def _eager_safe_variable_handle(shape, dtype, shared_name, name, graph_mode):
"""Creates a variable handle with information to do shape inference."""
container = ops.get_default_graph()._container # pylint: disable=protected-access
if container is None:
container = ""
handle = resource_variable_ops.var_handle_op(shape=shape, dtype=dtype,
shared_name=shared_name,
name=name,
container=container)
if graph_mode:
return handle
with context.graph_mode(), ops.Graph().as_default() as graph:
h = resource_variable_ops.var_handle_op(shape=shape, dtype=dtype,
shared_name=shared_name,
name=name,
container=container)
# Tensor._handle_data contains information for the shape-inference code to
# know the shape and dtype of the variable pointed to by a handle. Since
# shape inference doesn't run in eager mode we copy this data here for when
# the handle is captured by an eager mode function.
# pylint: disable=protected-access
handle._handle_data = resource_variable_ops.get_resource_handle_data(h)
# pylint: enable=protected-access
# Clean up op->graph->op reference cycles.
ops.dismantle_graph(graph)
return handle
def _eager_safe_variable_handle(shape, dtype, shared_name, name, graph_mode):
"""Creates a variable handle with information to do shape inference."""
container = ops.get_default_graph()._container # pylint: disable=protected-access
if container is None:
container = ""
handle = resource_variable_ops.var_handle_op(shape=shape, dtype=dtype,
shared_name=shared_name,
name=name,
container=container)
if graph_mode:
return handle
with context.graph_mode(), ops.Graph().as_default() as graph:
h = resource_variable_ops.var_handle_op(shape=shape, dtype=dtype,
shared_name=shared_name,
name=name,
container=container)
# Tensor._handle_data contains information for the shape-inference code to
# know the shape and dtype of the variable pointed to by a handle. Since
# shape inference doesn't run in eager mode we copy this data here for when
# the handle is captured by an eager mode function.
# pylint: disable=protected-access
handle._handle_data = resource_variable_ops.get_resource_handle_data(h)
# pylint: enable=protected-access
# Clean up op->graph->op reference cycles.
ops.dismantle_graph(graph)
return handle
def close(self):
"""
Close tensorflow session to free resources.
:return:
"""
self.sess.close()
try:
tf_ops.dismantle_graph(self.graph) # not implemented in older versions of tensorflow
except AttributeError:
pass
def dismantle_func_graph(func_graph):
"""Removes reference cycles in `func_graph` FuncGraph.
Helpful for making sure the garbage collector doesn't need to run when
the FuncGraph goes out of scope, e.g. in tests using defun with
@test_util.run_in_graph_and_eager_modes(assert_no_eager_garbage=True).
Args:
func_graph: A `FuncGraph` object to destroy. `func_graph` is unusable
after this function.
"""
func_graph.clear_captures()
ops.dismantle_graph(func_graph)
def export_meta_graph(obj, filename, signatures=None, options=None):
"""Exports the MetaGraph proto to a file."""
options = options or save_options.SaveOptions()
export_dir = os.path.dirname(filename)
meta_graph_def, exported_graph, _, _ = _build_meta_graph(
obj, export_dir, signatures, options)
file_io.atomic_write_string_to_file(
filename, meta_graph_def.SerializeToString(deterministic=True))
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector. Before this point, we need to keep references to captured
# constants in the saved graph.
ops.dismantle_graph(exported_graph)
def dismantle_func_graph(func_graph):
"""Removes reference cycles in `func_graph` FuncGraph.
Helpful for making sure the garbage collector doesn't need to run when
the FuncGraph goes out of scope, e.g. in tests using defun with
@test_util.run_in_graph_and_eager_modes(assert_no_eager_garbage=True).
Args:
func_graph: A `FuncGraph` object to destroy. `func_graph` is unusable
after this function.
"""
# TODO(b/115366440): Delete this method when a custom OrderedDict is added.
# Clearing captures using clear() leaves some cycles around.
while func_graph.captures:
func_graph.captures.popitem()
memory.dismantle_ordered_dict(func_graph.captures)
ops.dismantle_graph(func_graph)
def dismantle_func_graph(func_graph):
"""Removes reference cycles in `func_graph` FuncGraph.
Helpful for making sure the garbage collector doesn't need to run when
the FuncGraph goes out of scope, e.g. in tests using defun with
@test_util.run_in_graph_and_eager_modes(assert_no_eager_garbage=True).
Args:
func_graph: A `FuncGraph` object to destroy. `func_graph` is unusable
after this function.
"""
# TODO(b/115366440): Delete this method when a custom OrderedDict is added.
# Clearing captures using clear() leaves some cycles around.
while func_graph.captures:
func_graph.captures.popitem()
memory.dismantle_ordered_dict(func_graph.captures)
ops.dismantle_graph(func_graph)
def _make_graph_def(root, signature_functions, object_saver):
"""Generates and exports call ops for `signature_functions`."""
signatures = {}
# List objects from the eager context to make sure Optimizers give us the
# right Graph-dependent variables.
accessible_objects = util.list_objects(root)
exported_graph = ops.Graph()
with exported_graph.as_default():
object_map, resource_map = _map_resources(accessible_objects)
# Saving an object-based checkpoint again gathers variables. We need to do the
# gathering from the eager context so Optimizers save the right set of
# variables, but want any operations associated with the save/restore to be in
# the exported graph (thus the `to_graph` argument).
saver = object_saver.freeze(object_map=object_map, to_graph=exported_graph)
with exported_graph.as_default():
signatures = _generate_signatures(signature_functions, resource_map)
saver_def = saver.to_proto()
graph_def = exported_graph.as_graph_def(add_shapes=True)
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector.
