def _input_fn(params=None):
"""The `Dataset` input_fn which will be returned."""
del params # Unused.
input_features = tf_compat.make_one_shot_iterator(
tf.data.Dataset.from_tensors([features])).get_next()
if labels is not None:
input_labels = tf_compat.make_one_shot_iterator(
tf.data.Dataset.from_tensors([labels])).get_next()
else:
input_labels = None
return {"x": input_features}, input_labels
def build_subnetwork(self, features, labels, logits_dimension, training,
iteration_step, summary, previous_ensemble):
# We don't need an EVAL mode since AdaNet takes care of evaluation for us.
mode = tf.estimator.ModeKeys.PREDICT
if training:
mode = tf.estimator.ModeKeys.TRAIN
# Call in template to ensure that variables are created once and reused.
call_model_fn_template = tf.make_template("model_fn", self._call_model_fn)
subestimator_features, subestimator_labels = features, labels
if training and self._subestimator.train_input_fn:
# TODO: Consider tensorflow_estimator/python/estimator/util.py.
inputs = self._subestimator.train_input_fn()
if isinstance(inputs, (tf_compat.DatasetV1, tf_compat.DatasetV2)):
subestimator_features, subestimator_labels = (
tf_compat.make_one_shot_iterator(inputs).get_next())
else:
subestimator_features, subestimator_labels = inputs
# Construct subnetwork graph first because of dependencies on scope.
_, train_op = call_model_fn_template(subestimator_features,
subestimator_labels, mode, summary)
# Graph for ensemble learning gets model_fn_1 for scope.
logits, _ = call_model_fn_template(features, labels, mode, summary)
else:
logits, train_op = call_model_fn_template(features, labels, mode, summary)
# TODO: Replace with variance complexity measure.
complexity = tf.constant(0.)
return Subnetwork(
logits=logits,
last_layer=logits,
shared={"train_op": train_op},
complexity=complexity)
def build_subnetwork(self, features, labels, logits_dimension, training,
iteration_step, summary, previous_ensemble):
# We don't need an EVAL mode since AdaNet takes care of evaluation for us.
mode = tf.estimator.ModeKeys.PREDICT
if training:
mode = tf.estimator.ModeKeys.TRAIN
# Call in template to ensure that variables are created once and reused.
call_model_fn_template = tf.make_template("model_fn",
self._call_model_fn)
logits, train_op = call_model_fn_template(features, labels, mode)
if training and self._subestimator.train_input_fn:
# TODO: Consider tensorflow_estimator/python/estimator/util.py.
inputs = self._subestimator.train_input_fn()
if isinstance(inputs, (tf_compat.DatasetV1, tf_compat.DatasetV2)):
features, labels = tf_compat.make_one_shot_iterator(
inputs).get_next()
else:
features, labels = inputs
# Use different training set for train op only.
_, train_op = call_model_fn_template(features, labels, mode)
self._subnetwork_train_op = train_op
# TODO: Replace with variance complexity measure.
complexity = tf.constant(0.)
return Subnetwork(logits=logits,
last_layer=logits,
persisted_tensors={},
complexity=complexity)
def build_subnetwork(self, features, labels, logits_dimension, training,
iteration_step, summary, previous_ensemble):
# We don't need an EVAL mode since AdaNet takes care of evaluation for us.
mode = tf.estimator.ModeKeys.PREDICT
if training:
mode = tf.estimator.ModeKeys.TRAIN
# Call in template to ensure that variables are created once and reused.
call_model_fn_template = tf.compat.v1.make_template(
"model_fn", self._call_model_fn)
subestimator_features, subestimator_labels = features, labels
local_init_ops = []
if training and self._subestimator.train_input_fn:
# TODO: Consider tensorflow_estimator/python/estimator/util.py.
inputs = self._subestimator.train_input_fn()
if isinstance(inputs, (tf_compat.DatasetV1, tf_compat.DatasetV2)):
subestimator_features, subestimator_labels = (
tf_compat.make_one_shot_iterator(inputs).get_next())
else:
subestimator_features, subestimator_labels = inputs
# Construct subnetwork graph first because of dependencies on scope.
_, _, bagging_train_op_spec, sub_local_init_op = call_model_fn_template(
subestimator_features, subestimator_labels, mode, summary)
# Graph for ensemble learning gets model_fn_1 for scope.
logits, last_layer, _, ensemble_local_init_op = call_model_fn_template(
features, labels, mode, summary)
if sub_local_init_op:
local_init_ops.append(sub_local_init_op)
if ensemble_local_init_op:
local_init_ops.append(ensemble_local_init_op)
# Run train op in a hook so that exceptions can be intercepted by the
# AdaNet framework instead of the Estimator's monitored training session.
hooks = bagging_train_op_spec.hooks + (_SecondaryTrainOpRunnerHook(
bagging_train_op_spec.train_op), )
train_op_spec = TrainOpSpec(
train_op=tf.no_op(),
chief_hooks=bagging_train_op_spec.chief_hooks,
hooks=hooks)
else:
logits, last_layer, train_op_spec, local_init_op = call_model_fn_template(
features, labels, mode, summary)
if local_init_op:
local_init_ops.append(local_init_op)
# TODO: Replace with variance complexity measure.
complexity = tf.constant(0.)
return Subnetwork(logits=logits,
last_layer=last_layer,
shared={"train_op": train_op_spec},
complexity=complexity,
local_init_ops=local_init_ops)
