module `npy.onnx_numpy_compiler`#

Short summary#

module mlprodict.npy.onnx_numpy_compiler

Implements numpy functions with onnx and a runtime.

Classes#

class	truncated documentation
`OnnxNumpyCompiler`	Implements a class which runs onnx graph.
`OnnxNumpyFunction`	Class wrapping a function build with `OnnxNumpyCompiler`.
`OnnxNumpyFunctionInferenceSession`	Overwrites `OnnxNumpyFunction` to run an instance of InferenceSession from onnxruntime.
`OnnxNumpyFunctionOnnxInference`	Overwrites `OnnxNumpyFunction` to run an instance of `OnnxInference`.

Methods#

method	truncated documentation
`__call__`	Executes the function and returns the results.
`__call__`
`__call__`
`__getstate__`	Serializes everything but function fct_. Function fct_ is used to build the onnx graph and is not needed …
`__init__`
`__init__`
`__init__`
`__init__`
`__repr__`	usual
`__setstate__`	Restores serialized data.
`_build_runtime`	Creates the runtime for the ONNX graph.
`_check_`
`_check_`
`_check_`
`_find_hidden_algebras`	Subgraph are using inputs not linked to the others nodes. This function retrieves them as they are stored in …
`_parse_annotation`	Returns the annotations for function fct_.
`_to_onnx`	Returns the onnx graph produced by function fct_.
`_to_onnx_shape`
`to_onnx`	Returns the ONNX graph for the wrapped function. It takes additional arguments to distinguish between multiple graphs. …

Documentation#

Implements numpy functions with onnx and a runtime.

New in version 0.6.

source on GitHub

class mlprodict.npy.onnx_numpy_compiler.OnnxNumpyCompiler(fct, op_version=None, runtime=None, signature=None, version=None, fctsig=None)#

Bases: object

Implements a class which runs onnx graph.

Parameters:

fct – a function with annotations which returns an ONNX graph, it can also be an ONNX graph.
op_version – ONNX opset to use, None for the latest one
runtime – runtime to choose to execute the onnx graph, python, onnxruntime, onnxruntime1
signature – used when the function is not annotated
version – the same function can be instantiated with different type, this parameter is None or a numpy type if the signature allows multiple types, it must an instance of type FctVersion
fctsig – function used to overwrite the fct signature in case this one is using *args, **kwargs

New in version 0.6.

source on GitHub

__call__(*args, **kwargs)#

Executes the function and returns the results.

Parameters:: args – arguments
Returns:: results

source on GitHub

__getstate__()#

Serializes everything but function fct_. Function fct_ is used to build the onnx graph and is not needed anymore.

source on GitHub

__init__(fct, op_version=None, runtime=None, signature=None, version=None, fctsig=None)#

__repr__()#: usual

__setstate__(state)#

Restores serialized data.

source on GitHub

_build_runtime(op_version=None, runtime=None, signature=None, version=None)#

Creates the runtime for the ONNX graph.

Parameters:

op_version – ONNX opset to use, None for the latest one
runtime – runtime to choose to execute the onnx graph, python, onnxruntime, onnxruntime1
signature – used when the function is not annotated

source on GitHub

_find_hidden_algebras(onx_var, onx_algebra)#

Subgraph are using inputs not linked to the others nodes. This function retrieves them as they are stored in attributes alg_hidden_var_. The function looks into every node linked to the inputs and their predecessors.

Parameters:

onx_var – OnnxVar
onx_algebra – OnnxOperator

Returns:

tuple(dictionary {id(obj): (var, obj)}, all instance of OnnxVarGraph)

source on GitHub

_parse_annotation(signature, version)#

Returns the annotations for function fct_.

Parameters:

signature – needed if the annotation is missing, then version might be needed to specify which type to use if the signature allows many
version – version inside the many signatures possible

Returns:

tuple(inputs, outputs, kwargs), each of them is a list of tuple with the name and the dtype, kwargs is the list of additional parameters

source on GitHub

_to_onnx(op_version=None, signature=None, version=None)#

Returns the onnx graph produced by function fct_.

source on GitHub

_to_onnx_shape(shape)#

to_onnx(**kwargs)#

Returns the ONNX graph for the wrapped function. It takes additional arguments to distinguish between multiple graphs. This happens when a function needs to support multiple type.