ops.dismantle_graph(exported_graph)
return graph_def, signatures, saver_def
def _make_graph_def(root, signature_functions, object_saver):
"""Generates and exports call ops for `signature_functions`."""
signatures = {}
# List objects from the eager context to make sure Optimizers give us the
# right Graph-dependent variables.
accessible_objects = util.list_objects(root)
exported_graph = ops.Graph()
with exported_graph.as_default():
object_map, resource_map = _map_resources(accessible_objects)
# Saving an object-based checkpoint again gathers variables. We need to do the
# gathering from the eager context so Optimizers save the right set of
# variables, but want any operations associated with the save/restore to be in
# the exported graph (thus the `to_graph` argument).
saver = object_saver.freeze(object_map=object_map, to_graph=exported_graph)
with exported_graph.as_default():
signatures = _generate_signatures(signature_functions, resource_map)
saver_def = saver.to_proto()
graph_def = exported_graph.as_graph_def(add_shapes=True)
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector.
ops.dismantle_graph(exported_graph)
return graph_def, signatures, saver_def
def _fill_meta_graph_def(meta_graph_def, obj, signature_functions,
object_saver):
"""Generates a MetaGraph which calls `signature_functions`.
Args:
meta_graph_def: The MetaGraphDef proto to fill.
obj: The checkpointable object being exported.
signature_functions: A dictionary mapping signature keys to concrete
functions containing signatures to add to the MetaGraph.
object_saver: A CheckpointableSaver to add to the MetaGraph.
Returns:
An _AssetInfo, which contains information to help creating the SavedModel.
"""
signatures = {}
# List objects from the eager context to make sure Optimizers give us the
# right Graph-dependent variables.
accessible_objects = util.list_objects(obj)
resource_initializer_functions = _trace_resource_initializers(
accessible_objects)
exported_graph = ops.Graph()
resource_initializer_ops = []
with exported_graph.as_default():
object_map, resource_map, asset_info = _map_resources(
accessible_objects)
for resource_initializer_function in resource_initializer_functions:
asset_dependencies = []
for capture in resource_initializer_function.graph.external_captures:
asset_initializer = asset_info.asset_initializers_by_resource.get(
capture, None)
if asset_initializer is not None:
asset_dependencies.append(asset_initializer)
with ops.control_dependencies(asset_dependencies):
resource_initializer_ops.append(
_call_function_with_mapped_captures(
resource_initializer_function, [], resource_map))
with ops.control_dependencies(resource_initializer_ops):
init_op = control_flow_ops.no_op()
# Add the same op to the main_op collection and to the init_op
# signature. The collection is for compatibility with older loader APIs;
# only one will be executed.
meta_graph_def.collection_def[
constants.MAIN_OP_KEY].node_list.value.append(init_op.name)
meta_graph_def.signature_def[constants.INIT_OP_SIGNATURE_KEY].CopyFrom(
signature_def_utils.op_signature_def(
init_op, constants.INIT_OP_SIGNATURE_KEY))
# Saving an object-based checkpoint again gathers variables. We need to do the
# gathering from the eager context so Optimizers save the right set of
# variables, but want any operations associated with the save/restore to be in
# the exported graph (thus the `to_graph` argument).
saver = object_saver.freeze(object_map=object_map, to_graph=exported_graph)
# We must resolve the concrete function to add to MetaGraph while in eager
# mode.
concrete_functions = []
for accessible_object in accessible_objects:
for function in function_serialization.list_all_polymorphic_functions(
accessible_object).values():
concrete_functions.extend(
function_serialization.list_all_concrete_functions(function))
with exported_graph.as_default():
signatures = _generate_signatures(signature_functions, resource_map)
for concrete_function in concrete_functions:
concrete_function.add_to_graph()
saver_def = saver.to_proto()
meta_graph_def.saver_def.CopyFrom(saver_def)
graph_def = exported_graph.as_graph_def(add_shapes=True)
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector.
ops.dismantle_graph(exported_graph)
meta_graph_def.graph_def.CopyFrom(graph_def)
meta_graph_def.meta_info_def.tags.append(tag_constants.SERVING)
meta_graph_def.asset_file_def.extend(asset_info.asset_defs)
for signature_key, signature in signatures.items():
meta_graph_def.signature_def[signature_key].CopyFrom(signature)
meta_graph.strip_graph_default_valued_attrs(meta_graph_def)
return asset_info
def save(obj, export_dir, signatures=None, options=None):
# pylint: disable=line-too-long
"""Exports the Trackable object `obj` to [SavedModel format](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md).
Example usage:
```python
class Adder(tf.Module):
@tf.function(input_signature=[tf.TensorSpec(shape=None, dtype=tf.float32)])
def add(self, x):
return x + x + 1.
to_export = Adder()
tf.saved_model.save(to_export, '/tmp/adder')
```
The resulting SavedModel is then servable with an input named "x", its value
having any shape and dtype float32.
