def _dnn_tree_combined_model_fn(features,
labels,
mode,
head,
dnn_hidden_units,
dnn_feature_columns,
tree_learner_config,
num_trees,
tree_examples_per_layer,
config=None,
dnn_optimizer="Adagrad",
dnn_activation_fn=nn.relu,
dnn_dropout=None,
dnn_input_layer_partitioner=None,
dnn_input_layer_to_tree=True,
dnn_steps_to_train=10000,
tree_feature_columns=None,
tree_center_bias=False,
use_core_versions=False):
"""DNN and GBDT combined model_fn.
Args:
features: `dict` of `Tensor` objects.
labels: Labels used to train on.
mode: Mode we are in. (TRAIN/EVAL/INFER)
head: A `Head` instance.
dnn_hidden_units: List of hidden units per layer.
dnn_feature_columns: An iterable containing all the feature columns
used by the model's DNN.
tree_learner_config: A config for the tree learner.
num_trees: Number of trees to grow model to after training DNN.
tree_examples_per_layer: Number of examples to accumulate before
growing the tree a layer. This value has a big impact on model
quality and should be set equal to the number of examples in
training dataset if possible. It can also be a function that computes
the number of examples based on the depth of the layer that's
being built.
config: `RunConfig` of the estimator.
dnn_optimizer: string, `Optimizer` object, or callable that defines the
optimizer to use for training the DNN. If `None`, will use the Adagrad
optimizer with default learning rate of 0.001.
dnn_activation_fn: Activation function applied to each layer of the DNN.
If `None`, will use `tf.nn.relu`.
dnn_dropout: When not `None`, the probability to drop out a given
unit in the DNN.
dnn_input_layer_partitioner: Partitioner for input layer of the DNN.
Defaults to `min_max_variable_partitioner` with `min_slice_size` 64 << 20.
dnn_input_layer_to_tree: Whether to provide the DNN's input layer
as a feature to the tree.
dnn_steps_to_train: Number of steps to train dnn for before switching
to gbdt.
tree_feature_columns: An iterable containing all the feature columns
used by the model's boosted trees. If dnn_input_layer_to_tree is
set to True, these features are in addition to dnn_feature_columns.
tree_center_bias: Whether a separate tree should be created for
first fitting the bias.
use_core_versions: Whether feature columns and loss are from the core (as
opposed to contrib) version of tensorflow.
Returns:
A `ModelFnOps` object.
Raises:
ValueError: if inputs are not valid.
"""
if not isinstance(features, dict):
raise ValueError("features should be a dictionary of `Tensor`s. "
"Given type: {}".format(type(features)))
if not dnn_feature_columns:
raise ValueError("dnn_feature_columns must be specified")
# Build DNN Logits.
dnn_parent_scope = "dnn"
dnn_partitioner = dnn_input_layer_partitioner or (
partitioned_variables.min_max_variable_partitioner(
max_partitions=config.num_ps_replicas, min_slice_size=64 << 20))
with variable_scope.variable_scope(dnn_parent_scope,
values=tuple(six.itervalues(features)),
partitioner=dnn_partitioner):
with variable_scope.variable_scope(
"input_from_feature_columns",
values=tuple(six.itervalues(features)),
partitioner=dnn_partitioner) as input_layer_scope:
if use_core_versions:
input_layer = feature_column_lib.input_layer(
features=features,
feature_columns=dnn_feature_columns,
weight_collections=[dnn_parent_scope])
else:
input_layer = layers.input_from_feature_columns(
columns_to_tensors=features,
feature_columns=dnn_feature_columns,
weight_collections=[dnn_parent_scope],
scope=input_layer_scope)
previous_layer = input_layer
for layer_id, num_hidden_units in enumerate(dnn_hidden_units):
with variable_scope.variable_scope(
"hiddenlayer_%d" % layer_id,
values=(previous_layer, )) as hidden_layer_scope:
net = layers.fully_connected(
previous_layer,
num_hidden_units,
activation_fn=dnn_activation_fn,
variables_collections=[dnn_parent_scope],
scope=hidden_layer_scope)
if dnn_dropout is not None and mode == model_fn.ModeKeys.TRAIN:
net = layers.dropout(net, keep_prob=(1.0 - dnn_dropout))
_add_hidden_layer_summary(net, hidden_layer_scope.name)
previous_layer = net
with variable_scope.variable_scope(
"logits", values=(previous_layer, )) as logits_scope:
dnn_logits = layers.fully_connected(
previous_layer,
head.logits_dimension,
activation_fn=None,
variables_collections=[dnn_parent_scope],
scope=logits_scope)
_add_hidden_layer_summary(dnn_logits, logits_scope.name)
def _dnn_train_op_fn(loss):
"""Returns the op to optimize the loss."""
