the dataset if it is not already present - Exposes only the commonly tuned parameters in the public API

Parameters:

Name	Type	Description	Default
train	bool	Whether to load the training split. If False, loads the test split.	True
transform	callable	Optional transform applied to each image.	None
target_transform	callable	Optional transform applied to each label.	None
normalize	bool	Whether to apply standard CIFAR-10 mean/std normalization.	True
return_numpy	bool	Whether to return NumPy arrays. If False, callers may wrap outputs into tensors downstream.	False
dtype	str	Floating-point dtype used for image conversion.	"float32"
root_path	str or Path	Base directory for dataset storage. The CIFAR-10 files are stored under <root_path>/cifar10/raw.	"data"

Returns:

Type	Description
CIFAR10	An initialized CIFAR-10 dataset instance.

---

load_cifar100

Load the CIFAR-100 dataset with sensible defaults.

This is a thin convenience wrapper around :class:CIFAR100 that: - Uses a default dataset root (data/) - Automatically downloads the dataset if it is not already present - Exposes only the commonly tuned parameters in the public API

Parameters:

Name	Type	Description	Default
train	bool	Whether to load the training split. If False, loads the test split.	True
transform	callable	Optional transform applied to each image.	None
target_transform	callable	Optional transform applied to each label.	None
normalize	bool	Whether to apply standard CIFAR-100 mean/std normalization.	True
return_numpy	bool	Whether to return NumPy arrays. If False, callers may wrap outputs into tensors downstream.	False
dtype	str	Floating-point dtype used for image conversion.	"float32"
root_path	str or Path	Base directory for dataset storage. The CIFAR-100 files are stored under <root_path>/cifar100/raw.	"data"

Returns:

Type	Description
CIFAR100	An initialized CIFAR-100 dataset instance.

---

Notes

• Dataset loaders may download data automatically if not already present.
• Returned data is typically provided as NumPy arrays or Tensor objects, depending on configuration.
• For reproducibility, consider setting seeds and determinism before loading datasets.