.. DO NOT EDIT.
.. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY.
.. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE:
.. "auto_examples_patterns/plot_pattern_replay_cleanup.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_patterns_plot_pattern_replay_cleanup.py>`
        to download the full example code.

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

.. _sphx_glr_auto_examples_patterns_plot_pattern_replay_cleanup.py:


.. _l-example-plot-pattern-replay-cleanup:

Replaying graph cleanup modifications
=====================================

Graph cleanup algorithms also produce
:class:`~onnx_light.onnx_core.optimization.LocalRewriting` records. These
records replay identity removal, dead-end removal, and initializer
deduplication without running cleanup again.

.. GENERATED FROM PYTHON SOURCE LINES 12-22

.. code-block:: Python


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

    from __future__ import annotations

    from onnx_light.onnx import TensorProto
    import onnx_light.onnx.helper as oh
    from onnx_light.onnx_core.optimization import GraphBuilder, GraphGraph, replay
    from onnx_light.tools import pretty_onnx








.. GENERATED FROM PYTHON SOURCE LINES 23-28

Create and clean up the source model
++++++++++++++++++++++++++++++++++++

This graph contains an ``Identity`` node, a dead-end ``Neg`` node, and two
equal initializers used by retained nodes.

.. GENERATED FROM PYTHON SOURCE LINES 28-62

.. code-block:: Python


    model = oh.make_model(
        oh.make_graph(
            [
                oh.make_node("Add", ["x", "weight"], ["summed"]),
                oh.make_node("Identity", ["summed"], ["forwarded"]),
                oh.make_node("Add", ["forwarded", "duplicate_weight"], ["y"]),
                oh.make_node("Neg", ["x"], ["dead_end"]),
            ],
            "cleanup",
            [oh.make_tensor_value_info("x", TensorProto.FLOAT, [1])],
            [oh.make_tensor_value_info("y", TensorProto.FLOAT, [1])],
            initializer=[
                oh.make_tensor("weight", TensorProto.FLOAT, [1], [1.0]),
                oh.make_tensor("duplicate_weight", TensorProto.FLOAT, [1], [1.0]),
            ],
        ),
        opset_imports=[oh.make_opsetid("", 18)],
    )

    builder = GraphBuilder(model)
    graph = GraphGraph(builder, patterns=False)
    rewrites = list(graph.optimize())
    optimized_graph = builder.build_graph()

    assert {"RemoveIdentityNodes", "RemoveUnusedNodes", "RemoveDuplicateInitializers"} <= {
        rewrite.pattern_name for rewrite in rewrites
    }

    print("Original graph:")
    print(pretty_onnx(model))
    print("Optimized graph:")
    print(pretty_onnx(builder.to_onnx("model")))





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

 .. code-block:: none

    Original graph:
    opset: domain='' version=18
    graph: name='cleanup'
    input: float[1] x
    init: float[1] weight
    init: float[1] duplicate_weight
    0: Add(x, weight) -> summed
    1: Identity(summed) -> forwarded
    2: Add(forwarded, duplicate_weight) -> y
    3: Neg(x) -> dead_end
    output: float[1] y
    Optimized graph:
    opset: domain='ai.onnx' version=18
    graph: name='cleanup'
    input: float[1] x
    init: float[1] weight
    0: Add(x, weight) -> summed
    1: Add(summed, weight) -> y
    output: float[1] y




.. GENERATED FROM PYTHON SOURCE LINES 63-68

Inspect and replay the cleanup modifications
++++++++++++++++++++++++++++++++++++++++++++

Every cleanup operation is captured as a ``LocalRewriting`` record. Replay
applies the records to a fresh copy of the source model.

.. GENERATED FROM PYTHON SOURCE LINES 68-76

.. code-block:: Python


    for rewrite in rewrites:
        print(rewrite)

    replayed_graph = replay(model, rewrites)
    assert replayed_graph.SerializeToString() == optimized_graph.SerializeToString()
    print("Replayed graph:")
    print(pretty_onnx(replayed_graph))




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

 .. code-block:: none

    LocalRewriting(pattern=RemoveIdentityNodes, graph_path=<root>, matched_nodes=4, added_nodes=3)
    LocalRewriting(pattern=RemoveUnusedNodes, graph_path=<root>, matched_nodes=3, added_nodes=2)
    LocalRewriting(pattern=RemoveDuplicateInitializers, graph_path=<root>, matched_nodes=2, added_nodes=2)
    Replayed graph:
    graph: name='cleanup'
    input: float[1] x
    init: float[1] weight
    0: Add(x, weight) -> summed
    1: Add(summed, weight) -> y
    output: float[1] y





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

   **Total running time of the script:** (0 minutes 0.004 seconds)


.. _sphx_glr_download_auto_examples_patterns_plot_pattern_replay_cleanup.py:

.. only:: html

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

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

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

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

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

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

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


.. include:: plot_pattern_replay_cleanup.recommendations


.. only:: html

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

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