def _deserialize_keras_model(model_topology_json,
weight_entries=None,
use_unique_name_scope=False):
"""Internal helper method for deserializing a Keras Model.
Args:
model_topology_json: content of the JSON containing model topology, in
Keras (i.e., tfjs-layers) format. It can be any of the following types:
- A JSON object, i.e., a `dict`.
- A `str` or `buffer`, in which case it will be parsed as a JSON object.
- A `file` object or `file`-like object containing the JSON, in which
case it will be read with the `read()` method and the content parsed
as a JSON object.
weight_entries: Weight entries, in tensorflow.js format, as a `list`.
use_unique_name_scope: Use a unique ID as the name scope for the loaded
model. This may facilitate loading of multiple Keras models in the
same TensorFlow Graph or Session context. Default: `False`.
"""
if isinstance(model_topology_json, (six.string_types, bytes)):
model_topology_json = json.loads(tf.compat.as_text(model_topology_json))
elif not isinstance(model_topology_json, dict):
model_topology_json = json.load(model_topology_json)
is_tf_keras = ('keras_version' in model_topology_json and
model_topology_json['keras_version'].endswith('-tf'))
if 'model_config' in model_topology_json:
model_topology_json = model_topology_json['model_config']
unique_name_scope = uuid.uuid4().hex if use_unique_name_scope else None
with tf.compat.v1.name_scope(unique_name_scope):
with CustomObjectScope({'relu6': relu6}):
if is_tf_keras:
model = keras.models.model_from_json(json.dumps(model_topology_json))
else:
model = keras.models.model_from_json(json.dumps(model_topology_json))
if weight_entries:
weights_dict = dict()
for weight_entry in weight_entries:
weights_dict[weight_entry['name']] = weight_entry['data']
# Collect weight names from the model, in the same order as the internal
# ordering of model.set_weights() used below.
weight_names = []
for layer in model.layers:
for w in layer.weights:
weight_names.append(
keras_h5_conversion.normalize_weight_name(
w.name[len(unique_name_scope) + 1:])
if use_unique_name_scope
else keras_h5_conversion.normalize_weight_name(w.name))
# Prepare list of weight values for calling set_weights().
weights_list = [weights_dict[name] for name in weight_names]
model.set_weights(weights_list)
return model
def _deserialize_keras_model(model_topology_json,
weight_entries=None,
use_unique_name_scope=False):
"""Internal helper method for deserializing a Keras Model.
Args:
model_topology_json: content of the JSON containing model topology, in
Keras (i.e., tfjs-layers) format. It can be any of the following types:
- A JSON object, i.e., a `dict`.
- A `str` or `buffer`, in which case it will be parsed as a JSON object.
- A `file` object or `file`-like object containing the JSON, in which
case it will be read with the `read()` method and the content parsed
as a JSON object.
weight_entries: Weight entries, in tensorflow.js format, as a `list`.
use_unique_name_scope: Use a unique ID as the name scope for the loaded
model. This may facilitate loading of multiple Keras models in the
same TensorFlow Graph or Session context. Default: `False`.
"""
if isinstance(model_topology_json, (six.string_types, bytes)):
model_topology_json = json.loads(tf.compat.as_text(model_topology_json))
elif not isinstance(model_topology_json, dict):
model_topology_json = json.load(model_topology_json)
is_tf_keras = ('keras_version' in model_topology_json and
model_topology_json['keras_version'].endswith('-tf'))
if 'model_config' in model_topology_json:
model_topology_json = model_topology_json['model_config']
unique_name_scope = uuid.uuid4().hex if use_unique_name_scope else None
with tf.name_scope(unique_name_scope):
if is_tf_keras:
model = tf.keras.models.model_from_json(json.dumps(model_topology_json))
else:
model = keras.models.model_from_json(json.dumps(model_topology_json))
if weight_entries:
weights_dict = dict()
for weight_entry in weight_entries:
weights_dict[weight_entry['name']] = weight_entry['data']
