def build_inputs(input_features_def: List[Dict[str, Any]]) -> Dict[str, InputFeature]:
"""Builds and returns input features in topological order."""
input_features = OrderedDict()
input_features_def = topological_sort_feature_dependencies(input_features_def)
for input_feature_def in input_features_def:
input_features[input_feature_def[NAME]] = build_single_input(input_feature_def, input_features)
return input_features
def build_inputs(input_features_def, **kwargs):
input_features = OrderedDict()
input_features_def = topological_sort_feature_dependencies(
input_features_def)
for input_feature_def in input_features_def:
input_features[input_feature_def['name']] = build_single_input(
input_feature_def, input_features, **kwargs)
return input_features
def build_inputs(input_features_def: List[Dict[str, Any]],
**kwargs) -> Dict[str, InputFeature]:
input_features = OrderedDict()
input_features_def = topological_sort_feature_dependencies(
input_features_def)
for input_feature_def in input_features_def:
input_features[input_feature_def[NAME]] = build_single_input(
input_feature_def, input_features, **kwargs)
return input_features
def build_outputs(output_features_def, combiner, **kwargs):
output_features_def = topological_sort_feature_dependencies(output_features_def)
output_features = {}
for output_feature_def in output_features_def:
output_feature_def["input_size"] = combiner.output_shape[-1]
output_feature = build_single_output(output_feature_def, combiner, output_features, **kwargs)
output_features[output_feature_def[NAME]] = output_feature
return output_features
def build_outputs(output_features_def, combiner, **kwargs):
output_features_def = topological_sort_feature_dependencies(
output_features_def)
output_features = {}
for output_feature_def in output_features_def:
output_feature = build_single_output(output_feature_def, combiner,
output_features, **kwargs)
output_features[output_feature_def['name']] = output_feature
return output_features
def build_outputs(output_features_def: List[Dict[str, Any]], combiner: Combiner) -> Dict[str, OutputFeature]:
"""Builds and returns output features in topological order."""
output_features_def = topological_sort_feature_dependencies(output_features_def)
output_features = {}
for output_feature_def in output_features_def:
# TODO(Justin): Check that the semantics of input_size align with what the combiner's output shape returns for
# seq2seq.
output_feature_def["input_size"] = combiner.output_shape[-1]
output_feature = build_single_output(output_feature_def, output_features)
output_features[output_feature_def[NAME]] = output_feature
return output_features
def build_inputs(input_features,
regularizer,
dropout_rate,
is_training=True,
**kwargs):
# ================ Inputs =============
feature_representations = OrderedDict()
input_features = topological_sort_feature_dependencies(input_features)
for input_feature in input_features:
feature_representation = build_single_input(input_feature,
regularizer,
dropout_rate,
is_training=is_training,
**kwargs)
feature_representations[input_feature['name']] = feature_representation
return feature_representations
def build_outputs(output_features,
hidden,
hidden_size,
regularizer,
dropout_rate,
is_training=True,
**kwargs):
output_features = topological_sort_feature_dependencies(output_features)
output_tensors = OrderedDict()
final_hidden = {}
output_train_losses = []
output_eval_losses = []
for output_feature in output_features:
of_train_mean_loss, of_eval_loss, of_output_tensors = build_single_output(
output_feature, hidden, hidden_size, final_hidden, dropout_rate,
regularizer, is_training, **kwargs)
output_train_losses.append(of_train_mean_loss)
output_eval_losses.append(of_eval_loss)
output_tensors.update(of_output_tensors)
train_combined_mean_loss = tf.reduce_sum(tf.stack(output_train_losses),
name='train_combined_mean_loss')
if regularizer is not None:
regularization_losses = tf.get_collection(
tf.GraphKeys.REGULARIZATION_LOSSES)
if regularization_losses:
regularization_loss = tf.add_n(regularization_losses)
logger.debug(
'- Regularization losses: {0}'.format(regularization_losses))
else:
regularization_loss = tf.constant(0.0)
else:
regularization_loss = tf.constant(0.0)
train_reg_mean_loss = tf.add(train_combined_mean_loss,
regularization_loss,
name='train_combined_regularized_mean_loss')
eval_combined_loss = tf.reduce_sum(tf.stack(output_eval_losses),
axis=0,
name='eval_combined_loss')
return train_reg_mean_loss, eval_combined_loss, regularization_loss, output_tensors
