Updates default opset from 13 to 14

_doc/examples/plot_op_einsum.py

@@ -52,7 +52,7 @@
     torch_einsum = None
 
 
-def build_ort_einsum(equation, op_version=13):
+def build_ort_einsum(equation, op_version=14):  # opset=13, 14, ...
     node = OnnxEinsum('x', 'y', equation=equation,
                       op_version=op_version,
                       output_names=['z'])
@@ -63,7 +63,7 @@ def build_ort_einsum(equation, op_version=13):
     return lambda x, y: sess.run(None, {'x': x, 'y': y})
 
 
-def build_ort_decomposed(equation, op_version=13):
+def build_ort_decomposed(equation, op_version=14):  # opset=13, 14, ...
     cache = _einsum(equation, numpy.float32, opset=op_version,
                     optimize=True, verbose=True, runtime="python")
     if not hasattr(cache, 'onnx_'):

_doc/examples/plot_op_reducemax.py

@@ -43,7 +43,7 @@
     torch_max = None
 
 
-def build_ort_reducemax(axes, op_version=13):
+def build_ort_reducemax(axes, op_version=14):  # opset=13, 14, ...
     node = OnnxReduceMax('x', axes=axes, op_version=op_version,
                          output_names=['z'])
     onx = node.to_onnx(inputs=[('x', FloatTensorType())],

_doc/examples/plot_op_reducemean.py

@@ -43,7 +43,7 @@
     torch_mean = None
 
 
-def build_ort_reducemean(axes, op_version=13):
+def build_ort_reducemean(axes, op_version=14):  # opset=13, 14, ...
     node = OnnxReduceMean('x', axes=axes, op_version=op_version,
                           output_names=['z'])
     onx = node.to_onnx(inputs=[('x', FloatTensorType())],

_doc/examples/plot_op_reducesum.py

@@ -44,7 +44,7 @@
     torch_sum = None
 
 
-def build_ort_reducesum(axes, op_version=13):
+def build_ort_reducesum(axes, op_version=14):  # opset=13, 14, ...
     node = OnnxReduceSumApi11('x', axes=axes, op_version=op_version,
                               output_names=['z'])
     onx = node.to_onnx(inputs=[('x', FloatTensorType())],

_doc/examples/plot_op_reducesumsquare.py

@@ -43,7 +43,7 @@
     torch_sum = None
 
 
-def build_ort_reducesumsquare(axes, op_version=13):
+def build_ort_reducesumsquare(axes, op_version=14):  # opset=13, 14, ...
     node = OnnxReduceSumSquare('x', axes=axes, op_version=op_version,
                                output_names=['z'])
     onx = node.to_onnx(inputs=[('x', FloatTensorType())],

_doc/sphinxdoc/source/blog/2021/2021-05-05_numpyapionnx2.rst

@@ -46,7 +46,7 @@
 
         X_train, X_test, y_train, y_test = train_test_split(X, y)
 
-        @onnxsklearn_class("onnx_transform", op_version=13)
+        @onnxsklearn_class("onnx_transform", op_version=14)  # opset=13, 14, ...
         class DecorrelateTransformerOnnx(TransformerMixin, BaseEstimator):
             def __init__(self, alpha=0.):
                 BaseEstimator.__init__(self)
@@ -69,7 +69,7 @@
         model.fit(X_train)
         print(model.transform(X_test[:5]))
 
-        onx = to_onnx(model, X_test[:5], target_opset=13)
+        onx = to_onnx(model, X_test[:5], target_opset=14)  # opset=13, 14, ...
         print(onx)
 
     The tutorial :ref:`l-numpy-api-for-onnx` extends this example

_doc/sphinxdoc/source/tutorial/numpy_api_onnx.rst

@@ -322,7 +322,7 @@ is used. Let's see how to do it.
 
     X_train, X_test, y_train, y_test = train_test_split(X, y)
 
-    @onnxsklearn_class('onnx_graph', op_version=13)
+    @onnxsklearn_class('onnx_graph', op_version=14)  # opset=13, 14, ...
     class TwoLogisticRegressionOnnx(ClassifierMixin, BaseEstimator):
 
         def __init__(self):
@@ -377,7 +377,7 @@ is used. Let's see how to do it.
     model.fit(X_train, y_train)
     print(model.predict(X_test[:5]), model.predict_proba(X_test[:5]))
 
-    onx = to_onnx(model, X_test[:5], target_opset=13)
+    onx = to_onnx(model, X_test[:5], target_opset=14)  # opset=13, 14, ...
     # print(onx)  # too long to be displayed
 
