.. DO NOT EDIT. .. THIS FILE WAS AUTOMATICALLY GENERATED BY SPHINX-GALLERY. .. TO MAKE CHANGES, EDIT THE SOURCE PYTHON FILE: .. "auto_examples_core/plot_translate_comparison.py" .. LINE NUMBERS ARE GIVEN BELOW. .. only:: html .. note:: :class: sphx-glr-download-link-note :ref:`Go to the end ` to download the full example code. .. rst-class:: sphx-glr-example-title .. _sphx_glr_auto_examples_core_plot_translate_comparison.py: .. _l-plot-translate-comparison: Comparing the four ONNX translation APIs ========================================= :func:`translate ` converts an :class:`onnx.ModelProto` into Python source code that, when executed, recreates the same model. Four output APIs are available: * ``"onnx"`` — uses :mod:`onnx.helper` (``oh.make_node``, ``oh.make_graph``, …) via :class:`~yobx.translate.inner_emitter.InnerEmitter`. * ``"onnx-short"`` — same as ``"onnx"`` but replaces large initializers with random values to keep the snippet compact, via :class:`~yobx.translate.inner_emitter.InnerEmitterShortInitializer`. * ``"light"`` — fluent ``start(…).vin(…).…`` chain, via :class:`~yobx.translate.light_emitter.LightEmitter`. * ``"builder"`` — ``GraphBuilder``-based function wrapper, via :class:`~yobx.translate.builder_emitter.BuilderEmitter`. This example builds a small model, translates it with every API, shows the generated code, and verifies that the ``"onnx"`` snippet can be re-executed to reproduce the original model. .. GENERATED FROM PYTHON SOURCE LINES 25-32 .. code-block:: Python import numpy as np import onnx import onnx.helper as oh import onnx.numpy_helper as onh from yobx.translate import translate, translate_header .. GENERATED FROM PYTHON SOURCE LINES 33-38 Build the model ---------------- We use ``Z = Relu(X @ W + b)`` as a running example: a single ``Gemm`` followed by ``Relu``. .. GENERATED FROM PYTHON SOURCE LINES 38-59 .. code-block:: Python TFLOAT = onnx.TensorProto.FLOAT INT64 = onnx.TensorProto.INT64 W = onh.from_array(np.random.randn(8, 5).astype(np.float32), name="W") b = onh.from_array(np.random.randn(5).astype(np.float32), name="b") model = oh.make_model( oh.make_graph( [oh.make_node("Gemm", ["X", "W", "b"], ["T"]), oh.make_node("Relu", ["T"], ["Z"])], "gemm_relu", [oh.make_tensor_value_info("X", TFLOAT, [None, 8])], [oh.make_tensor_value_info("Z", TFLOAT, [None, 5])], [W, b], ), opset_imports=[oh.make_opsetid("", 17)], ir_version=9, ) print(f"Model: {len(model.graph.node)} node(s), {len(model.graph.initializer)} initializer(s)") .. rst-class:: sphx-glr-script-out .. code-block:: none Model: 2 node(s), 2 initializer(s) .. GENERATED FROM PYTHON SOURCE LINES 60-66 1. ``"onnx"`` API — full initializer values --------------------------------------------- The generated code uses :func:`onnx.helper.make_node`, :func:`onnx.helper.make_graph`, and :func:`onnx.helper.make_model`. Every initializer is serialised as an exact ``np.array(…)`` literal. .. GENERATED FROM PYTHON SOURCE LINES 66-71 .. code-block:: Python code_onnx = translate(model, api="onnx") print("=== api='onnx' ===") print(code_onnx) .. rst-class:: sphx-glr-script-out .. code-block:: none === api='onnx' === opset_imports = [ oh.make_opsetid('', 17), ] inputs = [] outputs = [] nodes = [] initializers = [] sparse_initializers = [] functions = [] initializers.append( onh.from_array( np.array([[0.09113407135009766, -1.9420450925827026, 0.29207441210746765, -0.10008358955383301, 1.1141226291656494], [-0.13821659982204437, -1.373523473739624, 1.2217375040054321, -0.563983142375946, 2.0154261589050293], [-0.2249707281589508, 0.8632537722587585, -0.44700318574905396, -1.6374095678329468, -1.8711106777191162], [0.4297427535057068, 1.4997031688690186, -0.3691343665122986, 0.05789622291922569, -0.10762215405702591], [-1.0069392919540405, 1.942077398300171, 0.05949755385518074, 0.3610239326953888, -0.19343183934688568], [-1.5105663537979126, 0.5791450142860413, -0.47895529866218567, -0.38696902990341187, -0.717380702495575], [0.040528614073991776, -1.2248902320861816, 0.6879969835281372, 1.190011978149414, 0.7228995561599731], [-0.25929516553878784, 0.16251587867736816, 0.7109960317611694, -0.12187383323907852, -0.9744327664375305]], dtype=np.float32), name='W' ) ) initializers.append( onh.from_array( np.array([2.4226601123809814, 1.3718501329421997, 0.5806605219841003, 1.3703954219818115, -1.2084741592407227], dtype=np.float32), name='b' ) ) inputs.append(oh.make_tensor_value_info('X', onnx.TensorProto.FLOAT, shape=(None, 