.. DO NOT EDIT.
.. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY.
.. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE:
.. "auto_examples_proto/plot_translate.py"
.. LINE NUMBERS ARE GIVEN BELOW.

.. only:: html

    .. note::
        :class: sphx-glr-download-link-note

        :ref:`Go to the end <sphx_glr_download_auto_examples_proto_plot_translate.py>`
        to download the full example code.

.. rst-class:: sphx-glr-example-title

.. _sphx_glr_auto_examples_proto_plot_translate.py:


.. _l-example-plot-translate:

translate: turn an ONNX model back into Python code
====================================================

:func:`~onnx_light.tools.translate` converts an existing ``ModelProto``
(or ``GraphProto``) into Python code that rebuilds an equivalent model.
Two *flavours* are available:

* ``api="onnx-compact"`` — a single nested
  :mod:`onnx_light.onnx.helper` expression
  (``oh.make_model(oh.make_graph([...], ...))``).
* ``api="builder"`` — an incremental script driving the
  :class:`~onnx_light.onnx_core.graph_builder.GraphBuilder`
  (``g.inp(...)``, ``g.init(...)``, ``g.op.<operator>(...)``,
  ``g.out(...)``, ``g.to_onnx(...)``).
* ``api="cpp"`` — a C++ function that rebuilds the model with
  :cpp:class:`~onnx_light::core::builder::GraphBuilder`.

:func:`~onnx_light.tools.translate_header` returns the matching import
header, so ``translate_header(api) + translate(model, api)`` is a fully
runnable Python snippet.  The example below builds a small model, prints
both flavours and then executes the generated code to rebuild the model.

.. GENERATED FROM PYTHON SOURCE LINES 26-43

.. code-block:: Python


    # sphinx_gallery_thumbnail_path = "_static/gallery_thumbnails/translate.png"

    from __future__ import annotations

    import numpy as np

    import onnx_light.onnx as onnx
    import onnx_light.onnx.defs as defs
    import onnx_light.onnx.helper as oh
    import onnx_light.onnx.numpy_helper as onh
    from onnx_light.tools import translate, translate_header

    # Built-in operator schemas are registered so the rebuilt models validate.
    defs.register_onnx_operator_set_schema()









.. GENERATED FROM PYTHON SOURCE LINES 44-49

Build the model
+++++++++++++++

A tiny graph ``Y = Add(Mul(X, W), B)`` with two initializers so the
translation exercises nodes, inputs/outputs and initializers.

.. GENERATED FROM PYTHON SOURCE LINES 49-66

.. code-block:: Python


    model = oh.make_model(
        oh.make_graph(
            [oh.make_node("Mul", ["X", "W"], ["XW"]), oh.make_node("Add", ["XW", "B"], ["Y"])],
            "linear",
            [oh.make_tensor_value_info("X", onnx.TensorProto.FLOAT, ["N", 3])],
            [oh.make_tensor_value_info("Y", onnx.TensorProto.FLOAT, ["N", 3])],
            [
                onh.from_array(np.array([1.0, 2.0, 3.0], dtype=np.float32), name="W"),
                onh.from_array(np.array([0.5, 0.5, 0.5], dtype=np.float32), name="B"),
            ],
        ),
        opset_imports=[oh.make_opsetid("", 18)],
        ir_version=8,
    )









.. GENERATED FROM PYTHON SOURCE LINES 67-73

onnx-compact flavour
++++++++++++++++++++

``translate_header("onnx-compact")`` returns the imports and
``translate(model, api="onnx-compact")`` the nested ``oh.make_model``
expression.

.. GENERATED FROM PYTHON SOURCE LINES 73-79

.. code-block:: Python


    compact_code = translate_header("onnx-compact") + translate(model, api="onnx-compact")
    print("=== onnx-compact ===")
    print(compact_code)






.. rst-class:: sphx-glr-script-out

 .. code-block:: none

    === onnx-compact ===
    import numpy as np
    import ml_dtypes
    import onnx_light.onnx as onnx
    import onnx_light.onnx.helper as oh
    import onnx_light.onnx.numpy_helper as onh
    model = oh.make_model(
        oh.make_graph(
            [
                oh.make_node('Mul', ['X', 'W'], ['XW']),
                oh.make_node('Add', ['XW', 'B'], ['Y']),
            ],
            'linear',
            [
                oh.make_tensor_value_info('X', onnx.TensorProto.FLOAT, ('N', 3)),
            ],
            [
                oh.make_tensor_value_info('Y', onnx.TensorProto.FLOAT, ('N', 3)),
            ],
            [
                onh.from_array(np.array([1.0, 2.0, 3.0], dtype=np.float32), name='W'),
                onh.from_array(np.array([0.5, 0.5, 0.5], dtype=np.float32), name='B'),
            ],
        ),
        opset_imports=[oh.make_opsetid('', 18)],
        ir_version=8,
    )





.. GENERATED FROM PYTHON SOURCE LINES 80-86

builder flavour
+++++++++++++++

The ``builder`` flavour rebuilds the same model step by step with the compact
:class:`~onnx_light.onnx_core.graph_builder.GraphBuilder` API
(``g.inp``, ``g.init``, ``g.op`` and ``g.out``).