The optional `signatures` argument controls which methods in `obj` will be
available to programs which consume `SavedModel`s, for example serving
APIs. Python functions may be decorated with
`@tf.function(input_signature=...)` and passed as signatures directly, or
lazily with a call to `get_concrete_function` on the method decorated with
`@tf.function`.
If the `signatures` argument is omitted, `obj` will be searched for
`@tf.function`-decorated methods. If exactly one `@tf.function` is found, that
method will be used as the default signature for the SavedModel. This behavior
is expected to change in the future, when a corresponding
`tf.saved_model.load` symbol is added. At that point signatures will be
completely optional, and any `@tf.function` attached to `obj` or its
dependencies will be exported for use with `load`.
When invoking a signature in an exported SavedModel, `Tensor` arguments are
identified by name. These names will come from the Python function's argument
names by default. They may be overridden by specifying a `name=...` argument
in the corresponding `tf.TensorSpec` object. Explicit naming is required if
multiple `Tensor`s are passed through a single argument to the Python
function.
The outputs of functions used as `signatures` must either be flat lists, in
which case outputs will be numbered, or a dictionary mapping string keys to
`Tensor`, in which case the keys will be used to name outputs.
Signatures are available in objects returned by `tf.saved_model.load` as a
`.signatures` attribute. This is a reserved attribute: `tf.saved_model.save`
on an object with a custom `.signatures` attribute will raise an exception.
Since `tf.keras.Model` objects are also Trackable, this function can be
used to export Keras models. For example, exporting with a signature
specified:
```python
class Model(tf.keras.Model):
@tf.function(input_signature=[tf.TensorSpec(shape=[None], dtype=tf.string)])
def serve(self, serialized):
...
m = Model()
tf.saved_model.save(m, '/tmp/saved_model/')
```
Exporting from a function without a fixed signature:
```python
class Model(tf.keras.Model):
@tf.function
def call(self, x):
...
m = Model()
tf.saved_model.save(
m, '/tmp/saved_model/',
signatures=m.call.get_concrete_function(
tf.TensorSpec(shape=[None, 3], dtype=tf.float32, name="inp")))
```
`tf.keras.Model` instances constructed from inputs and outputs already have a
signature and so do not require a `@tf.function` decorator or a `signatures`
argument. If neither are specified, the model's forward pass is exported.
```python
x = input_layer.Input((4,), name="x")
y = core.Dense(5, name="out")(x)
model = training.Model(x, y)
tf.saved_model.save(model, '/tmp/saved_model/')
# The exported SavedModel takes "x" with shape [None, 4] and returns "out"
# with shape [None, 5]
```
Variables must be tracked by assigning them to an attribute of a tracked
object or to an attribute of `obj` directly. TensorFlow objects (e.g. layers
from `tf.keras.layers`, optimizers from `tf.train`) track their variables
automatically. This is the same tracking scheme that `tf.train.Checkpoint`
uses, and an exported `Checkpoint` object may be restored as a training
checkpoint by pointing `tf.train.Checkpoint.restore` to the SavedModel's
"variables/" subdirectory. Currently variables are the only stateful objects
supported by `tf.saved_model.save`, but others (e.g. tables) will be supported
in the future.
`tf.function` does not hard-code device annotations from outside the function
body, instead using the calling context's device. This means for example that
exporting a model which runs on a GPU and serving it on a CPU will generally
work, with some exceptions. `tf.device` annotations inside the body of the
function will be hard-coded in the exported model; this type of annotation is
discouraged. Device-specific operations, e.g. with "cuDNN" in the name or with
device-specific layouts, may cause issues. Currently a `DistributionStrategy`
is another exception: active distribution strategies will cause device
placements to be hard-coded in a function. Exporting a single-device
computation and importing under a `DistributionStrategy` is not currently
supported, but may be in the future.
SavedModels exported with `tf.saved_model.save` [strip default-valued
attributes](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md#stripping-default-valued-attributes)
automatically, which removes one source of incompatibilities when the consumer
of a SavedModel is running an older TensorFlow version than the
producer. There are however other sources of incompatibilities which are not
handled automatically, such as when the exported model contains operations
which the consumer does not have definitions for.
Args:
obj: A trackable object to export.
export_dir: A directory in which to write the SavedModel.
signatures: Optional, either a `tf.function` with an input signature
specified or the result of `f.get_concrete_function` on a
`@tf.function`-decorated function `f`, in which case `f` will be used to
generate a signature for the SavedModel under the default serving
signature key. `signatures` may also be a dictionary, in which case it
maps from signature keys to either `tf.function` instances with input
signatures or concrete functions. The keys of such a dictionary may be
arbitrary strings, but will typically be from the
`tf.saved_model.signature_constants` module.
options: Optional, `tf.saved_model.SaveOptions` object that specifies
options for saving.
Raises:
ValueError: If `obj` is not trackable.
@compatibility(eager)
Not well supported when graph building. From TensorFlow 1.x,
`tf.compat.v1.enable_eager_execution()` should run first. Calling
tf.saved_model.save in a loop when graph building from TensorFlow 1.x will
add new save operations to the default graph each iteration.
May not be called from within a function body.
@end_compatibility
"""
if ops.inside_function():
raise AssertionError(
"tf.saved_model.save is not supported inside a traced "
"@tf.function. Move the call to the outer eagerly-executed "
"context.")
