def _construct_graph(self):
"""Actually load and parse the graph.
This is factored out from __init__ in case we want to support delayed-loads
in the future.
Raises:
ValueError: Could not find signature keyed with
DEFAULT_EVAL_SIGNATURE_DEF_KEY; or signature_def did not have exactly
one input; or there was a signature output with the metric prefix but an
unrecognised suffix.
"""
meta_graph_def = tf.saved_model.loader.load(self._session,
[constants.EVAL_TAG],
self._path)
with self._graph.as_default():
signature_def = meta_graph_def.signature_def.get(
constants.DEFAULT_EVAL_SIGNATURE_DEF_KEY)
if signature_def is None:
raise ValueError(
'could not find signature with name %s. signature_def '
'was %s' % (constants.EVAL_TAG, signature_def))
# If features and labels are not stored in the signature_def.inputs then
# only a single input will be present. We will use this as our flag to
# indicate whether the features and labels should be read using the legacy
# collections or using new signature_def.inputs.
if len(signature_def.inputs) == 1:
self._legacy_check_version(meta_graph_def)
self._input_map, self._input_refs_node = graph_ref.load_legacy_inputs(
meta_graph_def, signature_def, self._graph)
self._features_map, self._labels_map = (
graph_ref.load_legacy_features_and_labels(
meta_graph_def, self._graph))
else:
self._check_version(
graph_ref.load_tfma_version(signature_def, self._graph))
self._input_map, self._input_refs_node = graph_ref.load_inputs(
signature_def, self._graph)
self._features_map, self._labels_map = (
graph_ref.load_features_and_labels(signature_def,
self._graph))
self._predictions_map = graph_ref.load_predictions(
signature_def, self._graph)
# Create feed_list for metrics_reset_update_get_fn
#
# We need to save this because we need to update the
# metrics_reset_update_get_fn when additional metric ops are registered
# (the feed_list will stay the same though).
feed_list = []
feed_list_keys = []
for which_map, key, map_dict in (
self._iterate_fpl_maps_in_canonical_order()):
feed_list.append(map_dict[encoding.NODE_SUFFIX])
feed_list_keys.append((which_map, key))
self._perform_metrics_update_fn_feed_list = feed_list
# We also keep the associated keys for better error messages.
self._perform_metrics_update_fn_feed_list_keys = feed_list_keys
self._metric_names = []
self._metric_value_ops = []
self._metric_update_ops = []
self._metric_variable_nodes = []
self._metric_variable_placeholders = []
self._metric_variable_assign_ops = []
if self._include_default_metrics:
metrics_map = graph_ref.load_metrics(signature_def,
self._graph)
metric_ops = {}
for metric_name, ops in metrics_map.items():
metric_ops[metric_name] = (ops[encoding.VALUE_OP_SUFFIX],
ops[encoding.UPDATE_OP_SUFFIX])
self.register_additional_metric_ops(metric_ops)
# Make callable for predict_list. The callable for
# metrics_reset_update_get is updated in register_additional_metric_ops.
# Repeated calls to a callable made using make_callable are faster than
# doing repeated calls to session.run.
self._predict_list_fn = self._session.make_callable(
fetches=(self._features_map, self._predictions_map,
self._labels_map, self._input_refs_node),
feed_list=list(self._input_map.values()))
def _construct_graph(self):
"""Actually load and parse the graph.
This is factored out from __init__ in case we want to support delayed-loads
in the future.
Raises:
ValueError: Could not find signature keyed with
DEFAULT_EVAL_SIGNATURE_DEF_KEY; or signature_def did not have exactly
one input; or there was a signature output with the metric prefix but an
unrecognised suffix.
"""
meta_graph_def = tf.compat.v1.saved_model.loader.load(
self._session, self._tags, self._path)
with self._graph.as_default():
signature_def = meta_graph_def.signature_def.get(
constants.DEFAULT_EVAL_SIGNATURE_DEF_KEY)
if signature_def is None:
raise ValueError('could not find signature with name %s. signature_def '
'was %s' % (constants.EVAL_TAG, signature_def))
self._additional_fetches_map = {}
iterator_initializer = None
# If features and labels are not stored in the signature_def.inputs then
# only a single input will be present. We will use this as our flag to
# indicate whether the features and labels should be read using the legacy
# collections or using new signature_def.inputs.
# TODO(b/119308261): Remove once all exported EvalSavedModels are updated.
if len(signature_def.inputs) == 1:
self._legacy_check_version(meta_graph_def)
self._input_map, self._input_refs_node = graph_ref.load_legacy_inputs(
meta_graph_def, signature_def, self._graph)
self._features_map, self._labels_map = (
graph_ref.load_legacy_features_and_labels(meta_graph_def,
self._graph))
else:
self._check_version(
graph_ref.load_tfma_version(signature_def, self._graph))
self._input_map, self._input_refs_node = graph_ref.load_inputs(
signature_def, self._graph)
self._features_map = graph_ref.load_additional_inputs(
constants.FEATURES_NAME, signature_def, self._graph)
if self._blacklist_feature_fetches:
for feature_name in self._blacklist_feature_fetches:
self._features_map.pop(feature_name, None)
self._labels_map = graph_ref.load_additional_inputs(
constants.LABELS_NAME, signature_def, self._graph)
if self._additional_fetches:
for prefix in self._additional_fetches:
self._additional_fetches_map[prefix] = (
graph_ref.load_additional_inputs(prefix, signature_def,
self._graph))
iterator_initializer = self._get_op_from_tensor(
graph_ref.load_iterator_initializer_name(signature_def,
self._graph))
self._predictions_map = graph_ref.load_predictions(
signature_def, self._graph)
# Create feed_list for metrics_reset_update_get_fn
#
# We need to save this because we need to update the
# metrics_reset_update_get_fn when additional metric ops are registered
# (the feed_list will stay the same though).
self._perform_metrics_update_fn_feed_list = list(self._input_map.values())
self._metric_names = []
self._metric_value_ops = []
self._metric_update_ops = []
self._metric_variable_nodes = []
self._metric_variable_placeholders = []
self._metric_variable_assign_ops = []
if self._include_default_metrics:
metrics_map = graph_ref.load_metrics(signature_def, self._graph)
metric_ops = {}
for metric_name, ops in metrics_map.items():
metric_ops[metric_name] = (ops[encoding.VALUE_OP_SUFFIX],
ops[encoding.UPDATE_OP_SUFFIX])
self.register_additional_metric_ops(metric_ops)
# Make callable for predict_list. The callable for
# metrics_reset_update_get is updated in register_additional_metric_ops.
# Repeated calls to a callable made using make_callable are faster than
# doing repeated calls to session.run.
if iterator_initializer:
# When iterator is used, the initializer is used to feed the inputs. The
# values are then fetched by repeated calls to the predict_list_fn until
# OutOfRange is thrown.
self._iterator_initializer_fn = self._session.make_callable(
fetches=(iterator_initializer),
feed_list=list(self._input_map.values()))
self._predict_list_fn = self._session.make_callable(
fetches=(self._features_map, self._predictions_map,
self._labels_map, self._input_refs_node,
self._additional_fetches_map))
else:
self._iterator_initializer_fn = None
self._predict_list_fn = self._session.make_callable(
fetches=(self._features_map, self._predictions_map,
self._labels_map, self._input_refs_node,
self._additional_fetches_map),
feed_list=list(self._input_map.values()))