ONNX types and shapes, field by field

A graph declares the type of each input and output, and optionally of the values inside it, with ValueInfoProto messages. A type is a tensor of an element type and shape, or a sequence, map, optional value, sparse tensor or opaque type built from others. The last three messages on this page, SequenceProto, MapProto and OptionalProto, hold such values themselves; a model rarely contains them, but they are part of the same schema.

0A 01 78 12 0F 0A 0D 08 01 12 09 0A 02 08 01 0A 03 12 01 4E

A ValueInfoProto built for this page: name x, type a FLOAT tensor of shape [1, N]; the explorer shows it as x: FLOAT [1, "N"]

Bytes	Meaning
`0A 01 78`	ValueInfoProto `name`: `x`.
`12 0F`	`type`, a TypeProto of 15 bytes.
`0A 0D`	TypeProto `tensor_type`, 13 bytes.
`08 01`	TypeProto.Tensor `elem_type`: 1, FLOAT.
`12 09`	`shape`, a TensorShapeProto of 9 bytes.
`0A 02 08 01`	A `dim` with `dim_value` 1.
`0A 03 12 01 4E`	A `dim` with `dim_param` `N`.

ValueInfoProto

A named value and its type: a graph input, output or intermediate value.

Field	Name	Type	Wire type	Meaning
1	`name`	string	LEN	The value’s name, as nodes use it.
2	`type`	TypeProto	LEN	Its type and shape.
3	`doc_string`	string	LEN	Documentation.
4	`metadata_props`	repeated StringStringEntryProto	LEN, one per entry	Named metadata (IR 10).

TypeProto

A type. Fields 1, 4, 5, 7, 8 and 9 form a oneof: exactly one kind is set.

Field	Name	Type	Wire type	Meaning
1	`tensor_type`	TypeProto.Tensor	LEN	A tensor: element type and shape.
4	`sequence_type`	TypeProto.Sequence	LEN	A sequence of one element type.
5	`map_type`	TypeProto.Map	LEN	A map from integers or strings to a type.
9	`optional_type`	TypeProto.Optional	LEN	A value that may be absent (IR 8).
8	`sparse_tensor_type`	TypeProto.SparseTensor	LEN	A sparse tensor (IR 8).
7	`opaque_type`	TypeProto.Opaque	LEN	An opaque type, named by domain and name. Part of ONNX-ML until IR 14, which allows it everywhere.
6	`denotation`	string	LEN	A standard meaning for the whole type, such as `IMAGE`.

TypeProto.Tensor

Field	Name	Type	Wire type	Meaning
1	`elem_type`	int32	VARINT	The element type, a data type.
2	`shape`	TensorShapeProto	LEN	The shape, if known. Absent means any shape.

TypeProto.Sequence

Field	Name	Type	Wire type	Meaning
1	`elem_type`	TypeProto	LEN	The type of each element.

TypeProto.Map

Field	Name	Type	Wire type	Meaning
1	`key_type`	int32	VARINT	The key type: an integer data type or STRING.
2	`value_type`	TypeProto	LEN	The value type.

TypeProto.Optional

Field	Name	Type	Wire type	Meaning
1	`elem_type`	TypeProto	LEN	The type of the value, when present.

TypeProto.SparseTensor

Field	Name	Type	Wire type	Meaning
1	`elem_type`	int32	VARINT	The element type.
2	`shape`	TensorShapeProto	LEN	The shape of the dense tensor.

TypeProto.Opaque

Field	Name	Type	Wire type	Meaning
1	`domain`	string	LEN	The domain; the model’s own when missing.
2	`name`	string	LEN	The type’s name.

TensorShapeProto

A shape: a list of dimensions, outermost first. No dimensions is a scalar.

Field	Name	Type	Wire type	Meaning
1	`dim`	repeated Dimension	LEN, one per dimension	The dimensions.

TensorShapeProto.Dimension

One dimension: a number, or a name for a size not known in advance, such as the batch size. Fields 1 and 2 form a oneof; with neither the size is unknown.

Field	Name	Type	Wire type	Meaning
1	`dim_value`	int64	VARINT	The size.
2	`dim_param`	string	LEN	A symbolic size, such as `N` or `batch`; dimensions with the same name have the same size.
3	`denotation`	string	LEN	A standard meaning, such as `DATA_BATCH` or `DATA_CHANNEL`.

SequenceProto

A sequence of values of one kind, from onnx-data.proto. As a file of its own it is the ONNX sequence or optional value format.

Field	Name	Type	Wire type	Meaning
1	`name`	string	LEN	The value’s name.
2	`elem_type`	int32	VARINT	What the elements are: UNDEFINED 0, TENSOR 1, SPARSE_TENSOR 2, SEQUENCE 3, MAP 4, OPTIONAL 5.
3	`tensor_values`	repeated TensorProto	LEN, one per element	Elements, when TENSOR.
4	`sparse_tensor_values`	repeated SparseTensorProto	LEN, one per element	Elements, when SPARSE_TENSOR.
5	`sequence_values`	repeated SequenceProto	LEN, one per element	Elements, when SEQUENCE.
6	`map_values`	repeated MapProto	LEN, one per element	Elements, when MAP.
7	`optional_values`	repeated OptionalProto	LEN, one per element	Elements, when OPTIONAL.

MapProto

A map: keys of one integer type or strings, and the values as a sequence of the same length.

Field	Name	Type	Wire type	Meaning
1	`name`	string	LEN	The value’s name.
2	`key_type`	int32	VARINT	The key type: an integer data type or STRING.
3	`keys`	repeated int64	VARINT per key, or packed LEN	Integer keys.
4	`string_keys`	repeated bytes	LEN, one per key	String keys.
5	`values`	SequenceProto	LEN	The values, in the order of the keys.

OptionalProto

A value that may be absent. The same field numbers and types as SequenceProto, but each value field holds one value.

Field	Name	Type	Wire type	Meaning
1	`name`	string	LEN	The value’s name.
2	`elem_type`	int32	VARINT	What the value is, with the same numbers as SequenceProto’s; 0 when absent.
3	`tensor_value`	TensorProto	LEN	The value, when TENSOR.
4	`sparse_tensor_value`	SparseTensorProto	LEN	The value, when SPARSE_TENSOR.
5	`sequence_value`	SequenceProto	LEN	The value, when SEQUENCE.
6	`map_value`	MapProto	LEN	The value, when MAP.
7	`optional_value`	OptionalProto	LEN	The value, when OPTIONAL.

.ONNX File Format – Types and shapes