# pylint: enable=line-too-long
if not isinstance(obj, base.Trackable):
raise ValueError(
"Expected a Trackable object for export, got {}.".format(obj))
options = options or save_options.SaveOptions()
checkpoint_graph_view = _AugmentedGraphView(obj)
if signatures is None:
signatures = signature_serialization.find_function_to_export(
checkpoint_graph_view)
signatures = signature_serialization.canonicalize_signatures(signatures)
signature_serialization.validate_saveable_view(checkpoint_graph_view)
signature_map = signature_serialization.create_signature_map(signatures)
checkpoint_graph_view.add_object(
parent_node=checkpoint_graph_view.root,
name_in_parent=signature_serialization.SIGNATURE_ATTRIBUTE_NAME,
subgraph_root=signature_map)
# Use _SaveableView to provide a frozen listing of properties and functions.
# Note we run this twice since, while constructing the view the first time
# there can be side effects of creating variables.
_ = _SaveableView(checkpoint_graph_view)
saveable_view = _SaveableView(checkpoint_graph_view)
# TODO(allenl): Factor out some subset of SavedModelBuilder which is 2.x
# compatible (no sessions) and share it with this export API rather than
# making a SavedModel proto and writing it directly.
saved_model = saved_model_pb2.SavedModel()
meta_graph_def = saved_model.meta_graphs.add()
object_saver = util.TrackableSaver(checkpoint_graph_view)
asset_info, exported_graph = _fill_meta_graph_def(
meta_graph_def, saveable_view, signatures, options.namespace_whitelist)
saved_model.saved_model_schema_version = (
constants.SAVED_MODEL_SCHEMA_VERSION)
# So far we've just been generating protocol buffers with no I/O. Now we write
# the checkpoint, copy assets into the assets directory, and write out the
# SavedModel proto itself.
utils_impl.get_or_create_variables_dir(export_dir)
object_saver.save(utils_impl.get_variables_path(export_dir))
builder_impl.copy_assets_to_destination_dir(asset_info.asset_filename_map,
export_dir)
path = os.path.join(
compat.as_str(export_dir),
compat.as_str(constants.SAVED_MODEL_FILENAME_PB))
object_graph_proto = _serialize_object_graph(
saveable_view, asset_info.asset_index)
meta_graph_def.object_graph_def.CopyFrom(object_graph_proto)
# Save debug info, if requested.
if options.save_debug_info:
graph_debug_info = _export_debug_info(exported_graph)
file_io.atomic_write_string_to_file(
os.path.join(
utils_impl.get_or_create_debug_dir(export_dir),
constants.DEBUG_INFO_FILENAME_PB),
graph_debug_info.SerializeToString(deterministic=True))
# Note that this needs to be the last file operation when saving the
# SavedModel. Users rely on checking saved_model_dir/saved_model.pb as an
# indication that the SavedModel is completely written.
file_io.atomic_write_string_to_file(
path, saved_model.SerializeToString(deterministic=True))
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector. Before this point we need to keep references to captured
# constants in the saved graph.
ops.dismantle_graph(exported_graph)
def _fill_meta_graph_def(meta_graph_def, saveable_view, signature_functions,
object_saver):
"""Generates a MetaGraph which calls `signature_functions`.
Args:
meta_graph_def: The MetaGraphDef proto to fill.
saveable_view: The _SaveableView being exported.
signature_functions: A dictionary mapping signature keys to concrete
functions containing signatures to add to the MetaGraph.
object_saver: A CheckpointableSaver to add to the MetaGraph.
Returns:
An _AssetInfo, which contains information to help creating the SavedModel.
"""
# List objects from the eager context to make sure Optimizers give us the
# right Graph-dependent variables.
accessible_objects = saveable_view.nodes
resource_initializer_functions = _trace_resource_initializers(
accessible_objects)
exported_graph = ops.Graph()
resource_initializer_ops = []
with exported_graph.as_default():
object_map, resource_map, asset_info = _map_resources(accessible_objects)
for resource_initializer_function in resource_initializer_functions:
asset_dependencies = []
for capture in resource_initializer_function.graph.external_captures:
asset_initializer = asset_info.asset_initializers_by_resource.get(
capture, None)
if asset_initializer is not None:
asset_dependencies.append(asset_initializer)
with ops.control_dependencies(asset_dependencies):
resource_initializer_ops.append(
_call_function_with_mapped_captures(
resource_initializer_function, [], resource_map))
with ops.control_dependencies(resource_initializer_ops):
init_op = control_flow_ops.no_op()
# Add the same op to the main_op collection and to the init_op
# signature. The collection is for compatibility with older loader APIs;
# only one will be executed.
meta_graph_def.collection_def[constants.MAIN_OP_KEY].node_list.value.append(
init_op.name)
meta_graph_def.signature_def[constants.INIT_OP_SIGNATURE_KEY].CopyFrom(
signature_def_utils.op_signature_def(
init_op, constants.INIT_OP_SIGNATURE_KEY))
# Saving an object-based checkpoint again gathers variables. We need to do the
# gathering from the eager context so Optimizers save the right set of
# variables, but want any operations associated with the save/restore to be in
# the exported graph (thus the `to_graph` argument).
saver = object_saver.freeze(object_map=object_map, to_graph=exported_graph)
# We must instantiate and list all concrete functions of `Function`s while in
# eager mode so they end up added to the graph and can later be used by the
# object based saved model.
concrete_functions = []
for obj in accessible_objects:
for function in saveable_view.functions[obj].values():
if isinstance(function, defun.ConcreteFunction):
concrete_functions.append(function)
else:
concrete_functions.extend(
function._list_all_concrete_functions_for_serialization()) # pylint: disable=protected-access
with exported_graph.as_default():
signatures = _generate_signatures(signature_functions, resource_map)
for concrete_function in concrete_functions:
concrete_function.add_to_graph()
saver_def = saver.to_proto()
meta_graph_def.saver_def.CopyFrom(saver_def)
graph_def = exported_graph.as_graph_def(add_shapes=True)
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector.
