1# Copyright 2018 The TensorFlow Authors. All Rights Reserved. 2# 3# Licensed under the Apache License, Version 2.0 (the "License"); 4# you may not use this file except in compliance with the License. 5# You may obtain a copy of the License at 6# 7# http://www.apache.org/licenses/LICENSE-2.0 8# 9# Unless required by applicable law or agreed to in writing, software 10# distributed under the License is distributed on an "AS IS" BASIS, 11# WITHOUT WARRANTIES OR CONDITIONS OF ANY KIND, either express or implied. 12# See the License for the specific language governing permissions and 13# limitations under the License. 14# ============================================================================== 15"""Keras SavedModel serialization.""" 16 17import os 18 19from tensorflow.python.keras import backend as K 20from tensorflow.python.keras.protobuf import saved_metadata_pb2 21from tensorflow.python.keras.protobuf import versions_pb2 22from tensorflow.python.keras.saving import saving_utils 23from tensorflow.python.keras.saving.saved_model import constants 24from tensorflow.python.keras.saving.saved_model import save_impl 25from tensorflow.python.keras.saving.saved_model import utils 26from tensorflow.python.keras.utils.generic_utils import LazyLoader 27from tensorflow.python.keras.utils.io_utils import ask_to_proceed_with_overwrite 28from tensorflow.python.platform import gfile 29from tensorflow.python.saved_model import save as save_lib 30 31# To avoid circular dependencies between keras/engine and keras/saving, 32# code in keras/saving must delay imports. 33 34base_layer = LazyLoader( 35 "base_layer", globals(), 36 "tensorflow.python.keras.engine.base_layer") 37training_lib = LazyLoader( 38 "training_lib", globals(), 39 "tensorflow.python.keras.engine.training") 40 41 42def save(model, filepath, overwrite, include_optimizer, signatures=None, 43 options=None, save_traces=True): 44 """Saves a model as a SavedModel to the filepath. 45 46 Args: 47 model: Keras model instance to be saved. 48 filepath: String path to save the model. 49 overwrite: whether to overwrite the existing filepath. 50 include_optimizer: If True, save the model's optimizer state. 51 signatures: Signatures to save with the SavedModel. Applicable to the 'tf' 52 format only. Please see the `signatures` argument in `tf.saved_model.save` 53 for details. 54 options: (only applies to SavedModel format) `tf.saved_model.SaveOptions` 55 object that specifies options for saving to SavedModel. 56 save_traces: (only applies to SavedModel format) When enabled, the 57 SavedModel will store the function traces for each layer. This 58 can be disabled, so that only the configs of each layer are stored. 59 Defaults to `True`. Disabling this will decrease serialization time 60 and reduce file size, but it requires that all custom layers/models 61 implement a `get_config()` method. 62 63 Raises: 64 ValueError: if the model's inputs have not been defined. 65 """ 66 # If file exists and should not be overwritten. 67 if not overwrite and os.path.exists(filepath): 68 proceed = ask_to_proceed_with_overwrite(filepath) 69 if not proceed: 70 return 71 72 if save_traces: 73 if save_impl.should_skip_serialization(model): 74 saving_utils.raise_model_input_error(model) 75 76 if not include_optimizer: 77 orig_optimizer = model.optimizer 78 model.optimizer = None 79 # TODO(b/180760306) Change to del model.optimizer if Layer's __delattr__ 80 # calls AutoTrackable's __delattr__. 81 model._delete_tracking("optimizer") # pylint: disable=protected-access 82 83 # Trace all functions and signatures with `training=0` instead of using an 84 # already-set learning phase placeholder. 85 # This is needed for compatibility reasons until learning phase setting 86 # is removed from the public apis. 87 with K.deprecated_internal_learning_phase_scope(0): 88 with utils.keras_option_scope(save_traces): 89 saved_nodes, node_paths = save_lib.save_and_return_nodes( 90 model, filepath, signatures, options) 91 92 # Save all metadata to a separate file in the SavedModel directory. 93 metadata = generate_keras_metadata(saved_nodes, node_paths) 94 95 with gfile.GFile( 96 os.path.join(filepath, constants.SAVED_METADATA_PATH), "wb") as w: 97 w.write(metadata.SerializeToString(deterministic=True)) 98 99 if not include_optimizer: 100 model.optimizer = orig_optimizer 101 102 103def generate_keras_metadata(saved_nodes, node_paths): 104 """Constructs a KerasMetadata proto with the metadata of each keras object.""" 105 metadata = saved_metadata_pb2.SavedMetadata() 106 107 for node_id, node in enumerate(saved_nodes): 108 if isinstance(node, base_layer.Layer): 109 path = node_paths[node] 110 if not path: 111 node_path = "root" 112 else: 113 node_path = "root.{}".format( 114 ".".join([ref.name for ref in path])) 115 116 metadata.nodes.add( 117 node_id=node_id, 118 node_path=node_path, 119 version=versions_pb2.VersionDef( 120 producer=1, min_consumer=1, bad_consumers=[]), 121 identifier=node._object_identifier, # pylint: disable=protected-access 122 metadata=node._tracking_metadata) # pylint: disable=protected-access 123 124 return metadata 125