8))) nodes.append( oh.make_node( 'Gemm', ['X', 'W', 'b'], ['T'] ) ) nodes.append( oh.make_node( 'Relu', ['T'], ['Z'] ) ) outputs.append(oh.make_tensor_value_info('Z', onnx.TensorProto.FLOAT, shape=(None, 5))) graph = oh.make_graph( nodes, 'gemm_relu', inputs, outputs, initializers, sparse_initializer=sparse_initializers, ) model = oh.make_model( graph, functions=functions, opset_imports=opset_imports, ir_version=9, ) .. GENERATED FROM PYTHON SOURCE LINES 72-78 2. ``"onnx-short"`` API — large initializers replaced by random values ----------------------------------------------------------------------- Identical to ``"onnx"`` except that initializers with more than 16 elements are replaced by ``np.random.randn(…)`` / ``np.random.randint(…)`` calls. This keeps the snippet readable when dealing with large weight tensors. .. GENERATED FROM PYTHON SOURCE LINES 78-83 .. code-block:: Python code_short = translate(model, api="onnx-short") print("=== api='onnx-short' ===") print(code_short) .. rst-class:: sphx-glr-script-out .. code-block:: none === api='onnx-short' === opset_imports = [ oh.make_opsetid('', 17), ] inputs = [] outputs = [] nodes = [] initializers = [] sparse_initializers = [] functions = [] value = np.random.randn(8, 5).astype(np.float32) initializers.append( onh.from_array( np.array(value, dtype=np.float32), name='W' ) ) initializers.append( onh.from_array( np.array([2.4226601123809814, 1.3718501329421997, 0.5806605219841003, 1.3703954219818115, -1.2084741592407227], dtype=np.float32), name='b' ) ) inputs.append(oh.make_tensor_value_info('X', onnx.TensorProto.FLOAT, shape=(None, 8))) nodes.append( oh.make_node( 'Gemm', ['X', 'W', 'b'], ['T'] ) ) nodes.append( oh.make_node( 'Relu', ['T'], ['Z'] ) ) outputs.append(oh.make_tensor_value_info('Z', onnx.TensorProto.FLOAT, shape=(None, 5))) graph = oh.make_graph( nodes, 'gemm_relu', inputs, outputs, initializers, sparse_initializer=sparse_initializers, ) model = oh.make_model( graph, functions=functions, opset_imports=opset_imports, ir_version=9, ) .. GENERATED FROM PYTHON SOURCE LINES 84-85 Size comparison between the two onnx variants: .. GENERATED FROM PYTHON SOURCE LINES 85-89 .. code-block:: Python print(f"\nFull code length : {len(code_onnx):>6} characters") print(f"Short code length : {len(code_short):>6} characters") .. rst-class:: sphx-glr-script-out .. code-block:: none Full code length : 1906 characters Short code length : 1112 characters .. GENERATED FROM PYTHON SOURCE LINES 90-94 3. ``"light"`` API — fluent chain ----------------------------------- The output is a single method-chain expression (``start(…).vin(…).…``). .. GENERATED FROM PYTHON SOURCE LINES 94-99 .. code-block:: Python code_light = translate(model, api="light") print("=== api='light' ===") print(code_light) .. rst-class:: sphx-glr-script-out .. code-block:: none === api='light' === ( start(opset=17) .cst(np.array([[0.09113407135009766, -1.9420450925827026, 0.29207441210746765, -0.10008358955383301, 1.1141226291656494], [-0.13821659982204437, -1.373523473739624, 1.2217375040054321, -0.563983142375946, 2.0154261589050293], [-0.2249707281589508, 0.8632537722587585, -0.44700318574905396, -1.6374095678329468, -1.8711106777191162], [0.4297427535057068, 1.4997031688690186, -0.3691343665122986, 0.05789622291922569, -0.10762215405702591], [-1.0069392919540405, 1.942077398300171, 0.05949755385518074, 0.3610239326953888, -0.19343183934688568], [-1.5105663537979126, 0.5791450142860413, -0.47895529866218567, -0.38696902990341187, -0.717380702495575], [0.040528614073991776, -1.2248902320861816, 0.6879969835281372, 1.190011978149414, 0.7228995561599731], [-0.25929516553878784, 0.16251587867736816, 0.7109960317611694, -0.12187383323907852, -0.9744327664375305]], dtype=np.float32)) .rename('W') .cst(np.array([2.4226601123809814, 1.3718501329421997, 0.5806605219841003, 1.3703954219818115, -1.2084741592407227], dtype=np.float32)) .rename('b') .vin('X', elem_type=onnx.TensorProto.FLOAT, shape=(None, 8)) .bring('X', 'W', 'b') .Gemm() .rename('T') .bring('T') .Relu() .rename('Z') .bring('Z') .vout(elem_type=onnx.TensorProto.FLOAT, shape=(None, 5)) .to_onnx() ) .. GENERATED FROM PYTHON SOURCE LINES 100-104 4. ``"builder"`` API — GraphBuilder ------------------------------------- The output uses ``GraphBuilder`` to wrap the graph nodes in