Fixes #187, custom implementation for operator Einsum

_doc/examples/plot_experimental.py

@@ -0,0 +1,94 @@
+"""
+.. _l-example-experimental:
+
+Compares implementation of stanard function
+===========================================
+
+The following function benchmark different implementation
+of standard function.
+
+
+
+.. contents::
+    :local:
+
+Einsum
+++++++
+"""
+import numpy
+import pandas
+import matplotlib.pyplot as plt
+from tqdm import tqdm
+from cpyquickhelper.numbers.speed_measure import measure_time
+from mlprodict.testing.experimental_c import custom_einsum_double
+from onnxruntime import InferenceSession
+from skl2onnx.algebra.onnx_ops import OnnxEinsum
+from skl2onnx.common.data_types import DoubleTensorType
+import onnx
+
+
+def build_ort_einsum(equation, op_version=12):
+    node = OnnxEinsum('x', 'y', equation=equation,
+                      op_version=op_version,
+                      output_names=['z'])
+    onx = node.to_onnx(inputs=[('x', DoubleTensorType()), ('y', DoubleTensorType())],
+                       target_opset=op_version)
+    sess = InferenceSession(onx.SerializeToString())
+    return lambda x, y: sess.run(None, {'x': x, 'y': y})
+
+
+equation = "bsnh,btnh->bnts"
+ort_einsum = build_ort_einsum(equation)
+res = []
+for dim in tqdm([8, 16, 32, 64, 128, 256, 512]):
+    x = numpy.random.rand(1, dim, 12, 64)
+    y = numpy.random.rand(1, dim, 12, 64)
+
+    ort_einsum(x, y)
+
+    ctx = dict(equation=equation, x=x, y=y, einsum=numpy.einsum)
+    obs = measure_time("einsum(equation, x, y)", div_by_number=True, context=ctx,
+                       repeat=5, number=5)
+    obs['dim'] = dim
+    obs['fct'] = 'numpy.einsum'
+    res.append(obs)
+
+    ctx['einsum'] = ort_einsum
+    obs = measure_time("einsum(x, y)", div_by_number=True, context=ctx,
+                       repeat=5, number=5)
+    obs['dim'] = dim
+    obs['fct'] = 'ort_einsum'
+    res.append(obs)
+
+    ctx['einsum'] = custom_einsum_double
+    obs = measure_time("einsum(equation, x, y)", div_by_number=True, context=ctx,
+                       repeat=5, number=5)
+    obs['dim'] = dim
+    obs['fct'] = 'custom_einsum_double'
+    res.append(obs)
+
+df = pandas.DataFrame(res)
+df
+print(df.T)
+
+###########################################
+# Pivot
+
+piv = df.pivot('dim', 'fct', 'average')
+piv
+
+###########################################
+# Ratios
+
+rs = piv.copy()
+rs['custom_einsum_double'] = rs['numpy.einsum'] / rs['custom_einsum_double']
+rs['ort_einsum'] = rs['numpy.einsum'] / rs['ort_einsum']
+rs['numpy.einsum'] = 1.
+rs
+
+###########################################
+# Graphs.
+fig, ax = plt.subplots(1, 2, figsize=(12, 4))
+piv.plot(logx=True, logy=True, ax=ax[0], title="Einsum benchmark")
+rs.plot(logx=True, ax=ax[1], title="Einsum Speedup, baseline=numpy")
+plt.show()

_doc/sphinxdoc/source/api/index.rst

@@ -13,4 +13,5 @@ API
     sklapi
     asv
     validation
+    testing
     tools

_doc/sphinxdoc/source/api/testing.rst

@@ -0,0 +1,17 @@
+
+testing
+=======
+
+.. contents::
+    :local:
+
+Experimental
+++++++++++++
+
+Experimental implementations for algorithm.
+
+.. autosignature:: mlprodict.testing.experimental.custom_einsum
+
+.. autosignature:: mlprodict.testing.experimental_c.custom_einsum_double
+
+.. autosignature:: mlprodict.testing.experimental.custom_pad

_unittests/ut__skl2onnx/test_sklearn_cast_transformer.py

@@ -4,12 +4,12 @@
 import unittest
 import math
 import numpy
+from onnxruntime import InferenceSession
 from sklearn.datasets import make_regression
 from sklearn.model_selection import train_test_split
 from sklearn.pipeline import Pipeline
 from sklearn.preprocessing import StandardScaler
 from sklearn.tree import DecisionTreeRegressor
-from onnxruntime import InferenceSession
 from skl2onnx.sklapi import CastTransformer
 from skl2onnx import convert_sklearn, to_onnx
 from skl2onnx.common.data_types import (

_unittests/ut__skl2onnx/test_sklearn_gaussian_mixture_converter.py

@@ -3,10 +3,10 @@
 """
 import unittest
 import numpy as np
-from sklearn.datasets import load_iris
-from sklearn.mixture import GaussianMixture, BayesianGaussianMixture
 from onnxruntime import InferenceSession
 from onnxruntime.capi.onnxruntime_pybind11_state import Fail as OrtFail  # pylint: disable=E0611
+from sklearn.datasets import load_iris
+from sklearn.mixture import GaussianMixture, BayesianGaussianMixture
 from skl2onnx import convert_sklearn, to_onnx
 from skl2onnx.common.data_types import FloatTensorType
 from mlprodict.testing.test_utils import dump_data_and_model, TARGET_OPSET

