onnx_ir.schemas

The onnx_ir.schemas module provides classes for representing operator signatures and their parameters. These classes are useful for introspecting operator definitions and validating operator usage.

OpSignature

The OpSignature class represents the schema for an ONNX operator, including its inputs, outputs, and attributes.

class onnx_ir.schemas.OpSignature(domain, name, overload, params, outputs, since_version=1)[source]

Schema for an operator.

Parameters:
domain

Domain of the operator. E.g. “”.

Type:

str

name

Name of the operator. E.g. “Add”.

Type:

str

overload

Overload name of the operator.

Type:

str

params

Input parameters. When the op is an ONNX function definition, the order is according to the function signature. This mean we can interleave ONNX inputs and ONNX attributes in the list.

Type:

collections.abc.Sequence[onnx_ir.schemas.Parameter | onnx_ir.schemas.AttributeParameter]

outputs

Output parameters.

Type:

collections.abc.Sequence[onnx_ir.schemas.Parameter]

since_version

The version of the operator set. E.g. 1.

Type:

int

property attributes: Sequence[AttributeParameter]

Returns the attribute parameters.

domain: str
classmethod from_op_schema(op_schema)[source]

Produce an OpSignature from an ONNX OpSchema.

Parameters:

op_schema (OpSchema)

Return type:

OpSignature

get(name, default=None)[source]
Parameters:
Return type:

Parameter | AttributeParameter | None

property inputs: Sequence[Parameter]

Returns the input parameters.

name: str
outputs: Sequence[Parameter]
overload: str
params: Sequence[Parameter | AttributeParameter]
params_map: Mapping[str, Parameter | AttributeParameter]
since_version: int = 1

Parameter

The Parameter class represents a formal input parameter of an operator.

class onnx_ir.schemas.Parameter(name, type_constraint, required, variadic, homogeneous=True, min_arity=1, default=_EMPTY_DEFAULT)[source]

A formal parameter of an operator.

Parameters:
default: Any = _EMPTY_DEFAULT
has_default()[source]
Return type:

bool

homogeneous: bool = True
is_attribute()[source]
Return type:

bool

is_param()[source]

This parameter is an ONNX input or output parameter, as opposed to an ONNX attribute parameter.

Return type:

bool

min_arity: int = 1
name: str
required: bool
type_constraint: TypeConstraintParam
variadic: bool

AttributeParameter

The AttributeParameter class represents an attribute parameter in the operator signature.

class onnx_ir.schemas.AttributeParameter(name, type, required, default=None)[source]

A parameter in the function signature that represents an ONNX attribute.

Parameters:
default: Attr | None = None
has_default()[source]
Return type:

bool

is_attribute()[source]

This parameter is an ONNX attribute parameter, as opposed to an ONNX input or output parameter.

Return type:

bool

is_param()[source]
Return type:

bool

name: str
required: bool
type: AttributeType

TypeConstraintParam

The TypeConstraintParam class represents type constraints for parameters, specifying which types are allowed.

class onnx_ir.schemas.TypeConstraintParam(name, allowed_types, description='')[source]

Type constraint for a parameter.

Parameters:
name

Name of the parameter. E.g. “TFloat”

Type:

str

allowed_types

Allowed types for the parameter.

Type:

frozenset[onnx_ir.TypeProtocol]

description

Human-readable description of the type constraint.

Type:

str

allowed_types: frozenset[TypeProtocol]
classmethod any_tensor(name, description='')[source]
Parameters:

  • name (str)

  • description (str)

Return type:

TypeConstraintParam

classmethod any_value(name, description='')[source]
Parameters:

  • name (str)

  • description (str)

Return type:

TypeConstraintParam

description: str = ''
name: str