TensorFlow MetaGraphDef and SignatureDef, field by field

A SavedModel holds one or more MetaGraphDefs, each a graph together with what is needed to run it: tags that pick it out, the signatures a caller uses, and how to restore its variables. These messages are from TensorFlow’s meta_graph.proto and saver.proto, with Any from Protocol Buffers’ own any.proto (all proto3). Maps are written as lists of entry messages, as the Protocol Buffers reference describes.

MetaGraphDef

Field	Name	Type	Wire type	Meaning
1	`meta_info_def`	MetaInfoDef	LEN	Tags, versions and the ops the graph uses.
2	`graph_def`	GraphDef	LEN	The graph, with its function library.
3	`saver_def`	SaverDef	LEN	How to save and restore the variables.
4	`collection_def`	map<string, CollectionDef>	LEN, one per entry	Named collections, a TensorFlow 1 feature.
5	`signature_def`	map<string, SignatureDef>	LEN, one per entry	Named signatures, such as `serving_default`.
6	`asset_file_def`	repeated AssetFileDef	LEN, one per asset	Asset files the graph uses.
7	`object_graph_def`	SavedObjectGraph	LEN	The saved Python objects and their functions, in TensorFlow 2. The explorer does not include this message and shows it by number.

MetaGraphDef.MetaInfoDef

Metadata: the meta graph’s tags and the TensorFlow version that wrote it. The explorer shows the tags and the version of each meta graph.

Field	Name	Type	Wire type	Meaning
1	`meta_graph_version`	string	LEN	A version the user chose.
2	`stripped_op_list`	OpList	LEN	The definitions of the ops the graph uses.
3	`any_info`	Any	LEN	Extra information of any type.
4	`tags`	repeated string	LEN, one per tag	Tags that pick this meta graph out, such as `serve` or `train`.
5	`tensorflow_version`	string	LEN	The TensorFlow version that wrote it.
6	`tensorflow_git_version`	string	LEN	Its git version.
7	`stripped_default_attrs`	bool	VARINT	Whether attributes at their default values were left out of the nodes.
8	`function_aliases`	map<string, string>	LEN, one per entry	Aliases for functions.

MetaGraphDef.MetaInfoDef.FunctionAliasesEntry

Field	Name	Type	Wire type	Meaning
1	`key`	string	LEN	The key.
2	`value`	string	LEN	The value.

MetaGraphDef.CollectionDefEntry

Field	Name	Type	Wire type	Meaning
1	`key`	string	LEN	The collection’s name.
2	`value`	CollectionDef	LEN	Its contents.

MetaGraphDef.SignatureDefEntry

Field	Name	Type	Wire type	Meaning
1	`key`	string	LEN	The signature’s name, such as `serving_default`.
2	`value`	SignatureDef	LEN	Its definition.

CollectionDef

A collection: a list of one kind. Fields 1 to 5 form a oneof.

Field	Name	Type	Wire type	Meaning
1	`node_list`	NodeList	LEN	Node names.
2	`bytes_list`	BytesList	LEN	Byte strings.
3	`int64_list`	Int64List	LEN	Integers.
4	`float_list`	FloatList	LEN	Floats.
5	`any_list`	AnyList	LEN	Any messages.

CollectionDef.NodeList

Field	Name	Type	Wire type	Meaning
1	`value`	repeated string	LEN, one per name	The node names.

CollectionDef.BytesList

Field	Name	Type	Wire type	Meaning
1	`value`	repeated bytes	LEN, one per value	The values, often serialised messages.

CollectionDef.Int64List

Field	Name	Type	Wire type	Meaning
1	`value`	repeated int64, packed	LEN	The values.

CollectionDef.FloatList

Field	Name	Type	Wire type	Meaning
1	`value`	repeated float, packed	LEN	The values.

CollectionDef.AnyList

Field	Name	Type	Wire type	Meaning
1	`value`	repeated Any	LEN, one per value	The values.

SignatureDef

A signature: the named inputs a caller feeds and the named outputs it fetches. The explorer lists every signature of every meta graph with its inputs and outputs.

Field	Name	Type	Wire type	Meaning
1	`inputs`	map<string, TensorInfo>	LEN, one per entry	Inputs by name.
2	`outputs`	map<string, TensorInfo>	LEN, one per entry	Outputs by name.
3	`method_name`	string	LEN	Deprecated: the kind of call, such as `tensorflow/serving/predict`.
4	`defaults`	map<string, TensorProto>	LEN, one per entry	Default values for inputs.

SignatureDef.InputsEntry

Field	Name	Type	Wire type	Meaning
1	`key`	string	LEN	The input’s name, as a caller gives it.
2	`value`	TensorInfo	LEN	The graph tensor it feeds.

SignatureDef.OutputsEntry

Field	Name	Type	Wire type	Meaning
1	`key`	string	LEN	The output’s name.
2	`value`	TensorInfo	LEN	The graph tensor it comes from.

SignatureDef.DefaultsEntry

Field	Name	Type	Wire type	Meaning
1	`key`	string	LEN	The input’s name.
2	`value`	TensorProto	LEN	Its default value.

TensorInfo

Which tensor of the graph a signature input or output is, with its type and shape. Fields 1, 4 and 5 form the oneof encoding.

Field	Name	Type	Wire type	Meaning
1	`name`	string	LEN	The tensor’s name in the graph, as `node:output`, such as `serving_default_x:0`.
4	`coo_sparse`	CooSparse	LEN	A sparse tensor’s three tensors.
5	`composite_tensor`	CompositeTensor	LEN	A composite tensor.
2	`dtype`	DataType	VARINT	The data type.
3	`tensor_shape`	TensorShapeProto	LEN	The shape; −1 for a size not known, such as the batch.

TensorInfo.CooSparse

Field	Name	Type	Wire type	Meaning
1	`values_tensor_name`	string	LEN	The tensor of values.
2	`indices_tensor_name`	string	LEN	The tensor of indices.
3	`dense_shape_tensor_name`	string	LEN	The tensor of the dense shape.

TensorInfo.CompositeTensor

Field	Name	Type	Wire type	Meaning
1	`type_spec`	TypeSpecProto	LEN	The type spec. The explorer does not include this message and shows it by number.
2	`components`	repeated TensorInfo	LEN, one per component	The component tensors.

AssetFileDef

An asset file, such as a vocabulary, and the tensor that receives its path when the model is loaded.

Field	Name	Type	Wire type	Meaning
1	`tensor_info`	TensorInfo	LEN	The tensor fed the file’s path.
2	`filename`	string	LEN	The file’s name in the `assets` folder.

SaverDef

How the variables are saved to and restored from the variables folder.

Field	Name	Type	Wire type	Meaning
1	`filename_tensor_name`	string	LEN	The tensor that takes the checkpoint’s file name.
2	`save_tensor_name`	string	LEN	The op that saves.
3	`restore_op_name`	string	LEN	The op that restores.
4	`max_to_keep`	int32	VARINT	How many checkpoints to keep.
5	`sharded`	bool	VARINT	Whether to save in shards.
6	`keep_checkpoint_every_n_hours`	float	I32	Keep one checkpoint every so many hours.
7	`version`	CheckpointFormatVersion	VARINT	The checkpoint format: LEGACY (0), V1 (1) or V2 (2).

Any

Any message, named by a type URL; Protocol Buffers’ own google.protobuf.Any.

Field	Name	Type	Wire type	Meaning
1	`type_url`	string	LEN	The message type, such as `type.googleapis.com/tensorflow.Foo`.
2	`value`	bytes	LEN	The message, encoded.

.PB File Format – MetaGraphDef