TensorFlowModelDataset

TensorFlowModelDataset loads and saves TensorFlow models.

kedro_datasets.tensorflow.TensorFlowModelDataset

TensorFlowModelDataset(
    *,
    filepath,
    load_args=None,
    save_args=None,
    version=None,
    credentials=None,
    fs_args=None,
    metadata=None
)

Bases: AbstractVersionedDataset[Model, Model]

TensorFlowModelDataset loads and saves TensorFlow models. The underlying functionality is supported by, and passes input arguments through to, TensorFlow 2.X load_model and save_model methods.

TensorFlow does not currently support Python 3.14.

Warning

By default, models are loaded with safe_mode=True, which restricts deserialization to built-in Keras objects only. Loading untrusted model files without this flag can execute arbitrary code. If you need to load custom objects from a trusted source, set safe_mode: false in load_args.

Examples:

Using the YAML API:

tensorflow_model:
  type: tensorflow.TensorFlowModelDataset
  filepath: data/06_models/tensorflow_model.h5
  load_args:
    compile: False
  save_args:
    overwrite: True
    include_optimizer: False
  credentials: tf_creds

Using the Python API:

>>> import numpy as np
>>> import tensorflow as tf
>>> from kedro_datasets.tensorflow import TensorFlowModelDataset
>>>
>>> model = tf.keras.Sequential(
...     [tf.keras.layers.Dense(5, input_shape=(3,)), tf.keras.layers.Softmax()]
... )
>>> # x = tf.random.uniform((10, 3))
>>> # predictions = model.predict(x)
>>>
>>> dataset = TensorFlowModelDataset(
...     filepath=tmp_path / "data/06_models/tensorflow_model.h5"
... )
>>> dataset.save(model)
>>> loaded_model = dataset.load()

Parameters:

• filepath (str) –

  Filepath in POSIX format to a TensorFlow model directory 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.

• load_args (dict[str, Any] | None, default: None ) –

  TensorFlow options for loading models. Here you can find all available arguments: https://www.tensorflow.org/api_docs/python/tf/keras/models/load_model All defaults are preserved, except for "safe_mode", which is set to True. Set safe_mode: false to load models containing custom objects from trusted sources. A special key tf_device can be used to specify the device context for loading.

• save_args (dict[str, Any] | None, default: None ) –

  TensorFlow options for saving models. Here you can find all available arguments: https://www.tensorflow.org/api_docs/python/tf/keras/models/save_model All defaults are preserved, except for "save_format", which is set to "tf".

• version (Version | None, default: None ) –

  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 (dict[str, Any] | None, default: None ) –

  Credentials required to get access to the underlying filesystem. E.g. for GCSFileSystem it should look like {'token': None}.

• fs_args (dict[str, Any] | None, default: None ) –

  Extra arguments to pass into underlying filesystem class constructor (e.g. {"project": "my-project"} for GCSFileSystem).

• metadata (dict[str, Any] | None, default: None ) –

  Any arbitrary metadata. This is ignored by Kedro, but may be consumed by users or external plugins.

Source code in kedro_datasets/tensorflow/tensorflow_model_dataset.py

def __init__(  # noqa: PLR0913
    self,
    *,
    filepath: str,
    load_args: dict[str, Any] | None = None,
    save_args: dict[str, Any] | None = None,
    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 ``TensorFlowModelDataset``.

    Args:
        filepath: Filepath in POSIX format to a TensorFlow model directory 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.
        load_args: TensorFlow options for loading models.
            Here you can find all available arguments:
            https://www.tensorflow.org/api_docs/python/tf/keras/models/load_model
            All defaults are preserved, except for "safe_mode", which is set to True.
            Set ``safe_mode: false`` to load models containing custom objects from
            trusted sources. A special key ``tf_device`` can be used to specify the
            device context for loading.
        save_args: TensorFlow options for saving models.
            Here you can find all available arguments:
            https://www.tensorflow.org/api_docs/python/tf/keras/models/save_model
            All defaults are preserved, except for "save_format", which is set to "tf".
        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``).
        metadata: Any arbitrary metadata.
            This is ignored by Kedro, but may be consumed by users or external plugins.
    """
    _fs_args = copy.deepcopy(fs_args) or {}
    _credentials = copy.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._tmp_prefix = "kedro_tensorflow_tmp"  # temp prefix pattern

