def export(self, checkpoint_path=None, export_dir_base=None):
"""Exports a model.
Args:
checkpoint_path: The checkpoint path to export. If ``None``, the latest is used.
export_dir_base: The base directory in which a timestamped subdirectory
containing the exported model will be created. Defaults to
``$MODEL_DIR/export/manual``.
Returns:
The string path to the exported directory.
"""
estimator = self._make_estimator()
if checkpoint_path is not None and tf.gfile.IsDirectory(checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(checkpoint_path)
if export_dir_base is None:
export_dir_base = os.path.join(estimator.model_dir, "export", "manual")
kwargs = {}
if hasattr(estimator, "export_saved_model"):
export_fn = estimator.export_saved_model
else:
export_fn = estimator.export_savedmodel
if "strip_default_attrs" in fn_args(estimator.export_savedmodel):
# Set strip_default_attrs to True for TensorFlow 1.6+ to stay consistent
# with the behavior of tf.estimator.Exporter.
kwargs["strip_default_attrs"] = True
return export_fn(
export_dir_base,
estimator_util.make_serving_input_fn(self._model, metadata=self._config["data"]),
assets_extra=self._get_model_assets(),
checkpoint_path=checkpoint_path,
**kwargs)
def serving_input_fn(self, metadata=None):
"""Returns the serving input function.
Args:
metadata: A dictionary containing additional metadata set
by the user. Required if ``Model.initialize()`` has not been called.
Returns:
A callable that returns a ``tf.estimator.export.ServingInputReceiver``.
"""
if metadata is not None:
self.initialize(metadata)
return estimator.make_serving_input_fn(self)
def _build_eval_spec(self):
eval_spec = tf.estimator.EvalSpec(
input_fn=estimator_util.make_input_fn(
self._model,
tf.estimator.ModeKeys.EVAL,
self._config["eval"]["batch_size"],
features_file=self._config["data"]["eval_features_file"],
labels_file=self._config["data"].get("eval_labels_file"),
num_threads=self._config["eval"].get("num_threads"),
prefetch_buffer_size=self._config["eval"].get("prefetch_buffer_size")),
steps=None,
exporters=_make_exporters(
self._config["eval"]["exporters"],
estimator_util.make_serving_input_fn(self._model),
assets_extra=self._get_model_assets()),
throttle_secs=self._config["eval"]["eval_delay"])
return eval_spec
def export(self, checkpoint_path=None, export_dir_base=None):
"""Exports a model.
Args:
checkpoint_path: The checkpoint path to export. If ``None``, the latest is used.
export_dir_base: The base directory in which a timestamped subdirectory
containing the exported model will be created. Defaults to
``$MODEL_DIR/export/manual``.
Returns:
The string path to the exported directory.
"""
estimator = self._make_estimator()
if checkpoint_path is not None and tf.gfile.IsDirectory(checkpoint_path):
checkpoint_path = tf.train.latest_checkpoint(checkpoint_path)
if export_dir_base is None:
export_dir_base = os.path.join(estimator.model_dir, "export", "manual")
kwargs = {}
if hasattr(estimator, "export_saved_model"):
export_fn = estimator.export_saved_model
else:
export_fn = estimator.export_savedmodel
if "strip_default_attrs" in fn_args(estimator.export_savedmodel):
# Set strip_default_attrs to True for TensorFlow 1.6+ to stay consistent
# with the behavior of tf.estimator.Exporter.
kwargs["strip_default_attrs"] = True
# This is a hack for SequenceRecordInputter that currently infers the input
# depth from the data files.
# TODO: This method should not require the training data.
data_config = self._config["data"]
if "train_features_file" in data_config:
_ = model.features_inputter.make_dataset(data_config["train_features_file"])
return export_fn(
export_dir_base,
estimator_util.make_serving_input_fn(self._model),
assets_extra=self._get_model_assets(),
checkpoint_path=checkpoint_path,
**kwargs)