return optimizers.optimize_loss(
loss=loss,
global_step=training_util.get_global_step(),
learning_rate=_DNN_LEARNING_RATE,
optimizer=_get_optimizer(dnn_optimizer),
name=dnn_parent_scope,
variables=ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES,
scope=dnn_parent_scope),
# Empty summaries to prevent optimizers from logging training_loss.
summaries=[])
# Build Tree Logits.
global_step = training_util.get_global_step()
with ops.device(global_step.device):
ensemble_handle = model_ops.tree_ensemble_variable(
stamp_token=0,
tree_ensemble_config="", # Initialize an empty ensemble.
name="ensemble_model")
tree_features = features.copy()
if dnn_input_layer_to_tree:
tree_features["dnn_input_layer"] = input_layer
tree_feature_columns.append(
layers.real_valued_column("dnn_input_layer"))
gbdt_model = gbdt_batch.GradientBoostedDecisionTreeModel(
is_chief=config.is_chief,
num_ps_replicas=config.num_ps_replicas,
ensemble_handle=ensemble_handle,
center_bias=tree_center_bias,
examples_per_layer=tree_examples_per_layer,
learner_config=tree_learner_config,
feature_columns=tree_feature_columns,
logits_dimension=head.logits_dimension,
features=tree_features)
with ops.name_scope("gbdt"):
predictions_dict = gbdt_model.predict(mode)
tree_logits = predictions_dict["predictions"]
def _tree_train_op_fn(loss):
"""Returns the op to optimize the loss."""
update_op = gbdt_model.train(loss, predictions_dict, labels)
with ops.control_dependencies(
[update_op]), (ops.colocate_with(global_step)):
update_op = state_ops.assign_add(global_step, 1).op
return update_op
tree_train_logits = dnn_logits + tree_logits
def _no_train_op_fn(loss):
"""Returns a no-op."""
del loss
return control_flow_ops.no_op()
if use_core_versions:
model_fn_ops = head.create_estimator_spec(features=features,
mode=mode,
labels=labels,
train_op_fn=_no_train_op_fn,
logits=tree_train_logits)
dnn_train_op = head.create_estimator_spec(features=features,
mode=mode,
labels=labels,
train_op_fn=_dnn_train_op_fn,
logits=dnn_logits)
dnn_train_op = estimator_utils.estimator_spec_to_model_fn_ops(
dnn_train_op).train_op
tree_train_op = head.create_estimator_spec(
features=tree_features,
mode=mode,
labels=labels,
train_op_fn=_tree_train_op_fn,
logits=tree_train_logits)
tree_train_op = estimator_utils.estimator_spec_to_model_fn_ops(
tree_train_op).train_op
model_fn_ops = estimator_utils.estimator_spec_to_model_fn_ops(
model_fn_ops)
else:
model_fn_ops = head.create_model_fn_ops(features=features,
mode=mode,
labels=labels,
train_op_fn=_no_train_op_fn,
logits=tree_train_logits)
dnn_train_op = head.create_model_fn_ops(features=features,
mode=mode,
labels=labels,
train_op_fn=_dnn_train_op_fn,
logits=dnn_logits).train_op
tree_train_op = head.create_model_fn_ops(
features=tree_features,
mode=mode,
labels=labels,
train_op_fn=_tree_train_op_fn,
logits=tree_train_logits).train_op
if tree_center_bias:
num_trees += 1
finalized_trees, attempted_trees = gbdt_model.get_number_of_trees_tensor()
model_fn_ops.training_hooks.extend([
trainer_hooks.SwitchTrainOp(dnn_train_op, dnn_steps_to_train,
tree_train_op),
trainer_hooks.StopAfterNTrees(num_trees, attempted_trees,
finalized_trees)
])
return model_fn_ops
def _dnn_tree_combined_model_fn(
features,
labels,
mode,
head,
dnn_hidden_units,
dnn_feature_columns,
tree_learner_config,
num_trees,
tree_examples_per_layer,
config=None,
dnn_optimizer="Adagrad",
dnn_activation_fn=nn.relu,
dnn_dropout=None,
dnn_input_layer_partitioner=None,
dnn_input_layer_to_tree=True,
dnn_steps_to_train=10000,
predict_with_tree_only=False,
tree_feature_columns=None,
tree_center_bias=False,
dnn_to_tree_distillation_param=None,
use_core_versions=False,
output_type=model.ModelBuilderOutputType.MODEL_FN_OPS,
override_global_step_value=None):
"""DNN and GBDT combined model_fn.
