TensorFlow SavedModel: SavedObjectGraph

In a TensorFlow 2 SavedModel, a MetaGraphDef’s object_graph_def holds a SavedObjectGraph: the Python objects of the program, flattened to a list, so that the model can be loaded with its structure and not only its graph. The first entry in the list, nodes[0], is the root object. The schema is saved_object_graph.proto, and the explorer decodes every field below.

SavedObjectGraph

Field	Name	Type	Meaning
1	`nodes`	repeated SavedObject	The objects. An object’s position in this list is its ID.
2	`concrete_functions`	map<string, SavedConcreteFunction>	Details of each concrete function, by name.

SavedObject

One object: named edges to others, what kind it is, and how it is saved. Number 2 and the name attributes are reserved.

Field	Name	Type	Meaning
1	`children`	repeated ObjectReference	Named edges to objects. Every kind may have children except `constant` and `captured_tensor`.
15	`dependencies`	repeated ObjectReference	Objects that must load before this one.
3	`slot_variables`	repeated SlotVariableReference	Slot variables this object owns; only for `user_object`.
4	`user_object`	SavedUserObject	Kind (oneof): an object of a class the loader handles by name.
5	`asset`	SavedAsset	Kind: a reference to an asset file.
6	`function`	SavedFunction	Kind: a `tf.function`.
7	`variable`	SavedVariable	Kind: a variable.
8	`bare_concrete_function`	SavedBareConcreteFunction	Kind: a concrete function with argument names.
9	`constant`	SavedConstant	Kind: a constant.
10	`resource`	SavedResource	Kind: a stateful resource.
12	`captured_tensor`	CapturedTensor	Kind: a tensor a concrete function captures.
11	`saveable_objects`	map<string, SaveableObject>	Functions that save and restore this object.
13	`registered_name`	string	The registered class, `package.class_name`, used to find the class when loading.
14	`serialized_user_proto`	Any	Metadata passed to the class’s `_deserialize_from_proto`.
16	`registered_saver`	string	The registered saver’s name. At most one of this and `saveable_objects` is set.

ObjectReference and SlotVariableReference

Both come from trackable_object_graph.proto.

Field	Name	Type	Meaning
1	`node_id`	int32	The object referred to, an index into `nodes`.
2	`local_name`	string	The name of this edge.

SlotVariableReference
Field	Name	Type	Meaning
1	`original_variable_node_id`	int32	The variable this slot belongs to.
2	`slot_name`	string	The slot’s name, such as `m` or `v`.
3	`slot_variable_node_id`	int32	The object holding the slot’s value.

The kinds of object

SavedUserObject: an object of a class no other kind covers
Field	Name	Type	Meaning
1	`identifier`	string	The registered type identifier the loader uses.
2	`version`	VersionDef	The producer’s version.
3	`metadata`	string	Deprecated: used only by Keras.

SavedAsset: evaluates to a tensor holding the absolute file name
Field	Name	Type	Meaning
1	`asset_file_def_index`	int32	Index into `MetaGraphDef.asset_file_def`; only the filename is used.

SavedFunction: a `tf.function`
Field	Name	Type	Meaning
1	`concrete_functions`	repeated string	The concrete functions, by name.
2	`function_spec`	FunctionSpec	The function’s specification.

CapturedTensor
Field	Name	Type	Meaning
1	`name`	string	The captured tensor’s name.
2	`concrete_function`	string	The concrete function whose graph computes it.

SavedConcreteFunction
Field	Name	Type	Meaning
2	`bound_inputs`	repeated int32	The objects the function captures, by ID.
3	`canonicalized_input_signature`	StructuredValue	The input signature used to create it.
4	`output_signature`	StructuredValue	The return structure, with tensors replaced by TensorSpecs.

SavedBareConcreteFunction
Field	Name	Type	Meaning
1	`concrete_function_name`	string	The concrete function.
2	`argument_keywords`	repeated string	Unique argument names, one per tensor.
3	`allowed_positional_arguments`	int64	How many keywords may be given by position.
4	`function_spec`	FunctionSpec	The original function spec, if any.

SavedConstant
Field	Name	Type	Meaning
1	`operation`	string	The `Const` op’s name in the meta graph.

SavedVariable: a variable initialised from the checkpoint
Field	Name	Type	Meaning
1	`dtype`	DataType	The data type.
2	`shape`	TensorShapeProto	The shape.
3	`trainable`	bool	Whether it is trained.
4	`synchronization`	VariableSynchronization	When a distributed variable is synchronised.
5	`aggregation`	VariableAggregation	How it is aggregated across replicas.
6	`name`	string	The variable’s name.
7	`device`	string	Where it is placed.
8	`experimental_distributed_variable_components`	repeated SavedVariable	Components; non-empty means a distributed variable.

SavedResource: an object that holds state
Field	Name	Type	Meaning
1	`device`	string	Where it must be created; empty lets the user choose.

SaveableObject: node IDs of functions that save and restore from tensors
Field	Name	Type	Meaning
2	`save_function`	int32	The save function’s node.
3	`restore_function`	int32	The restore function’s node.

FunctionSpec

The specification of a function wrapped as a TensorFlow Function. Numbers 3 and 4 are reserved.

Field	Name	Type	Meaning
1	`fullargspec`	StructuredValue	Arguments, from Python’s `inspect.getfullargspec`.
2	`is_method`	bool	Whether it is a class method.
5	`input_signature`	StructuredValue	The input signature, if given.
6	`jit_compile`	JitCompile	Whether XLA compiles it: `DEFAULT` 0, `ON` 1, `OFF` 2.

VariableSynchronization and VariableAggregation

From variable.proto.

Value	VariableSynchronization	Meaning
0	`VARIABLE_SYNCHRONIZATION_AUTO`	The distribution strategy decides.
1	`VARIABLE_SYNCHRONIZATION_NONE`	Only one copy; no syncing.
2	`VARIABLE_SYNCHRONIZATION_ON_WRITE`	Copies update on every write.
3	`VARIABLE_SYNCHRONIZATION_ON_READ`	Copies are aggregated on read.

Value	VariableAggregation	Meaning
0	`VARIABLE_AGGREGATION_NONE`	Default: an error if updated across replicas.
1	`VARIABLE_AGGREGATION_SUM`	Sums the replicas’ updates.
2	`VARIABLE_AGGREGATION_MEAN`	Averages them.
3	`VARIABLE_AGGREGATION_ONLY_FIRST_REPLICA`	Every replica runs the update, but only the first applies it.