ops.dismantle_graph(exported_graph)
meta_graph_def.graph_def.CopyFrom(graph_def)
meta_graph_def.meta_info_def.tags.append(tag_constants.SERVING)
meta_graph_def.asset_file_def.extend(asset_info.asset_defs)
for signature_key, signature in signatures.items():
meta_graph_def.signature_def[signature_key].CopyFrom(signature)
meta_graph.strip_graph_default_valued_attrs(meta_graph_def)
return asset_info
def save(obj, export_dir, signatures=None):
# pylint: disable=line-too-long
"""Exports the Trackable object `obj` to [SavedModel format](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md).
Example usage:
```python
class Adder(tf.train.Checkpoint):
@tf.function(input_signature=[tf.TensorSpec(shape=None, dtype=tf.float32)])
def add(self, x):
return x + x + 1.
to_export = Adder()
tf.saved_model.save(to_export, '/tmp/adder')
```
The resulting SavedModel is then servable with an input named "x", its value
having any shape and dtype float32.
The optional `signatures` argument controls which methods in `obj` will be
available to programs which consume `SavedModel`s, for example serving
APIs. Python functions may be decorated with
`@tf.function(input_signature=...)` and passed as signatures directly, or
lazily with a call to `get_concrete_function` on the method decorated with
`@tf.function`.
If the `signatures` argument is omitted, `obj` will be searched for
`@tf.function`-decorated methods. If exactly one `@tf.function` is found, that
method will be used as the default signature for the SavedModel. This behavior
is expected to change in the future, when a corresponding
`tf.saved_model.load` symbol is added. At that point signatures will be
completely optional, and any `@tf.function` attached to `obj` or its
dependencies will be exported for use with `load`.
When invoking a signature in an exported SavedModel, `Tensor` arguments are
identified by name. These names will come from the Python function's argument
names by default. They may be overridden by specifying a `name=...` argument
in the corresponding `tf.TensorSpec` object. Explicit naming is required if
multiple `Tensor`s are passed through a single argument to the Python
function.
The outputs of functions used as `signatures` must either be flat lists, in
which case outputs will be numbered, or a dictionary mapping string keys to
`Tensor`, in which case the keys will be used to name outputs.
Signatures are available in objects returned by `tf.saved_model.load` as a
`.signatures` attribute. This is a reserved attribute: `tf.saved_model.save`
on an object with a custom `.signatures` attribute will raise an exception.
Since `tf.keras.Model` objects are also Trackable, this function can be
used to export Keras models. For example, exporting with a signature
specified:
```python
class Model(tf.keras.Model):
@tf.function(input_signature=[tf.TensorSpec(shape=[None], dtype=tf.string)])
def serve(self, serialized):
...
m = Model()
tf.saved_model.save(m, '/tmp/saved_model/')
```
Exporting from a function without a fixed signature:
```python
class Model(tf.keras.Model):
@tf.function
def call(self, x):
...
m = Model()
tf.saved_model.save(
m, '/tmp/saved_model/',
signatures=m.call.get_concrete_function(
tf.TensorSpec(shape=[None, 3], dtype=tf.float32, name="inp")))
```
`tf.keras.Model` instances constructed from inputs and outputs already have a
signature and so do not require a `@tf.function` decorator or a `signatures`
argument. If neither are specified, the model's forward pass is exported.
```python
x = input_layer.Input((4,), name="x")
y = core.Dense(5, name="out")(x)
model = training.Model(x, y)
tf.saved_model.save(model, '/tmp/saved_model/')
# The exported SavedModel takes "x" with shape [None, 4] and returns "out"
# with shape [None, 5]
```
Variables must be tracked by assigning them to an attribute of a tracked
object or to an attribute of `obj` directly. TensorFlow objects (e.g. layers
from `tf.keras.layers`, optimizers from `tf.train`) track their variables
automatically. This is the same tracking scheme that `tf.train.Checkpoint`
uses, and an exported `Checkpoint` object may be restored as a training
checkpoint by pointing `tf.train.Checkpoint.restore` to the SavedModel's
"variables/" subdirectory. Currently variables are the only stateful objects
supported by `tf.saved_model.save`, but others (e.g. tables) will be supported
in the future.
`tf.function` does not hard-code device annotations from outside the function
body, instead using the calling context's device. This means for example that
exporting a model which runs on a GPU and serving it on a CPU will generally
work, with some exceptions. `tf.device` annotations inside the body of the
function will be hard-coded in the exported model; this type of annotation is
discouraged. Device-specific operations, e.g. with "cuDNN" in the name or with
device-specific layouts, may cause issues. Currently a `DistributionStrategy`
is another exception: active distribution strategies will cause device
placements to be hard-coded in a function. Exporting a single-device
computation and importing under a `DistributionStrategy` is not currently
supported, but may be in the future.