Args:
features: `dict` of `Tensor` objects.
labels: Labels used to train on.
mode: Mode we are in. (TRAIN/EVAL/INFER)
head: A `Head` instance.
dnn_hidden_units: List of hidden units per layer.
dnn_feature_columns: An iterable containing all the feature columns
used by the model's DNN.
tree_learner_config: A config for the tree learner.
num_trees: Number of trees to grow model to after training DNN.
tree_examples_per_layer: Number of examples to accumulate before
growing the tree a layer. This value has a big impact on model
quality and should be set equal to the number of examples in
training dataset if possible. It can also be a function that computes
the number of examples based on the depth of the layer that's
being built.
config: `RunConfig` of the estimator.
dnn_optimizer: string, `Optimizer` object, or callable that defines the
optimizer to use for training the DNN. If `None`, will use the Adagrad
optimizer with default learning rate of 0.001.
dnn_activation_fn: Activation function applied to each layer of the DNN.
If `None`, will use `tf.nn.relu`.
dnn_dropout: When not `None`, the probability to drop out a given
unit in the DNN.
dnn_input_layer_partitioner: Partitioner for input layer of the DNN.
Defaults to `min_max_variable_partitioner` with `min_slice_size` 64 << 20.
dnn_input_layer_to_tree: Whether to provide the DNN's input layer
as a feature to the tree.
dnn_steps_to_train: Number of steps to train dnn for before switching
to gbdt.
predict_with_tree_only: Whether to use only the tree model output as the
final prediction.
tree_feature_columns: An iterable containing all the feature columns
used by the model's boosted trees. If dnn_input_layer_to_tree is
set to True, these features are in addition to dnn_feature_columns.
tree_center_bias: Whether a separate tree should be created for
first fitting the bias.
dnn_to_tree_distillation_param: A Tuple of (float, loss_fn), where the
float defines the weight of the distillation loss, and the loss_fn, for
computing distillation loss, takes dnn_logits, tree_logits and weight
tensor. If the entire tuple is None, no distillation will be applied. If
only the loss_fn is None, we will take the sigmoid/softmax cross entropy
loss be default. When distillation is applied, `predict_with_tree_only`
will be set to True.
use_core_versions: Whether feature columns and loss are from the core (as
opposed to contrib) version of tensorflow.
output_type: Whether to return ModelFnOps (old interface) or EstimatorSpec
(new interface).
override_global_step_value: If after the training is done, global step
value must be reset to this value. This is particularly useful for hyper
parameter tuning, which can't recognize early stopping due to the number
of trees. If None, no override of global step will happen.
Returns:
A `ModelFnOps` object.
Raises:
ValueError: if inputs are not valid.
