def from_session(cls, sess, input_tensors, output_tensors):
"""Creates a TFLiteConverter class from a TensorFlow Session.
Args:
sess: TensorFlow Session.
input_tensors: List of input tensors. Type and shape are computed using
`foo.shape` and `foo.dtype`.
output_tensors: List of output tensors (only .name is used from this).
Returns:
TFLiteConverter class.
"""
graph_def = _freeze_graph(sess, input_tensors, output_tensors)
return cls(graph_def,
input_tensors,
output_tensors,
experimental_debug_info_func=_build_debug_info_func(
sess.graph))
def from_saved_model(cls,
saved_model_dir,
input_arrays=None,
input_shapes=None,
output_arrays=None,
tag_set=None,
signature_key=None):
"""Creates a TFLiteConverter class from a SavedModel.
Args:
saved_model_dir: SavedModel directory to convert.
input_arrays: List of input tensors to freeze graph with. Uses input
arrays from SignatureDef when none are provided. (default None)
input_shapes: Dict of strings representing input tensor names to list of
integers representing input shapes (e.g., {"foo" : [1, 16, 16, 3]}).
Automatically determined when input shapes is None (e.g., {"foo" :
None}). (default None)
output_arrays: List of output tensors to freeze graph with. Uses output
arrays from SignatureDef when none are provided. (default None)
tag_set: Set of tags identifying the MetaGraphDef within the SavedModel to
analyze. All tags in the tag set must be present. (default set("serve"))
signature_key: Key identifying SignatureDef containing inputs and outputs.
(default DEFAULT_SERVING_SIGNATURE_DEF_KEY)
Returns:
TFLiteConverter class.
"""
if tag_set is None:
tag_set = set([_tag_constants.SERVING])
if signature_key is None:
signature_key = _signature_constants.DEFAULT_SERVING_SIGNATURE_DEF_KEY
result = _freeze_saved_model(saved_model_dir, input_arrays,
input_shapes, output_arrays, tag_set,
signature_key)
return cls(graph_def=result[0],
input_tensors=result[1],
output_tensors=result[2],
experimental_debug_info_func=_build_debug_info_func(
result[3]))
def from_keras_model_file(cls,
model_file,
input_arrays=None,
input_shapes=None,
output_arrays=None,
custom_objects=None):
"""Creates a TFLiteConverter class from a tf.keras model file.
Args:
model_file: Full filepath of HDF5 file containing the tf.keras model.
input_arrays: List of input tensors to freeze graph with. Uses input
arrays from SignatureDef when none are provided. (default None)
input_shapes: Dict of strings representing input tensor names to list of
integers representing input shapes (e.g., {"foo" : [1, 16, 16, 3]}).
Automatically determined when input shapes is None (e.g., {"foo" :
None}). (default None)
output_arrays: List of output tensors to freeze graph with. Uses output
arrays from SignatureDef when none are provided. (default None)
custom_objects: Dict mapping names (strings) to custom classes or
functions to be considered during model deserialization. (default None)
Returns:
TFLiteConverter class.
"""
# Handles Keras when Eager mode is enabled.
if context.executing_eagerly():
if input_arrays or output_arrays:
raise ValueError(
"`input_arrays` and `output_arrays` are unsupported "
"with Eager mode. If your model requires any of these "
"parameters, please use disable_eager_execution().")
_keras.backend.set_learning_phase(False)
keras_model = _keras.models.load_model(model_file, custom_objects)
function = _saving_utils.trace_model_call(keras_model)
concrete_func = function.get_concrete_function()
frozen_func = _convert_to_constants.convert_variables_to_constants_v2(
concrete_func, lower_control_flow=False)
_set_tensor_shapes(frozen_func.inputs, input_shapes)
return cls(frozen_func.graph.as_graph_def(),
frozen_func.inputs,
frozen_func.outputs,
experimental_debug_info_func=_build_debug_info_func(
frozen_func.graph))
# Handles Keras when Eager mode is disabled.
_keras.backend.clear_session()
_keras.backend.set_learning_phase(False)
keras_model = _keras.models.load_model(model_file, custom_objects)
sess = _keras.backend.get_session()
# Get input and output tensors.
if input_arrays:
input_tensors = _get_tensors_from_tensor_names(
sess.graph, input_arrays)
else:
input_tensors = keras_model.inputs
if output_arrays:
output_tensors = _get_tensors_from_tensor_names(
sess.graph, output_arrays)
else:
output_tensors = keras_model.outputs
_set_tensor_shapes(input_tensors, input_shapes)
graph_def = _freeze_graph(sess, input_tensors, output_tensors)
return cls(graph_def,
input_tensors,
output_tensors,
experimental_debug_info_func=_build_debug_info_func(
sess.graph))
def convert(self):
"""Converts a TensorFlow GraphDef based on instance variables.
