from __future__ import annotations
import importlib
from copy import deepcopy
from pathlib import PurePosixPath
from typing import Any
import fsspec
from kedro.io.core import (
AbstractVersionedDataset,
DatasetError,
Version,
get_filepath_str,
get_protocol_and_path,
)
[docs]
class SafetensorsDataset(AbstractVersionedDataset[Any, Any]):
"""``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 <https://docs.kedro.org/en/stable/data/\
data_catalog_yaml_examples.html>`_:
.. code-block:: yaml
test_model:
type: safetensors.SafetensorsDataset
filepath: data/07_model_output/test_model.safetensors
Example usage for the
`Python API <https://docs.kedro.org/en/stable/data/\
advanced_data_catalog_usage.html>`_:
.. code-block:: pycon
>>> 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)
"""
DEFAULT_LOAD_ARGS: dict[str, Any] = {}
DEFAULT_SAVE_ARGS: dict[str, Any] = {}
DEFAULT_FS_ARGS: dict[str, Any] = {"open_args_save": {"mode": "wb"}}
[docs]
def __init__( # noqa: PLR0913
self,
*,
filepath: str,
backend: str = "numpy",
version: Version | None = None,
credentials: dict[str, Any] | None = None,
fs_args: dict[str, Any] | None = None,
metadata: dict[str, Any] | None = None,
) -> None:
"""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`
Args:
filepath: 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: The backend library to use for serialising/deserialising objects.
The default backend is 'numpy'.
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: Credentials required to get access to the underlying filesystem.
E.g. for ``GCSFileSystem`` it should look like `{"token": None}`.
fs_args: 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: 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.
"""
try:
importlib.import_module(f"safetensors.{backend}")
except ImportError as exc:
raise ImportError(
f"Selected backend '{backend}' could not be imported. "
"Make sure it is installed and importable."
) from exc
_fs_args = deepcopy(fs_args) or {}
_fs_open_args_load = _fs_args.pop("open_args_load", {})
_fs_open_args_save = _fs_args.pop("open_args_save", {})
_credentials = deepcopy(credentials) or {}
protocol, path = get_protocol_and_path(filepath, version)
if protocol == "file":
_fs_args.setdefault("auto_mkdir", True)
self._protocol = protocol
self._fs = fsspec.filesystem(self._protocol, **_credentials, **_fs_args)
self.metadata = metadata
super().__init__(
filepath=PurePosixPath(path),
version=version,
exists_function=self._fs.exists,
glob_function=self._fs.glob,
)
self._backend = backend
self._fs_open_args_load = {
**self.DEFAULT_FS_ARGS.get("open_args_load", {}),
**(_fs_open_args_load or {}),
}
self._fs_open_args_save = {
**self.DEFAULT_FS_ARGS.get("open_args_save", {}),
**(_fs_open_args_save or {}),
}
[docs]
def load(self) -> Any:
load_path = get_filepath_str(self._get_load_path(), self._protocol)
with self._fs.open(load_path, **self._fs_open_args_load) as fs_file:
imported_backend = importlib.import_module(f"safetensors.{self._backend}")
return imported_backend.load(fs_file.read())
[docs]
def save(self, data: Any) -> None:
save_path = get_filepath_str(self._get_save_path(), self._protocol)
with self._fs.open(save_path, **self._fs_open_args_save) as fs_file:
try:
imported_backend = importlib.import_module(f"safetensors.{self._backend}")
imported_backend.save_file(data, fs_file.name)
except Exception as exc:
raise DatasetError(
f"{data.__class__} was not serialised due to: {exc}"
) from exc
self._invalidate_cache()
def _describe(self) -> dict[str, Any]:
return {
"filepath": self._filepath,
"backend": self._backend,
"protocol": self._protocol,
"version": self._version,
}
def _exists(self) -> bool:
try:
load_path = get_filepath_str(self._get_load_path(), self._protocol)
except DatasetError:
return False
return self._fs.exists(load_path)
def _release(self) -> None:
super()._release()
self._invalidate_cache()
def _invalidate_cache(self) -> None:
"""Invalidate underlying filesystem caches."""
filepath = get_filepath_str(self._filepath, self._protocol)
self._fs.invalidate_cache(filepath)
