documentation + arrayratureextraction

_doc/sphinxdoc/source/_exts/generate_automated_pages.py

@@ -95,7 +95,7 @@ def shorten(text):
             title = "Available of scikit-learn model for runtime {0}".format(
                 runtime)
             f.write(dedent('''
-            _l-onnx-bench-{0}:
+            .. _l-onnx-bench-{0}:
 
             {1}
             {2}

_doc/sphinxdoc/source/onnx_bench.rst

@@ -9,7 +9,12 @@ models into :epkg:`ONNX`. Every of them is tested against
 a couple of runtimes. The following pages shows
 which models are correctly converted and compares
 the predictions obtained by every runtime. It also
-displays some benchmark.
+displays some benchmark. The benchmark evaluates
+every model on a dataset inspired from the :epkg:`Iris`
+dataset, so with four features, and different number of
+observations *N= 1, 10, 100, 1000, 100.00, 100.000*.
+The measures for high values of *N* may be missing
+because  the first one took too long.
 
 .. toctree::
     :maxdepth: 1

_doc/sphinxdoc/source/onnx_python.rst

@@ -114,8 +114,8 @@ the cause of the error if it does not work.
 
             piv["ERROR-msg"] = piv["ERROR-msg"].apply(shorten)
 
-        print(df2rst(piv))
+        print(df2rst(piv, number_format=2))
 
     build_table()
 
-Full results are available at :ref:`l-onnx-bench-python`.
+Full results are available at :ref:`l-onnx-bench-CPU`.

_unittests/ut_onnxrt/test_onnxrt_python_runtime_.py

@@ -5,12 +5,16 @@
 from logging import getLogger
 import numpy
 from pyquickhelper.pycode import ExtTestCase
+from pyquickhelper.texthelper.version_helper import compare_module_version
 from skl2onnx.algebra.onnx_ops import (  # pylint: disable=E0611
-    OnnxAdd, OnnxArgMax, OnnxArgMin, OnnxDiv,
+    OnnxAdd, OnnxArgMax, OnnxArgMin,
+    OnnxArrayFeatureExtractor, OnnxDiv,
     OnnxGemm, OnnxMatMul, OnnxMean, OnnxMul,
     OnnxReduceSum, OnnxReduceSumSquare, OnnxSqrt,
     OnnxSub,
 )
+from skl2onnx.common.data_types import FloatTensorType
+from skl2onnx import __version__ as skl2onnx_version
 from mlprodict.onnxrt import OnnxInference
 
 
@@ -82,6 +86,19 @@ def test_onnxt_runtime_argmin(self):
         self.assertEqualArray(numpy.argmin(X, axis=1).ravel(),
                               got['Y'].ravel())
 
+    @unittest.skipIf(compare_module_version(skl2onnx_version, "1.5.0") <= 0,
+                     reason="int64 not implemented for constants")
+    def test_onnxt_runtime_array_feature_extractor(self):
+        onx = OnnxArrayFeatureExtractor('X', numpy.array([1], dtype=numpy.int64),
+                                        output_names=['Y'])
+        X = numpy.array([[1, 2], [3, 4]], dtype=numpy.float64)
+        model_def = onx.to_onnx({'X': X.astype(numpy.float32)},
+                                outputs=[('Y', FloatTensorType([2]))])
+        oinf = OnnxInference(model_def)
+        got = oinf.run({'X': X})
+        self.assertEqual(list(sorted(got)), ['Y'])
+        self.assertEqualArray(X[:, 1], got['Y'], decimal=6)
+
     def test_onnxt_runtime_div(self):
         idi = numpy.identity(2)
         onx = OnnxDiv('X', idi, output_names=['Y'])

mlprodict/onnxrt/ops_cpu/_op_list.py

@@ -9,6 +9,7 @@
 from .op_add import Add
 from .op_argmax import ArgMax
 from .op_argmin import ArgMin
+from .op_array_feature_extractor import ArrayFeatureExtractor
 from .op_cast import Cast
 from .op_div import Div
 from .op_gemm import Gemm

mlprodict/onnxrt/ops_cpu/op_array_feature_extractor.py

@@ -0,0 +1,22 @@
+# -*- encoding: utf-8 -*-
+# pylint: disable=E0203,E1101,C0111
+"""
+@file
+@brief Runtime operator.
+"""
+from ._op import OpRun
+
+
+class ArrayFeatureExtractor(OpRun):
+
+    def __init__(self, onnx_node, desc=None, **options):
+        if desc is None:
+            raise ValueError("desc should not be None.")
+        OpRun.__init__(self, onnx_node, desc=desc,
+                       **options)
+
+    def _run(self, data, indices):  # pylint: disable=W0221
+        index = tuple(indices.tolist())
+        sl = slice(0, data.shape[0])
+        index = (sl, ) + index
+        return (data[index], )

mlprodict/onnxrt/validate.py

@@ -910,15 +910,16 @@ def fct():
     return mes
 
 
-def benchmark_fct(fct, X):
+def benchmark_fct(fct, X, time_limit=4):
     """
     Benchmarks a function which takes an array
     as an input and changes the number of rows.
 
-    @param      fct     function to benchmark, signature
-                        is fct(xo)
-    @param      X       array
-    @return             dictionary with the results
+    @param      fct         function to benchmark, signature
+                            is fct(xo)
+    @param      X           array
+    @param      time_limit  above this time, measurement as stopped
+    @return                 dictionary with the results
     """
 
     def make(x, n):
@@ -948,7 +949,7 @@ def make(x, n):
             number = 1
         res[N] = measure_time(fct, x, repeat=repeat,
                               number=number, div_by_number=True)
-        if res[N] is not None and res[N].get('total', 2.) >= 2.:
+        if res[N] is not None and res[N].get('total', time_limit) >= time_limit:
             # too long
             break
     return res