_unittests/ut__skl2onnx/test_sklearn_gaussian_process.py

@@ -7,6 +7,7 @@
 import numpy as np
 import pandas as pd
 from numpy.testing import assert_almost_equal
+from onnxruntime import __version__ as ort_version
 from sklearn.datasets import load_iris
 from sklearn.gaussian_process import GaussianProcessRegressor
 from sklearn.gaussian_process.kernels import (
@@ -18,7 +19,6 @@
 from pyquickhelper.texthelper import compare_module_version
 from skl2onnx.common.data_types import FloatTensorType, DoubleTensorType
 from skl2onnx import to_onnx, __version__ as skl2_vers
-from onnxruntime import __version__ as ort_version
 from mlprodict.onnxrt import OnnxInference
 from mlprodict.testing.test_utils import (
     dump_data_and_model, fit_regression_model, TARGET_OPSET)

_unittests/ut__skl2onnx/test_sklearn_pipeline.py

@@ -8,6 +8,7 @@
 import numpy
 from numpy.testing import assert_almost_equal
 import pandas
+from onnxruntime import __version__ as ort_version, InferenceSession
 from sklearn import __version__ as sklearn_version
 from sklearn import datasets
 from sklearn.compose import ColumnTransformer
@@ -23,7 +24,6 @@
 from skl2onnx import convert_sklearn
 from skl2onnx.common.data_types import (
     FloatTensorType, Int64TensorType, StringTensorType)
-from onnxruntime import __version__ as ort_version, InferenceSession
 from mlprodict.testing.test_utils import (
     dump_data_and_model, fit_classification_model)
 

_unittests/ut_documentation/test_run_notebooks_onnx_sbs.py

@@ -4,6 +4,7 @@
 """
 import os
 import unittest
+from onnxruntime import __version__ as ort_version
 from sklearn.exceptions import ConvergenceWarning
 try:
     from sklearn.utils._testing import ignore_warnings
@@ -16,7 +17,6 @@
     add_missing_development_version, ExtTestCase
 )
 from skl2onnx import __version__ as skl2onnx_version
-from onnxruntime import __version__ as ort_version
 import mlprodict
 
 

_unittests/ut_onnx_conv/test_onnx_conv_knn.py

@@ -7,6 +7,7 @@
 import numpy
 from pandas import DataFrame
 from scipy.spatial.distance import cdist as scipy_cdist
+from onnxruntime.capi.onnxruntime_pybind11_state import InvalidArgument as OrtInvalidArgument  # pylint: disable=E0611
 from pyquickhelper.pycode import ExtTestCase
 from sklearn.calibration import CalibratedClassifierCV
 from sklearn.datasets import load_iris, make_regression
@@ -24,7 +25,6 @@
 from skl2onnx.common.data_types import Int64TensorType
 import skl2onnx
 from skl2onnx.algebra.complex_functions import onnx_cdist
-from onnxruntime.capi.onnxruntime_pybind11_state import InvalidArgument as OrtInvalidArgument  # pylint: disable=E0611
 from mlprodict.onnx_conv import (
     register_converters, to_onnx)
 from mlprodict.onnxrt import OnnxInference

_unittests/ut_onnxrt/test_onnxrt_side_by_side.py

@@ -6,10 +6,10 @@
 from logging import getLogger
 import numpy
 import pandas
+from onnxruntime import __version__ as ort_version
 from sklearn.gaussian_process.kernels import RBF, ConstantKernel as CK, Sum
 from pyquickhelper.pycode import ExtTestCase
 from pyquickhelper.texthelper.version_helper import compare_module_version
-from onnxruntime import __version__ as ort_version
 from skl2onnx.common.data_types import FloatTensorType
 try:
     from skl2onnx.operator_converters.gaussian_process import convert_kernel

_unittests/ut_onnxrt/test_onnxrt_validate_bug.py

@@ -5,9 +5,9 @@
 import unittest
 import numpy
 import onnx
+from onnxruntime import InferenceSession
 from pyquickhelper.pycode import ExtTestCase
 from skl2onnx.algebra.onnx_ops import OnnxAdd, OnnxMatMul  # pylint: disable=E0611
-from onnxruntime import InferenceSession
 from mlprodict.onnxrt import OnnxInference
 from mlprodict.tools import get_opset_number_from_onnx
 

_unittests/ut_onnxrt/test_rt_valid_model_gaussian_process_ort.py

@@ -4,6 +4,7 @@
 import unittest
 from logging import getLogger
 import numpy
+from onnxruntime import __version__ as ort_version
 from pyquickhelper.loghelper import fLOG
 from pyquickhelper.pycode import ExtTestCase, skipif_circleci
 from pyquickhelper.texthelper.version_helper import compare_module_version
@@ -16,7 +17,6 @@
 from sklearn.gaussian_process.kernels import RBF, ExpSineSquared
 from skl2onnx import __version__ as skl2onnx_version
 from skl2onnx.common.data_types import FloatTensorType
-from onnxruntime import __version__ as ort_version
 from mlprodict.onnxrt.validate import enumerate_validated_operator_opsets
 from mlprodict.onnxrt import OnnxInference
 from mlprodict.tools.asv_options_helper import get_ir_version_from_onnx

_unittests/ut_onnxrt/test_rt_valid_model_gaussian_process_ort2.py

@@ -3,6 +3,7 @@
 """
 import unittest
 from logging import getLogger
+from onnxruntime import __version__ as ort_version
 from pyquickhelper.loghelper import fLOG
 from pyquickhelper.pycode import ExtTestCase, skipif_circleci
 from pyquickhelper.texthelper.version_helper import compare_module_version
@@ -12,7 +13,6 @@
 except ImportError:
     from sklearn.utils.testing import ignore_warnings
 from skl2onnx import __version__ as skl2onnx_version
-from onnxruntime import __version__ as ort_version
 from mlprodict.onnxrt.validate import enumerate_validated_operator_opsets