# Collect weight names from the model, in the same order as the internal
# ordering of model.set_weights() used below.
weight_names = []
for layer in model.layers:
for w in layer.weights:
weight_names.append(
keras_h5_conversion.normalize_weight_name(
w.name[len(unique_name_scope) + 1:])
if use_unique_name_scope
else keras_h5_conversion.normalize_weight_name(w.name))
# Prepare list of weight values for calling set_weights().
weights_list = [weights_dict[name] for name in weight_names]
model.set_weights(weights_list)
return model
def load_keras_model(config_json_path,
weights_path_prefix=None,
weights_data_buffers=None,
load_weights=True,
use_unique_name_scope=False):
"""Load a Keras Model from TensorFlow.js-format artifacts.
Args:
config_json_path: Path to the TensorFlow.js-format JSON file that includes
the model topology and weights manifest.
weights_path_prefix: Optional path prefix for the weights files.
If not specified (`None`), will assume the prefix is the same directory
as the dirname of `config_json_path`.
weights_data_buffers: A buffer of a `list` of buffers containing the weight
values concatenated and sharded in the order as specified by the
weights manifest at `config_json_path`. This argument is mutually
exclusive with `weights_path_prefix`.
load_weights: Whether the weights are to be loaded according
to the weights manifest at `config_json_path`. Default: `True`.
use_unique_name_scope: Use a unique ID as the name scope for the loaded
model. This may facilitate loading of multiple Keras models in the
same TensorFlow Graph or Session context. Default: `False`.
Returns:
The loaded instance of `keras.Model`.
Raises:
TypeError, if the format of the JSON content of `config_json_path` has an
invalid format.
KeyError, if required keys do not exist in the JSON content of
`config_json_path`.
ValueError, if both `weights_data_buffers` and `weights_path_prefix` are
provided.
"""
with open(config_json_path, 'rt') as f:
model_and_weights_manifest = json.load(f)
if not isinstance(model_and_weights_manifest, dict):
raise TypeError(
'The JSON content of %s is required to be a `dict`, but found %s' %
(config_json_path, type(model_and_weights_manifest)))
if 'modelTopology' not in model_and_weights_manifest:
raise KeyError(
'Field "modelTopology" is missing from the JSON content in %s' %
config_json_path)
model_json = model_and_weights_manifest['modelTopology']
if 'model_config' in model_json:
model_json = model_json['model_config']
unique_name_scope = uuid.uuid4().hex if use_unique_name_scope else None
with tf.name_scope(unique_name_scope):
model = keras.models.model_from_json(json.dumps(model_json))
if load_weights:
if 'weightsManifest' not in model_and_weights_manifest:
raise KeyError(
'Field "weightsManifest" is missing from the JSON content in %s'
% config_json_path)
weights_manifest = model_and_weights_manifest['weightsManifest']
if weights_data_buffers:
if weights_path_prefix:
raise ValueError(
'The arguments weights_data_buffers and weights_path_prefix are '
'mutually exclusive and should not be both specified.')
weight_entries = read_weights.decode_weights(weights_manifest,
weights_data_buffers,
flatten=True)
else:
weight_names = [
keras_h5_conversion.normalize_weight_name(
w.name[len(unique_name_scope) +
1:]) if use_unique_name_scope else
keras_h5_conversion.normalize_weight_name(w.name[:-2])
for w in model.weights
]
if not weights_path_prefix:
weights_path_prefix = os.path.dirname(config_json_path)
if not os.path.isdir(weights_path_prefix):
raise ValueError(
'Weights path prefix is not an existing directory: %s' %
weights_path_prefix)
weight_entries = read_weights.read_weights(weights_manifest,
weights_path_prefix,
flatten=True)
weights_dict = dict()
for weight_entry in weight_entries:
weights_dict[weight_entry['name']] = weight_entry['data']
weights_list = []
for weight_name in weight_names:
weights_list.append(weights_dict[weight_name])
model.set_weights(weights_list)
return model