a Python function. .. GENERATED FROM PYTHON SOURCE LINES 104-109 .. code-block:: Python code_builder = translate(model, api="builder") print("=== api='builder' ===") print(code_builder) .. rst-class:: sphx-glr-script-out .. code-block:: none === api='builder' === def gemm_relu( op: "GraphBuilder", X: "FLOAT[None, 8]", ): W = np.array([[0.09113407135009766, -1.9420450925827026, 0.29207441210746765, -0.10008358955383301, 1.1141226291656494], [-0.13821659982204437, -1.373523473739624, 1.2217375040054321, -0.563983142375946, 2.0154261589050293], [-0.2249707281589508, 0.8632537722587585, -0.44700318574905396, -1.6374095678329468, -1.8711106777191162], [0.4297427535057068, 1.4997031688690186, -0.3691343665122986, 0.05789622291922569, -0.10762215405702591], [-1.0069392919540405, 1.942077398300171, 0.05949755385518074, 0.3610239326953888, -0.19343183934688568], [-1.5105663537979126, 0.5791450142860413, -0.47895529866218567, -0.38696902990341187, -0.717380702495575], [0.040528614073991776, -1.2248902320861816, 0.6879969835281372, 1.190011978149414, 0.7228995561599731], [-0.25929516553878784, 0.16251587867736816, 0.7109960317611694, -0.12187383323907852, -0.9744327664375305]], dtype=np.float32) b = np.array([2.4226601123809814, 1.3718501329421997, 0.5806605219841003, 1.3703954219818115, -1.2084741592407227], dtype=np.float32) T = op.Gemm(X, W, b, outputs=['T']) Z = op.Relu(T, outputs=['Z']) op.Identity(Z, outputs=["Z"]) return Z g = GraphBuilder({'': 17}, ir_version=9) g.make_tensor_input("X", onnx.TensorProto.FLOAT, (None, 8)) gemm_relu(g.op, "X") g.make_tensor_output("Z", onnx.TensorProto.FLOAT, (None, 5), is_dimension=False, indexed=False) model = g.to_onnx() .. GENERATED FROM PYTHON SOURCE LINES 110-115 Round-trip verification ----------------------- The ``"onnx"`` snippet is fully self-contained and executable. Running it should recreate a model with the same graph structure. .. GENERATED FROM PYTHON SOURCE LINES 115-132 .. code-block:: Python header = translate_header("onnx") full_code = header + "\n" + code_onnx ns: dict = {} exec(compile(full_code, "", "exec"), ns) # noqa: S102 recreated = ns["model"] assert isinstance(recreated, onnx.ModelProto) assert len(recreated.graph.node) == len( model.graph.node ), f"Expected {len(model.graph.node)} nodes, got {len(recreated.graph.node)}" assert len(recreated.graph.initializer) == len(model.graph.initializer), ( f"Expected {len(model.graph.initializer)} initializers, " f"got {len(recreated.graph.initializer)}" ) print("\nRound-trip succeeded ✓") .. rst-class:: sphx-glr-script-out .. code-block:: none Round-trip succeeded ✓ .. GENERATED FROM PYTHON SOURCE LINES 133-139 Plot: code size by API ----------------------- The bar chart compares the number of characters produced by each API for the same model. ``"onnx-short"`` is always ≤ ``"onnx"`` because it compresses large initializers. .. GENERATED FROM PYTHON SOURCE LINES 139-160 .. code-block:: Python import matplotlib.pyplot as plt # noqa: E402 api_labels = ["onnx", "onnx-short", "light", "builder"] code_sizes = [len(code_onnx), len(code_short), len(code_light), len(code_builder)] fig, ax = plt.subplots(figsize=(7, 4)) bars = ax.bar(api_labels, code_sizes, color=["#4c72b0", "#dd8452", "#55a868", "#c44e52"]) ax.set_ylabel("Generated code size (characters)") ax.set_title("ONNX translation: code size by API") for bar, size in zip(bars, code_sizes): ax.text( bar.get_x() + bar.get_width() / 2, bar.get_height() * 1.01, str(size), ha="center", va="bottom", fontsize=9, ) plt.tight_layout() plt.show() .. image-sg:: /auto_examples_core/images/sphx_glr_plot_translate_comparison_001.png :alt: ONNX translation: code size by API :srcset: /auto_examples_core/images/sphx_glr_plot_translate_comparison_001.png :class: sphx-glr-single-img .. rst-class:: sphx-glr-timing **Total running time of the script:** (0 minutes 0.156 seconds) .. _sphx_glr_download_auto_examples_core_plot_translate_comparison.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_comparison.ipynb ` .. container:: sphx-glr-download sphx-glr-download-python :download:`Download Python source code: plot_translate_comparison.py ` .. container:: sphx-glr-download sphx-glr-download-zip :download:`Download zipped: plot_translate_comparison.zip ` .. include:: plot_translate_comparison.recommendations .. only:: html .. rst-class:: sphx-glr-signature `Gallery generated by Sphinx-Gallery `_