SavedModels exported with `tf.saved_model.save` [strip default-valued
attributes](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md#stripping-default-valued-attributes)
automatically, which removes one source of incompatibilities when the consumer
of a SavedModel is running an older TensorFlow version than the
producer. There are however other sources of incompatibilities which are not
handled automatically, such as when the exported model contains operations
which the consumer does not have definitions for.
The current implementation of `tf.saved_model.save` targets serving use-cases,
but omits information which will be necessary for the planned future
implementation of `tf.saved_model.load`. Exported models using the current
`save` implementation, and other existing SavedModels, will not be compatible
with `tf.saved_model.load` when it is implemented. Further, `save` will in the
future attempt to export `@tf.function`-decorated methods which it does not
currently inspect, so some objects which are exportable today will raise
exceptions on export in the future (e.g. due to complex/non-serializable
default arguments). Such backwards-incompatible API changes are expected only
prior to the TensorFlow 2.0 release.
Args:
obj: A trackable object to export.
export_dir: A directory in which to write the SavedModel.
signatures: Optional, either a `tf.function` with an input signature
specified or the result of `f.get_concrete_function` on a
`@tf.function`-decorated function `f`, in which case `f` will be used to
generate a signature for the SavedModel under the default serving
signature key. `signatures` may also be a dictionary, in which case it
maps from signature keys to either `tf.function` instances with input
signatures or concrete functions. The keys of such a dictionary may be
arbitrary strings, but will typically be from the
`tf.saved_model.signature_constants` module.
Raises:
ValueError: If `obj` is not trackable.
@compatibility(eager)
Not well supported when graph building. From TensorFlow 1.x,
`tf.enable_eager_execution()` should run first. Calling tf.saved_model.save in
a loop when graph building from TensorFlow 1.x will add new save operations to
the default graph each iteration.
May not be called from within a function body.
@end_compatibility
"""
if ops.inside_function():
raise AssertionError(
"tf.saved_model.save is not supported inside a traced "
"@tf.function. Move the call to the outer eagerly-executed "
"context.")
# pylint: enable=line-too-long
if not isinstance(obj, base.Trackable):
raise ValueError(
"Expected a Trackable object for export, got {}.".format(obj))
checkpoint_graph_view = _AugmentedGraphView(obj)
if signatures is None:
signatures = signature_serialization.find_function_to_export(
checkpoint_graph_view)
signatures = signature_serialization.canonicalize_signatures(signatures)
signature_serialization.validate_saveable_view(checkpoint_graph_view)
signature_map = signature_serialization.create_signature_map(signatures)
checkpoint_graph_view.add_object(
parent_node=checkpoint_graph_view.root,
name_in_parent=signature_serialization.SIGNATURE_ATTRIBUTE_NAME,
subgraph_root=signature_map)
# Use _SaveableView to provide a frozen listing of properties and functions.
# Note we run this twice since, while constructing the view the first time
# there can be side effects of creating variables.
_ = _SaveableView(checkpoint_graph_view)
saveable_view = _SaveableView(checkpoint_graph_view)
# TODO(allenl): Factor out some subset of SavedModelBuilder which is 2.x
# compatible (no sessions) and share it with this export API rather than
# making a SavedModel proto and writing it directly.
saved_model = saved_model_pb2.SavedModel()
meta_graph_def = saved_model.meta_graphs.add()
object_saver = util.TrackableSaver(checkpoint_graph_view)
asset_info, exported_graph = _fill_meta_graph_def(
meta_graph_def, saveable_view, signatures)
saved_model.saved_model_schema_version = (
constants.SAVED_MODEL_SCHEMA_VERSION)
# So far we've just been generating protocol buffers with no I/O. Now we write
# the checkpoint, copy assets into the assets directory, and write out the
# SavedModel proto itself.
utils_impl.get_or_create_variables_dir(export_dir)
object_saver.save(utils_impl.get_variables_path(export_dir))
builder_impl.copy_assets_to_destination_dir(asset_info.asset_filename_map,
export_dir)
path = os.path.join(
compat.as_bytes(export_dir),
compat.as_bytes(constants.SAVED_MODEL_FILENAME_PB))
object_graph_proto = _serialize_object_graph(
saveable_view, asset_info.asset_index)
meta_graph_def.object_graph_def.CopyFrom(object_graph_proto)
file_io.write_string_to_file(path, saved_model.SerializeToString())
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector. Before this point we need to keep references to captured
# constants in the saved graph.
ops.dismantle_graph(exported_graph)
def save(obj, export_dir, signatures=None, options=None):
# pylint: disable=line-too-long
"""Exports the Trackable object `obj` to [SavedModel format](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md).
Example usage:
```python
class Adder(tf.Module):
@tf.function(input_signature=[tf.TensorSpec(shape=None, dtype=tf.float32)])
def add(self, x):
return x + x + 1.
to_export = Adder()
tf.saved_model.save(to_export, '/tmp/adder')
```
The resulting SavedModel is then servable with an input named "x", its value
having any shape and dtype float32.
The optional `signatures` argument controls which methods in `obj` will be
available to programs which consume `SavedModel`s, for example, serving
APIs. Python functions may be decorated with
`@tf.function(input_signature=...)` and passed as signatures directly, or
lazily with a call to `get_concrete_function` on the method decorated with
`@tf.function`.
