kedro_datasets_experimental.safetensors.SafetensorsDataset¶

class kedro_datasets_experimental.safetensors.SafetensorsDataset(*, filepath, backend='numpy', version=None, credentials=None, fs_args=None, metadata=None)[source]¶

SafetensorsDataset loads/saves data from/to a Safetensors file using an underlying filesystem (e.g.: local, S3, GCS). The underlying functionality is supported by the specified backend library passed in (defaults to the numpy library), so it supports all allowed options for loading and Safetensors files.

Example usage for the YAML API:

test_model:
  type: safetensors.SafetensorsDataset
  filepath: data/07_model_output/test_model.safetensors

Example usage for the Python API:

 from kedro_datasets_experimental.safetensors import SafetensorsDataset
 import numpy as np

 data = {
...     "embedding": np.zeros((512, 1024)),
...     "attention": np.zeros((256, 256))
... }
 dataset = SafetensorsDataset(
...     filepath="test.safetensors",
... )
 dataset.save(data)
 reloaded = dataset.load()
 assert all(np.array_equal(data[key], reloaded[key]) for key in data)

Attributes

`DEFAULT_FS_ARGS`
`DEFAULT_LOAD_ARGS`
`DEFAULT_SAVE_ARGS`

Methods

`exists`()	Checks whether a dataset's output already exists by calling the provided _exists() method.
`from_config`(name, config[, load_version, ...])	Create a dataset instance using the configuration provided.
`load`()	Loads data by delegation to the provided load method.
`release`()	Release any cached data.
`resolve_load_version`()	Compute the version the dataset should be loaded with.
`resolve_save_version`()	Compute the version the dataset should be saved with.
`save`(data)	Saves data by delegation to the provided save method.
`to_config`()	Converts the dataset instance into a dictionary-based configuration for serialization.

DEFAULT_FS_ARGS: dict[str, Any] = {'open_args_save': {'mode': 'wb'}}¶

DEFAULT_LOAD_ARGS: dict[str, Any] = {}¶

DEFAULT_SAVE_ARGS: dict[str, Any] = {}¶

__init__(*, filepath, backend='numpy', version=None, credentials=None, fs_args=None, metadata=None)[source]¶

Creates a new instance of SafetensorsDataset pointing to a concrete Safetensors file on a specific filesystem. SafetensorsDataset supports custom backends to serialise/deserialise objects.

The following backends are supported:

numpy
torch
tensorflow
paddle
flax

Parameters:

filepath (str) – Filepath in POSIX format to a Safetensors file prefixed with a protocol like s3://. If prefix is not provided, file protocol (local filesystem) will be used. The prefix should be any protocol supported by fsspec. Note: http(s) doesn’t support versioning.
backend (str) – The backend library to use for serialising/deserialising objects. The default backend is ‘numpy’.
version (Optional[Version]) – If specified, should be an instance of kedro.io.core.Version. If its load attribute is None, the latest version will be loaded. If its save attribute is None, save version will be autogenerated.
credentials (Optional[dict[str, Any]]) – Credentials required to get access to the underlying filesystem. E.g. for GCSFileSystem it should look like {“token”: None}.
fs_args (Optional[dict[str, Any]]) – Extra arguments to pass into underlying filesystem class constructor (e.g. {“project”: “my-project”} for GCSFileSystem), as well as to pass to the filesystem’s open method through nested keys open_args_load and open_args_save. Here you can find all available arguments for open: https://filesystem-spec.readthedocs.io/en/latest/api.html#fsspec.spec.AbstractFileSystem.open All defaults are preserved, except mode, which is set to wb when saving.
metadata (Optional[dict[str, Any]]) – Any arbitrary metadata. This is ignored by Kedro, but may be consumed by users or external plugins.

Raises:

ImportError – If the backend module could not be imported.

exists()[source]¶

Checks whether a dataset’s output already exists by calling the provided _exists() method.

Return type:: bool
Returns:: Flag indicating whether the output already exists.
Raises:: DatasetError – when underlying exists method raises error.

classmethod from_config(name, config, load_version=None, save_version=None)[source]¶

Create a dataset instance using the configuration provided.

Parameters:

name (str) – Data set name.
config (dict[str, Any]) – Data set config dictionary.
load_version (Optional[str]) – Version string to be used for load operation if the dataset is versioned. Has no effect on the dataset if versioning was not enabled.
save_version (Optional[str]) – Version string to be used for save operation if the dataset is versioned. Has no effect on the dataset if versioning was not enabled.

Return type:

AbstractDataset

Returns:

An instance of an AbstractDataset subclass.

Raises:

DatasetError – When the function fails to create the dataset from its config.

load()[source]¶

Loads data by delegation to the provided load method.

Return type:: Any
Returns:: Data returned by the provided load method.
Raises:: DatasetError – When underlying load method raises error.

release()[source]¶

Release any cached data.

Raises:: DatasetError – when underlying release method raises error.
Return type:: None

resolve_load_version()[source]¶

Compute the version the dataset should be loaded with.

Return type:: Optional[str]

resolve_save_version()[source]¶

Compute the version the dataset should be saved with.

Return type:: Optional[str]

save(data)[source]¶

Saves data by delegation to the provided save method.

Parameters:

data (Any) – the value to be saved by provided save method.

Raises:

DatasetError – when underlying save method raises error.
FileNotFoundError – when save method got file instead of dir, on Windows.
NotADirectoryError – when save method got file instead of dir, on Unix.

Return type:

None

to_config()[source]¶

Converts the dataset instance into a dictionary-based configuration for serialization. Ensures that any subclass-specific details are handled, with additional logic for versioning and caching implemented for CachedDataset.

Adds a key for the dataset’s type using its module and class name and includes the initialization arguments.

For CachedDataset it extracts the underlying dataset’s configuration, handles the versioned flag and removes unnecessary metadata. It also ensures the embedded dataset’s configuration is appropriately flattened or transformed.

If the dataset has a version key, it sets the versioned flag in the configuration.

Removes the metadata key from the configuration if present.

Return type:: dict[str, Any]
Returns:: A dictionary containing the dataset’s type and initialization arguments.