Returns:
The converted data in serialized format.
Raises:
ValueError:
Multiple concrete functions are specified.
Input shape is not specified.
Invalid quantization parameters.
"""
# TODO(b/130297984): Add support for converting multiple function.
if len(self._funcs) != 1:
raise ValueError(
"This converter can only convert a single "
"ConcreteFunction. Converting multiple functions is "
"under development.")
frozen_func = _convert_to_constants.convert_variables_to_constants_v2(
self._funcs[0], lower_control_flow=False)
input_tensors = [
tensor for tensor in frozen_func.inputs
if tensor.dtype != _dtypes.resource
]
output_tensors = frozen_func.outputs
# Run a Grappler pass.
graph_def = frozen_func.graph.as_graph_def()
graph_def = _run_graph_optimizations(graph_def,
input_tensors,
output_tensors,
config=self._grappler_config(),
graph=frozen_func.graph)
# Checks dimensions in input tensor.
for tensor in input_tensors:
# Note that shape_list might be empty for scalar shapes.
shape_list = tensor.shape.as_list()
if None in shape_list[1:]:
raise ValueError(
"None is only supported in the 1st dimension. Tensor '{0}' has "
"invalid shape '{1}'.".format(_get_tensor_name(tensor),
shape_list))
elif shape_list and shape_list[0] is None:
# Set the batch size to 1 if undefined.
shape = tensor.shape.as_list()
shape[0] = 1
tensor.set_shape(shape)
self._validate_quantization()
self._validate_representative_dataset()
self._debug_info = _get_debug_info(
_build_debug_info_func(self._funcs[0].graph), graph_def)
converter_kwargs = self._get_base_converter_args()
# Converts model.
result = _toco_convert_impl(input_data=graph_def,
input_tensors=input_tensors,
output_tensors=output_tensors,
**converter_kwargs)
if self._is_calibration_quantize():
result = self._calibrate_quantize_model(result, constants.FLOAT,
constants.FLOAT)
return result
def convert(self):
"""Converts a TensorFlow GraphDef based on instance variables.
Returns:
The converted data in serialized format.
Raises:
ValueError:
Multiple concrete functions are specified.
Input shape is not specified.
Invalid quantization parameters.
"""
# TODO(b/130297984): Add support for converting multiple function.
if len(self._funcs) != 1:
raise ValueError("This converter can only convert a single "
"ConcreteFunction. Converting multiple functions is "
"under development.")
# graph_def is used here to preserve the node bug information
frozen_func, graph_def = (
_convert_to_constants.convert_variables_to_constants_v2_as_graph(
self._funcs[0], lower_control_flow=False))
input_tensors = [
tensor for tensor in frozen_func.inputs
if tensor.dtype != _dtypes.resource
]
output_tensors = frozen_func.outputs
# Run a Grappler pass.
graph_def = _run_graph_optimizations(
graph_def,
input_tensors,
output_tensors,
config=self._grappler_config(),
graph=frozen_func.graph)
# Checks dimensions in input tensor.
for tensor in input_tensors:
# Note that shape_list might be empty for scalar shapes.
shape_list = tensor.shape.as_list()
if None in shape_list[1:]:
raise ValueError(
"None is only supported in the 1st dimension. Tensor '{0}' has "
"invalid shape '{1}'.".format(_get_tensor_name(tensor), shape_list))
elif shape_list and shape_list[0] is None:
# Set the batch size to 1 if undefined.
shape = tensor.shape.as_list()
shape[0] = 1
tensor.set_shape(shape)
self._validate_quantization()
self._validate_representative_dataset()
if self._trackable_obj is None:
self._debug_info = _get_debug_info(
_build_debug_info_func(self._funcs[0].graph), graph_def)
else:
self._debug_info = _get_debug_info(
_convert_debug_info_func(self._trackable_obj.graph_debug_info),
graph_def)
converter_kwargs = self._get_base_converter_args()
if not self.experimental_new_converter:
logging.warning(
"Please consider switching to use new converter by setting "
"experimental_new_converter to true. "
"Old converter (TOCO) is deprecated and flow will be switched on "
"by default to use new converter soon.")
else:
logging.info("Using experimental converter: If you encountered a problem "
"please file a bug. You can opt-out "
"by setting experimental_new_converter=False")
# Converts model.
result = _toco_convert_impl(
input_data=graph_def,
input_tensors=input_tensors,
output_tensors=output_tensors,
**converter_kwargs)
if self._is_calibration_quantize():
result = self._calibrate_quantize_model(
result, constants.FLOAT, constants.FLOAT,
self.experimental_new_quantizer)
return result