If the `signatures` argument is omitted, `obj` will be searched for
`@tf.function`-decorated methods. If exactly one `@tf.function` is found, that
method will be used as the default signature for the SavedModel. This behavior
is expected to change in the future, when a corresponding
`tf.saved_model.load` symbol is added. At that point signatures will be
completely optional, and any `@tf.function` attached to `obj` or its
dependencies will be exported for use with `load`.
When invoking a signature in an exported SavedModel, `Tensor` arguments are
identified by name. These names will come from the Python function's argument
names by default. They may be overridden by specifying a `name=...` argument
in the corresponding `tf.TensorSpec` object. Explicit naming is required if
multiple `Tensor`s are passed through a single argument to the Python
function.
The outputs of functions used as `signatures` must either be flat lists, in
which case outputs will be numbered, or a dictionary mapping string keys to
`Tensor`, in which case the keys will be used to name outputs.
Signatures are available in objects returned by `tf.saved_model.load` as a
`.signatures` attribute. This is a reserved attribute: `tf.saved_model.save`
on an object with a custom `.signatures` attribute will raise an exception.
Since `tf.keras.Model` objects are also Trackable, this function can be
used to export Keras models. For example, exporting with a signature
specified:
```python
class Model(tf.keras.Model):
@tf.function(input_signature=[tf.TensorSpec(shape=[None], dtype=tf.string)])
def serve(self, serialized):
...
m = Model()
tf.saved_model.save(m, '/tmp/saved_model/')
```
Exporting from a function without a fixed signature:
```python
class Model(tf.keras.Model):
@tf.function
def call(self, x):
...
m = Model()
tf.saved_model.save(
m, '/tmp/saved_model/',
signatures=m.call.get_concrete_function(
tf.TensorSpec(shape=[None, 3], dtype=tf.float32, name="inp")))
```
`tf.keras.Model` instances constructed from inputs and outputs already have a
signature and so do not require a `@tf.function` decorator or a `signatures`
argument. If neither are specified, the model's forward pass is exported.
```python
x = input_layer.Input((4,), name="x")
y = core.Dense(5, name="out")(x)
model = training.Model(x, y)
tf.saved_model.save(model, '/tmp/saved_model/')
# The exported SavedModel takes "x" with shape [None, 4] and returns "out"
# with shape [None, 5]
```
Variables must be tracked by assigning them to an attribute of a tracked
object or to an attribute of `obj` directly. TensorFlow objects (e.g. layers
from `tf.keras.layers`, optimizers from `tf.train`) track their variables
automatically. This is the same tracking scheme that `tf.train.Checkpoint`
uses, and an exported `Checkpoint` object may be restored as a training
checkpoint by pointing `tf.train.Checkpoint.restore` to the SavedModel's
"variables/" subdirectory. Currently, variables are the only stateful objects
supported by `tf.saved_model.save`, but others (e.g. tables) will be supported
in the future.
`tf.function` does not hard-code device annotations from outside the function
body, instead of using the calling context's device. This means for example
that exporting a model that runs on a GPU and serving it on a CPU will
generally work, with some exceptions. `tf.device` annotations inside the body
of the function will be hard-coded in the exported model; this type of
annotation is discouraged. Device-specific operations, e.g. with "cuDNN" in
the name or with device-specific layouts, may cause issues. Currently a
`DistributionStrategy` is another exception: active distribution strategies
will cause device placements to be hard-coded in a function. Exporting a
single-device computation and importing under a `DistributionStrategy` is
not currently supported, but may be in the future.
SavedModels exported with `tf.saved_model.save` [strip default-valued
attributes](https://github.com/tensorflow/tensorflow/blob/master/tensorflow/python/saved_model/README.md#stripping-default-valued-attributes)
automatically, which removes one source of incompatibilities when the consumer
of a SavedModel is running an older TensorFlow version than the
producer. There are however other sources of incompatibilities which are not
handled automatically, such as when the exported model contains operations
which the consumer does not have definitions for.
A single tf.function can generate many ConcreteFunctions. If a downstream tool
wants to refer to all concrete functions generated by a single tf.function you
can use the `function_aliases` argument to store a map from the alias name to
all concrete function names.
E.g.
```python
class MyModel:
@tf.function
def func():
...
@tf.function
def serve():
...
func()
model = MyModel()
signatures = {
'serving_default': model.serve.get_concrete_function(),
}
options = tf.saved_model.SaveOptions(function_aliases={
'my_func': func,
})
tf.saved_model.save(model, export_dir, signatures, options)
```
Args:
obj: A trackable object to export.
export_dir: A directory in which to write the SavedModel.
signatures: Optional, either a `tf.function` with an input signature
specified or the result of `f.get_concrete_function` on a
`@tf.function`-decorated function `f`, in which case `f` will be used to
generate a signature for the SavedModel under the default serving
signature key. `signatures` may also be a dictionary, in which case it
maps from signature keys to either `tf.function` instances with input
signatures or concrete functions. The keys of such a dictionary may be
arbitrary strings, but will typically be from the
`tf.saved_model.signature_constants` module.
options: Optional, `tf.saved_model.SaveOptions` object that specifies
options for saving.
Raises:
ValueError: If `obj` is not trackable.
