def construct(): # pylint: disable=invalid-name
"""Function for constructing shared ModelTypes."""
start_time = datetime.datetime.now()
saved_model = None
keras_model = None
eval_saved_model = None
if model_path:
if tf.version.VERSION.split('.')[0] == '1':
saved_model = tf.compat.v1.saved_model.load_v2(model_path,
tags=[tag])
else:
saved_model = tf.saved_model.load(model_path, tags=[tag])
try:
keras_model = tf.keras.experimental.load_from_saved_model(
model_path)
except tf.errors.NotFoundError:
pass
if eval_saved_model_path:
eval_saved_model = load.EvalSavedModel(
eval_saved_model_path,
include_default_metrics,
additional_fetches=additional_fetches,
blacklist_feature_fetches=blacklist_feature_fetches)
if add_metrics_callbacks:
eval_saved_model.register_add_metric_callbacks(
add_metrics_callbacks)
eval_saved_model.graph_finalize()
end_time = datetime.datetime.now()
model_load_seconds_callback(
int((end_time - start_time).total_seconds()))
return types.ModelTypes(saved_model=saved_model,
keras_model=keras_model,
eval_saved_model=eval_saved_model)
def construct(): # pylint: disable=invalid-name
"""Function for constructing shared ModelTypes."""
start_time = datetime.datetime.now()
saved_model = None
keras_model = None
eval_saved_model = None
if tags == [eval_constants.EVAL_TAG]:
eval_saved_model = load.EvalSavedModel(
eval_saved_model_path,
include_default_metrics,
additional_fetches=additional_fetches,
blacklist_feature_fetches=blacklist_feature_fetches)
if add_metrics_callbacks:
eval_saved_model.register_add_metric_callbacks(add_metrics_callbacks)
eval_saved_model.graph_finalize()
else:
try:
keras_model = tf.keras.models.load_model(eval_saved_model_path)
except Exception: # pylint: disable=broad-except
saved_model = tf.compat.v1.saved_model.load_v2(
eval_saved_model_path, tags=tags)
end_time = datetime.datetime.now()
model_load_seconds_callback(int((end_time - start_time).total_seconds()))
return types.ModelTypes(
saved_model=saved_model,
keras_model=keras_model,
eval_saved_model=eval_saved_model)
def construct(): # pylint: disable=invalid-name
"""Function for constructing a model agnostic eval graph."""
start_time = datetime.datetime.now()
model_agnostic_eval = ModelAgnosticEvaluateGraph(add_metrics_callbacks,
config)
end_time = datetime.datetime.now()
model_load_seconds_callback(int((end_time - start_time).total_seconds()))
return types.ModelTypes(
saved_model=None,
keras_model=None,
eval_saved_model=model_agnostic_eval)
def construct(): # pylint: disable=invalid-name
"""Function for constructing shared ModelTypes."""
start_time = datetime.datetime.now()
saved_model = None
keras_model = None
eval_saved_model = None
# If we are evaluating on TPU, initialize the TPU.
# TODO(b/143484017): Add model warmup for TPU.
if tf.saved_model.TPU in tags:
tf.tpu.experimental.initialize_tpu_system()
if eval_constants.EVAL_TAG in tags:
eval_saved_model = load.EvalSavedModel(
eval_saved_model_path,
include_default_metrics,
additional_fetches=additional_fetches,
blacklist_feature_fetches=blacklist_feature_fetches,
tags=tags)
if add_metrics_callbacks:
eval_saved_model.register_add_metric_callbacks(
add_metrics_callbacks)
eval_saved_model.graph_finalize()
else:
# TODO(b/141524386, b/141566408): TPU Inference is not supported
# for Keras saved_model yet.
try:
keras_model = tf.keras.models.load_model(
eval_saved_model_path)
# In some cases, tf.keras.models.load_model can successfully load a
# saved_model but it won't actually be a keras model.
if not isinstance(keras_model, tf.keras.models.Model):
keras_model = None
except Exception: # pylint: disable=broad-except
keras_model = None
if keras_model is None:
saved_model = tf.compat.v1.saved_model.load_v2(
eval_saved_model_path, tags=tags)
end_time = datetime.datetime.now()
model_load_seconds_callback(
int((end_time - start_time).total_seconds()))
return types.ModelTypes(saved_model=saved_model,
keras_model=keras_model,
eval_saved_model=eval_saved_model)