    # Handle default load and save arguments
    self._load_args = {**self.DEFAULT_LOAD_ARGS, **(load_args or {})}
    self._save_args = {**self.DEFAULT_SAVE_ARGS, **(save_args or {})}

    self._is_h5 = self._save_args.get("save_format") == "h5"

DEFAULT_LOAD_ARGS class-attribute instance-attribute

DEFAULT_LOAD_ARGS = {'safe_mode': True}

DEFAULT_SAVE_ARGS class-attribute instance-attribute

DEFAULT_SAVE_ARGS = {}

_fs instance-attribute

_fs = fsspec.filesystem(
    self._protocol, **_credentials, **_fs_args
)

_is_h5 instance-attribute

_is_h5 = self._save_args.get('save_format') == 'h5'

_load_args instance-attribute

_load_args = {
    None: self.DEFAULT_LOAD_ARGS,
    None: load_args or {},
}

_protocol instance-attribute

_protocol = protocol

_save_args instance-attribute

_save_args = {
    None: self.DEFAULT_SAVE_ARGS,
    None: save_args or {},
}

_tmp_prefix instance-attribute

_tmp_prefix = 'kedro_tensorflow_tmp'

metadata instance-attribute

metadata = metadata

_describe

_describe()

Source code in kedro_datasets/tensorflow/tensorflow_model_dataset.py

def _describe(self) -> dict[str, Any]:
    return {
        "filepath": self._filepath,
        "protocol": self._protocol,
        "load_args": self._load_args,
        "save_args": self._save_args,
        "version": self._version,
    }

_exists

_exists()

Source code in kedro_datasets/tensorflow/tensorflow_model_dataset.py

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)

_invalidate_cache

_invalidate_cache()

Invalidate underlying filesystem caches.

Source code in kedro_datasets/tensorflow/tensorflow_model_dataset.py

def _invalidate_cache(self) -> None:
    """Invalidate underlying filesystem caches."""
    filepath = get_filepath_str(self._filepath, self._protocol)
    self._fs.invalidate_cache(filepath)

_release

_release()

Source code in kedro_datasets/tensorflow/tensorflow_model_dataset.py

def _release(self) -> None:
    super()._release()
    self._invalidate_cache()

load

load()

Source code in kedro_datasets/tensorflow/tensorflow_model_dataset.py

def load(self) -> tf.keras.Model:
    load_path = get_filepath_str(self._get_load_path(), self._protocol)

    with tempfile.TemporaryDirectory(prefix=self._tmp_prefix) as tempdir:
        if self._is_h5:
            path = str(PurePath(tempdir) / TEMPORARY_H5_FILE)  # noqa: PLW2901
        else:
            # We assume .keras
            path = str(PurePath(tempdir) / TEMPORARY_KERAS_FILE)  # noqa: PLW2901

        self._fs.get(load_path, path)

        load_args = {k: v for k, v in self._load_args.items() if k != "tf_device"}
        device_name = self._load_args.get("tf_device")
        if device_name:
            with tf.device(device_name):
                model = tf.keras.models.load_model(path, **load_args)
        else:
            model = tf.keras.models.load_model(path, **load_args)
        return model

save

save(data)

Source code in kedro_datasets/tensorflow/tensorflow_model_dataset.py

def save(self, data: tf.keras.Model) -> None:
    save_path = get_filepath_str(self._get_save_path(), self._protocol)

    with tempfile.TemporaryDirectory(prefix=self._tmp_prefix) as tempdir:
        if self._is_h5:
            path = str(PurePath(tempdir) / TEMPORARY_H5_FILE)  # noqa: PLW2901
        else:
            # We assume .keras
            path = str(PurePath(tempdir) / TEMPORARY_KERAS_FILE)  # noqa: PLW2901

        tf.keras.models.save_model(data, path, **self._save_args)

        # Use fsspec to take from local tempfile directory/file and
        # put in ArbitraryFileSystem
        self._fs.put(path, save_path)