"""
if not isinstance(features, dict):
raise ValueError("features should be a dictionary of `Tensor`s. "
"Given type: {}".format(type(features)))
if not dnn_feature_columns:
raise ValueError("dnn_feature_columns must be specified")
if dnn_to_tree_distillation_param:
if not predict_with_tree_only:
logging.warning(
"update predict_with_tree_only to True since distillation"
"is specified.")
predict_with_tree_only = True
# Build DNN Logits.
dnn_parent_scope = "dnn"
dnn_partitioner = dnn_input_layer_partitioner or (
partitioned_variables.min_max_variable_partitioner(
max_partitions=config.num_ps_replicas, min_slice_size=64 << 20))
if (output_type == model.ModelBuilderOutputType.ESTIMATOR_SPEC
and not use_core_versions):
raise ValueError("You must use core versions with Estimator Spec")
with variable_scope.variable_scope(dnn_parent_scope,
values=tuple(six.itervalues(features)),
partitioner=dnn_partitioner):
with variable_scope.variable_scope(
"input_from_feature_columns",
values=tuple(six.itervalues(features)),
partitioner=dnn_partitioner) as input_layer_scope:
if use_core_versions:
input_layer = feature_column_lib.input_layer(
features=features,
feature_columns=dnn_feature_columns,
weight_collections=[dnn_parent_scope])
else:
input_layer = layers.input_from_feature_columns(
columns_to_tensors=features,
feature_columns=dnn_feature_columns,
weight_collections=[dnn_parent_scope],
scope=input_layer_scope)
previous_layer = input_layer
for layer_id, num_hidden_units in enumerate(dnn_hidden_units):
with variable_scope.variable_scope(
"hiddenlayer_%d" % layer_id,
values=(previous_layer, )) as hidden_layer_scope:
net = layers.fully_connected(
previous_layer,
num_hidden_units,
activation_fn=dnn_activation_fn,
variables_collections=[dnn_parent_scope],
scope=hidden_layer_scope)
if dnn_dropout is not None and mode == model_fn.ModeKeys.TRAIN:
net = layers.dropout(net, keep_prob=(1.0 - dnn_dropout))
_add_hidden_layer_summary(net, hidden_layer_scope.name)
previous_layer = net
with variable_scope.variable_scope(
"logits", values=(previous_layer, )) as logits_scope:
dnn_logits = layers.fully_connected(
previous_layer,
head.logits_dimension,
activation_fn=None,
variables_collections=[dnn_parent_scope],
scope=logits_scope)
_add_hidden_layer_summary(dnn_logits, logits_scope.name)
def _dnn_train_op_fn(loss):
"""Returns the op to optimize the loss."""
return optimizers.optimize_loss(
loss=loss,
global_step=training_util.get_global_step(),
learning_rate=_DNN_LEARNING_RATE,
optimizer=_get_optimizer(dnn_optimizer),
name=dnn_parent_scope,
variables=ops.get_collection(ops.GraphKeys.TRAINABLE_VARIABLES,
scope=dnn_parent_scope),
# Empty summaries to prevent optimizers from logging training_loss.
summaries=[])
# Build Tree Logits.
global_step = training_util.get_global_step()
with ops.device(global_step.device):
ensemble_handle = model_ops.tree_ensemble_variable(
stamp_token=0,
tree_ensemble_config="", # Initialize an empty ensemble.
name="ensemble_model")
tree_features = features.copy()
if dnn_input_layer_to_tree:
tree_features["dnn_input_layer"] = input_layer
tree_feature_columns.append(
layers.real_valued_column("dnn_input_layer"))
gbdt_model = gbdt_batch.GradientBoostedDecisionTreeModel(
is_chief=config.is_chief,
num_ps_replicas=config.num_ps_replicas,
ensemble_handle=ensemble_handle,
center_bias=tree_center_bias,
examples_per_layer=tree_examples_per_layer,
learner_config=tree_learner_config,
feature_columns=tree_feature_columns,
logits_dimension=head.logits_dimension,
features=tree_features,
use_core_columns=use_core_versions)
with ops.name_scope("gbdt"):
predictions_dict = gbdt_model.predict(mode)
tree_logits = predictions_dict["predictions"]
def _tree_train_op_fn(loss):
"""Returns the op to optimize the loss."""