@compatibility(eager)
Not well supported when graph building. From TensorFlow 1.x,
`tf.compat.v1.enable_eager_execution()` should run first. Calling
tf.saved_model.save in a loop when graph building from TensorFlow 1.x will
add new save operations to the default graph each iteration.
May not be called from within a function body.
@end_compatibility
"""
options = options or save_options.SaveOptions()
# TODO(allenl): Factor out some subset of SavedModelBuilder which is 2.x
# compatible (no sessions) and share it with this export API rather than
# making a SavedModel proto and writing it directly.
saved_model = saved_model_pb2.SavedModel()
meta_graph_def = saved_model.meta_graphs.add()
_, exported_graph, object_saver, asset_info = _build_meta_graph(
obj, export_dir, signatures, options, meta_graph_def)
saved_model.saved_model_schema_version = constants.SAVED_MODEL_SCHEMA_VERSION
# Write the checkpoint, copy assets into the assets directory, and write out
# the SavedModel proto itself.
utils_impl.get_or_create_variables_dir(export_dir)
ckpt_options = checkpoint_options.CheckpointOptions(
experimental_io_device=options.experimental_io_device)
object_saver.save(utils_impl.get_variables_path(export_dir),
options=ckpt_options)
builder_impl.copy_assets_to_destination_dir(asset_info.asset_filename_map,
export_dir)
# Note that this needs to be the last file operation when saving the
# SavedModel. Users rely on checking saved_model_dir/saved_model.pb as an
# indication that the SavedModel is completely written.
if context.executing_eagerly():
try:
context.async_wait() # Ensure save operations have completed.
except errors.NotFoundError as err:
raise FileNotFoundError(
str(err) +
"\n If trying to save on a different device from the "
"computational device, consider using setting the "
"`experimental_io_device` option on tf.saved_model.SaveOptions "
"to the io_device such as '/job:localhost'.")
path = os.path.join(compat.as_str(export_dir),
compat.as_str(constants.SAVED_MODEL_FILENAME_PB))
file_io.atomic_write_string_to_file(
path, saved_model.SerializeToString(deterministic=True))
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector. Before this point, we need to keep references to captured
# constants in the saved graph.
ops.dismantle_graph(exported_graph)
def _fill_meta_graph_def(meta_graph_def, obj, signature_functions,
object_saver):
"""Generates a MetaGraph which calls `signature_functions`.
Args:
meta_graph_def: The MetaGraphDef proto to fill.
obj: The checkpointable object being exported.
signature_functions: A dictionary mapping signature keys to concrete
functions containing signatures to add to the MetaGraph.
object_saver: A CheckpointableSaver to add to the MetaGraph.
Returns:
asset_filename_map, a dictionary mapping from asset base names to
user-specified full asset paths, which should be copied to the SavedModel's
assets/ directory.
"""
signatures = {}
# List objects from the eager context to make sure Optimizers give us the
# right Graph-dependent variables.
accessible_objects = util.list_objects(obj)
resource_initializer_functions = _trace_resource_initializers(
accessible_objects)
exported_graph = ops.Graph()
resource_initializer_ops = []
with exported_graph.as_default():
object_map, resource_map, asset_info = _map_resources(accessible_objects)
for resource_initializer_function in resource_initializer_functions:
asset_dependencies = []
for capture in resource_initializer_function.graph.external_captures:
asset_initializer = asset_info.asset_initializers_by_resource.get(
capture, None)
if asset_initializer is not None:
asset_dependencies.append(asset_initializer)
with ops.control_dependencies(asset_dependencies):
resource_initializer_ops.append(
_call_function_with_mapped_captures(
resource_initializer_function, [], resource_map))
with ops.control_dependencies(resource_initializer_ops):
init_op = control_flow_ops.no_op()
# Add the same op to the main_op collection and to the init_op
# signature. The collection is for compatibility with older loader APIs;
# only one will be executed.
meta_graph_def.collection_def[constants.MAIN_OP_KEY].node_list.value.append(
init_op.name)
meta_graph_def.signature_def[constants.INIT_OP_SIGNATURE_KEY].CopyFrom(
signature_def_utils.op_signature_def(
init_op, constants.INIT_OP_SIGNATURE_KEY))
# Saving an object-based checkpoint again gathers variables. We need to do the
# gathering from the eager context so Optimizers save the right set of
# variables, but want any operations associated with the save/restore to be in
# the exported graph (thus the `to_graph` argument).
saver = object_saver.freeze(object_map=object_map, to_graph=exported_graph)
with exported_graph.as_default():
signatures = _generate_signatures(signature_functions, resource_map)
saver_def = saver.to_proto()
meta_graph_def.saver_def.CopyFrom(saver_def)
graph_def = exported_graph.as_graph_def(add_shapes=True)
# Clean reference cycles so repeated export()s don't make work for the garbage
# collector.
ops.dismantle_graph(exported_graph)
meta_graph_def.graph_def.CopyFrom(graph_def)
meta_graph_def.meta_info_def.tags.append(tag_constants.SERVING)
meta_graph_def.asset_file_def.extend(asset_info.asset_defs)
for signature_key, signature in signatures.items():
meta_graph_def.signature_def[signature_key].CopyFrom(signature)
meta_graph.strip_graph_default_valued_attrs(meta_graph_def)
return asset_info.asset_filename_map