.. GENERATED FROM PYTHON SOURCE LINES 86-92

.. code-block:: Python


    builder_code = translate_header("builder") + translate(model, api="builder")
    print("\n=== builder ===")
    print(builder_code)






.. rst-class:: sphx-glr-script-out

 .. code-block:: none


    === builder ===
    import numpy as np
    import ml_dtypes
    import onnx_light.onnx as onnx
    import onnx_light.onnx.helper as oh
    import onnx_light.onnx.numpy_helper as onh
    from onnx_light.onnx_core.graph_builder import GraphBuilder
    g = GraphBuilder('linear')
    g.set_opset_version('', 18)
    g.inp('X', onnx.TensorProto.FLOAT, ('N', 3))
    g.init(np.array([1.0, 2.0, 3.0], dtype=np.float32), name='W')
    g.init(np.array([0.5, 0.5, 0.5], dtype=np.float32), name='B')
    g.op.Mul('X', 'W', outputs=['XW'])
    g.op.Add('XW', 'B', outputs=['Y'])
    g.out('Y', onnx.TensorProto.FLOAT, ('N', 3))
    model = g.to_onnx('model', ir_version=8)





.. GENERATED FROM PYTHON SOURCE LINES 93-98

C++ flavour
+++++++++++

The ``cpp`` flavour emits a standalone ``BuildModel`` function using the
native :cpp:class:`~onnx_light::core::builder::GraphBuilder`.

.. GENERATED FROM PYTHON SOURCE LINES 98-104

.. code-block:: Python


    cpp_code = translate_header("cpp") + translate(model, api="cpp")
    print("\n=== cpp ===")
    print(cpp_code)






.. rst-class:: sphx-glr-script-out

 .. code-block:: none


    === cpp ===
    #include "onnx_core/builder/graph_builder.h"
    #include "onnx_op/operator_sets.h"

    #include <limits>
    #include <string>
    #include <vector>

    namespace onnx_light = ONNX_LIGHT_NAMESPACE;
    onnx_light::ModelProto BuildModel() {
      onnx_light::core::builder::GraphBuilder g(
          "linear", [](const std::string &op_type) {
            return onnx_light::onnx_op::GetAllOnnxOpSchemasWithHistory(op_type, false);
          });
      g.SetOpsetVersion("", 18);
      g.MakeInput("X", onnx_light::core::symbolic::TensorType::kFloat, onnx_light::core::symbolic::SymShape{onnx_light::core::symbolic::SymDim('N'), onnx_light::core::symbolic::SymDim(3)});
      onnx_light::TensorProto initializer_0;
      initializer_0.set_name("W");
      initializer_0.set_data_type(onnx_light::TensorProto::DataType::FLOAT);
      initializer_0.add_dims(3);
      initializer_0.set_raw_data(std::string({static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(128)), static_cast<char>(static_cast<unsigned char>(63)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(64)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(64)), static_cast<char>(static_cast<unsigned char>(64))}));
      g.MakeInitializer(initializer_0);
      onnx_light::TensorProto initializer_1;
      initializer_1.set_name("B");
      initializer_1.set_data_type(onnx_light::TensorProto::DataType::FLOAT);
      initializer_1.add_dims(3);
      initializer_1.set_raw_data(std::string({static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(63)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(63)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(0)), static_cast<char>(static_cast<unsigned char>(63))}));
      g.MakeInitializer(initializer_1);
      g.MakeNode("Mul", {"X", "W"}, {"XW"});
      g.MakeNode("Add", {"XW", "B"}, {"Y"});
      g.MakeOutput("Y", onnx_light::core::symbolic::TensorType::kFloat, onnx_light::core::symbolic::SymShape{onnx_light::core::symbolic::SymDim('N'), onnx_light::core::symbolic::SymDim(3)});
      return g.ToModel(8);
    }





.. GENERATED FROM PYTHON SOURCE LINES 105-111

Round-trip
++++++++++

The generated code is plain Python: executing it rebuilds an equivalent
model.  Here we run the ``builder`` snippet and check that the
rebuilt graph has the same nodes as the original.

.. GENERATED FROM PYTHON SOURCE LINES 111-122

.. code-block:: Python


    namespace: dict = {}
    exec(builder_code, namespace)  # noqa: S102
    rebuilt = namespace["model"]

    original_ops = [node.op_type for node in model.graph.node]
    rebuilt_ops = [node.op_type for node in rebuilt.graph.node]
    print("\n=== round-trip ===")
    print("original ops:", original_ops)
    print("rebuilt ops :", rebuilt_ops)
    assert original_ops == rebuilt_ops




.. rst-class:: sphx-glr-script-out

 .. code-block:: none


    === round-trip ===
    original ops: ['Mul', 'Add']
    rebuilt ops : ['Mul', 'Add']





.. rst-class:: sphx-glr-timing

   **Total running time of the script:** (0 minutes 0.011 seconds)


.. _sphx_glr_download_auto_examples_proto_plot_translate.py:

.. only:: html

  .. container:: sphx-glr-footer sphx-glr-footer-example

    .. container:: sphx-glr-download sphx-glr-download-jupyter

      :download:`Download Jupyter notebook: plot_translate.ipynb <plot_translate.ipynb>`

    .. container:: sphx-glr-download sphx-glr-download-python

      :download:`Download Python source code: plot_translate.py <plot_translate.py>`

    .. container:: sphx-glr-download sphx-glr-download-zip

      :download:`Download zipped: plot_translate.zip <plot_translate.zip>`


.. include:: plot_translate.recommendations


.. only:: html

 .. rst-class:: sphx-glr-signature

    `Gallery generated by Sphinx-Gallery <https://sphinx-gallery.github.io>`_