if dnn_to_tree_distillation_param:
loss_weight, loss_fn = dnn_to_tree_distillation_param
weight_tensor = head_lib._weight_tensor( # pylint: disable=protected-access
features, head.weight_column_name)
dnn_logits_fixed = array_ops.stop_gradient(dnn_logits)
if loss_fn is None:
# we create the loss_fn similar to the head loss_fn for
# multi_class_head used previously as the default one.
n_classes = 2 if head.logits_dimension == 1 else head.logits_dimension
loss_fn = distillation_loss.create_dnn_to_tree_cross_entropy_loss_fn(
n_classes)
dnn_to_tree_distillation_loss = loss_weight * loss_fn(
dnn_logits_fixed, tree_logits, weight_tensor)
summary.scalar("dnn_to_tree_distillation_loss",
dnn_to_tree_distillation_loss)
loss += dnn_to_tree_distillation_loss
update_op = gbdt_model.train(loss, predictions_dict, labels)
with ops.control_dependencies(
[update_op]), (ops.colocate_with(global_step)):
update_op = state_ops.assign_add(global_step, 1).op
return update_op
if predict_with_tree_only:
if mode == model_fn.ModeKeys.TRAIN or mode == model_fn.ModeKeys.INFER:
tree_train_logits = tree_logits
else:
tree_train_logits = control_flow_ops.cond(
global_step > dnn_steps_to_train, lambda: tree_logits,
lambda: dnn_logits)
else:
tree_train_logits = dnn_logits + tree_logits
def _no_train_op_fn(loss):
"""Returns a no-op."""
del loss
return control_flow_ops.no_op()
if tree_center_bias:
num_trees += 1
finalized_trees, attempted_trees = gbdt_model.get_number_of_trees_tensor()
if output_type == model.ModelBuilderOutputType.MODEL_FN_OPS:
if use_core_versions:
model_fn_ops = head.create_estimator_spec(
features=features,
mode=mode,
labels=labels,
train_op_fn=_no_train_op_fn,
logits=tree_train_logits)
dnn_train_op = head.create_estimator_spec(
features=features,
mode=mode,
labels=labels,
train_op_fn=_dnn_train_op_fn,
logits=dnn_logits)
dnn_train_op = estimator_utils.estimator_spec_to_model_fn_ops(
dnn_train_op).train_op
tree_train_op = head.create_estimator_spec(
features=tree_features,
mode=mode,
labels=labels,
train_op_fn=_tree_train_op_fn,
logits=tree_train_logits)
tree_train_op = estimator_utils.estimator_spec_to_model_fn_ops(
tree_train_op).train_op
model_fn_ops = estimator_utils.estimator_spec_to_model_fn_ops(
model_fn_ops)
else:
model_fn_ops = head.create_model_fn_ops(
features=features,
mode=mode,
labels=labels,
train_op_fn=_no_train_op_fn,
logits=tree_train_logits)
dnn_train_op = head.create_model_fn_ops(
features=features,
mode=mode,
labels=labels,
train_op_fn=_dnn_train_op_fn,
logits=dnn_logits).train_op
tree_train_op = head.create_model_fn_ops(
features=tree_features,
mode=mode,
labels=labels,
train_op_fn=_tree_train_op_fn,
logits=tree_train_logits).train_op
# Add the hooks
model_fn_ops.training_hooks.extend([
trainer_hooks.SwitchTrainOp(dnn_train_op, dnn_steps_to_train,
tree_train_op),
trainer_hooks.StopAfterNTrees(num_trees, attempted_trees,
finalized_trees,
override_global_step_value)
])
return model_fn_ops
elif output_type == model.ModelBuilderOutputType.ESTIMATOR_SPEC:
fusion_spec = head.create_estimator_spec(features=features,
mode=mode,
labels=labels,
train_op_fn=_no_train_op_fn,
logits=tree_train_logits)
dnn_spec = head.create_estimator_spec(features=features,
mode=mode,
labels=labels,
train_op_fn=_dnn_train_op_fn,
logits=dnn_logits)
tree_spec = head.create_estimator_spec(features=tree_features,
mode=mode,
labels=labels,
train_op_fn=_tree_train_op_fn,
logits=tree_train_logits)
training_hooks = [
trainer_hooks.SwitchTrainOp(dnn_spec.train_op, dnn_steps_to_train,
tree_spec.train_op),
trainer_hooks.StopAfterNTrees(num_trees, attempted_trees,
finalized_trees,
override_global_step_value)
]
fusion_spec = fusion_spec._replace(training_hooks=training_hooks +
list(fusion_spec.training_hooks))
return fusion_spec