 The decorator ``@onnxsklearn_class('onnx_graph')``
@@ -396,7 +396,7 @@ When one of them is called, it follows the steps:
 * Create an instance with a runtime if it does not exist
 * Returns the output of the runtime
 
-The instruction ``to_onnx(model, X_test[:5], target_opset=13)`` creates
+The instruction ``to_onnx(model, X_test[:5], target_opset=?)`` creates
 an ONNX graph by calling method *onnx_graph* registered as a converter
 in *skl2onnx*. It is equivalent to something like
 ``model.onnx_graph(X_test[:5]).to_algebra()[0].to_onnx({'X': X})``.
@@ -410,7 +410,7 @@ Custom Transformer
 ^^^^^^^^^^^^^^^^^^
 
 The syntax is the same. The decorator
-``@onnxsklearn_class("onnx_transform", op_version=13)`` detects
+``@onnxsklearn_class("onnx_transform", op_version=?)`` detects
 the class is a transformer and automatically adds method
 *transform*.
 
@@ -434,7 +434,7 @@ the class is a transformer and automatically adds method
 
     X_train, X_test, y_train, y_test = train_test_split(X, y)
 
-    @onnxsklearn_class("onnx_transform", op_version=13)
+    @onnxsklearn_class("onnx_transform", op_version=14)  # opset=13, 14, ...
     class DecorrelateTransformerOnnx(TransformerMixin, BaseEstimator):
         def __init__(self, alpha=0.):
             BaseEstimator.__init__(self)
@@ -457,7 +457,7 @@ the class is a transformer and automatically adds method
     model.fit(X_train)
     print(model.transform(X_test[:5]))
 
-    onx = to_onnx(model, X_test[:5], target_opset=13)
+    onx = to_onnx(model, X_test[:5], target_opset=14)  # opset=13, 14, ...
     print(onx)
 
 More options

_unittests/ut__skl2onnx/test_sklearn_label_encoder_converter.py

@@ -50,7 +50,7 @@ def test_model_label_encoder_int(self):
         model = LabelEncoder()
         data = numpy.array([10, 3, 5, -34, 0], dtype=numpy.int64)
         model.fit(data)
-        for op in sorted(set([9, 10, 11, 12, 13, TARGET_OPSET])):
+        for op in sorted(set([9, 10, 11, 12, 13, 14, TARGET_OPSET])):  # opset=13, 14, ...
             if op > TARGET_OPSET:
                 continue
             with self.subTest(opset=op):

_unittests/ut_asv_benchmark/test_create_asv_benchmark_pyspy.py

@@ -56,6 +56,7 @@ def test_create_asv_benchmark_pyspy_knn(self):
         self.assertNotEmpty(created)
 
         verif = False
+        target_opset = get_opset_number_from_onnx()
         allnames = []
         for path, _, files in os.walk(os.path.join(temp, 'pyspy')):
             for zoo in files:
@@ -67,7 +68,7 @@ def test_create_asv_benchmark_pyspy_knn(self):
                     content = f.read()
                 if (zoo.endswith(
                         "bench_KNNClas_default_k3_b_cl_64_algorithmbrute_n_neighbors3"
-                        "_10000_20_13_double_optcdist-zm0.py") and
+                        "_10000_20_%d_double_optcdist-zm0.py" % target_opset) and
                         compare_module_version(sklearn.__version__, "0.21") >= 0):
                     if "setup_profile" not in content:
                         raise AssertionError(content)

_unittests/ut_onnx_conv/test_onnx_conv_knn.py

@@ -244,7 +244,7 @@ def onnx_test_knn_single_classreg(self, dtype, n_targets=1, debug=False,
 
         if target_opset is None:
             opsets = list(sorted(set([
-                9, 10, 11, 12, 13, get_opset_number_from_onnx()])))
+                9, 10, 11, 12, 13, 14, get_opset_number_from_onnx()])))  # opset=13, 14, ...
         else:
             opsets = [target_opset]
         for ops in opsets:
@@ -475,7 +475,7 @@ def test_onnx_test_knn_transform(self):
         clr = NearestNeighbors(n_neighbors=3)
         clr.fit(X_train)
 
-        for to in (10, 11, 12, 13):
+        for to in (10, 11, 12, 13, 14):  # opset=13, 14, ...
             if to > get_opset_number_from_onnx():
                 break
             try:

_unittests/ut_onnxrt/test_benchmark_replay.py

@@ -17,8 +17,8 @@ def test_benchmark_replay(self):
             enumerate_benchmark_replay(temp, runtime='python')),
             FileNotFoundError)
         res = list(enumerate_validated_operator_opsets(
-            0, fLOG=None, models={"LogisticRegression"}, opset_min=13,
-            opset_max=13, benchmark=False, store_models=True, dump_all=True,
+            0, fLOG=None, models={"LogisticRegression"}, opset_min=14,  # opset=13, 14, ...
+            opset_max=14, benchmark=False, store_models=True, dump_all=True,
             dump_folder=temp, filter_exp=lambda m, p: (
                 "64" not in p and "b-cl" in p and "dec" not in p)))
         self.assertNotEmpty(res)

_unittests/ut_onnxrt/test_onnx_profiling.py

@@ -41,7 +41,7 @@ def test_profile_onnxruntime1(self):
         del model_def.opset_import[:]  # pylint: disable=E1101
         op_set = model_def.opset_import.add()  # pylint: disable=E1101
         op_set.domain = ''
-        op_set.version = 13
+        op_set.version = 14  # opset=13, 14, ...
 
         X = (numpy.random.randn(4, 2) * 100000).astype(  # pylint: disable=E1101
             numpy.float32)

_unittests/ut_onnxrt/test_onnxrt_python_runtime_.py

@@ -15,6 +15,7 @@
 from onnx import TensorProto, __version__ as onnx_version
 from onnx.helper import make_sparse_tensor, make_tensor
 from onnx.defs import onnx_opset_version
+from onnx.numpy_helper import from_array
 from pyquickhelper.pycode import ExtTestCase
 from pyquickhelper.texthelper import compare_module_version
 from sklearn.utils.extmath import softmax
@@ -73,7 +74,7 @@
     from skl2onnx.algebra.onnx_ops import OnnxBatchNormalization_14
 except ImportError:
     OnnxBatchNormalization_14 = None
-from skl2onnx import __version__ as skl2onnx_version
+from skl2onnx import __version__ as skl2onnx_version, __max_supported_opset__
 from mlprodict.onnxrt import OnnxInference
 from mlprodict.tools.asv_options_helper import (
     get_opset_number_from_onnx, get_ir_version_from_onnx)
@@ -82,13 +83,15 @@
     _batchnorm_test_mode, _batchnorm_training_mode)
 from mlprodict.onnxrt.ops_cpu.op_global_average_pool import _global_average_pool
 from mlprodict.onnxrt.ops_cpu._op_onnx_numpy import (  # pylint: disable=E0611,E0401
-    topk_element_min_double, topk_element_max_double, topk_element_fetch_double,
+    topk_element_min_double, topk_element_max_double,
+    topk_element_fetch_double,
     topk_element_min_float, topk_element_max_float, topk_element_fetch_float,
     topk_element_min_int64, topk_element_max_int64, topk_element_fetch_int64)
 from mlprodict.onnxrt.ops_cpu.op_celu import _vcelu1, pycelu
 from mlprodict.onnxrt.ops_cpu.op_topk import topk_sorted_implementation
 from mlprodict.onnxrt.ops_cpu.op_pad import _pad_impl
-from mlprodict.onnxrt.ops_cpu.op_max_pool import _pool_get_output_shape, _pool_impl
+from mlprodict.onnxrt.ops_cpu.op_max_pool import (
+    _pool_get_output_shape, _pool_impl)
 from mlprodict.onnxrt.ops_cpu.op_dropout import _dropout
 from mlprodict.onnxrt.ops_cpu._op_helper import proto2dtype
 from mlprodict.onnx_tools.onnx2py_helper import (
@@ -190,6 +193,11 @@ def setUp(self):
         logger = getLogger('skl2onnx')
         logger.disabled = True
 
+    def test_opset_skl2onnx(self):
+        opset_mlprodict = get_opset_number_from_onnx()
+        opset_skl2onnx = __max_supported_opset__
+        self.assertGreater(opset_skl2onnx, opset_mlprodict)
+
     @ignore_warnings(category=(RuntimeWarning, DeprecationWarning,
                                SparseEfficiencyWarning, PendingDeprecationWarning))
     def common_test_onnxt_runtime_unary(self, onnx_cl, np_fct,
@@ -2866,7 +2874,7 @@ def test_onnxt_runtime_reduce_prod(self):
     def test_onnxt_runtime_reduce_sum(self):
         X = numpy.array([[2, 1], [0, 1]], dtype=float)
 
-        for opset in (10, 11, 12, 13, get_opset_number_from_onnx()):
+        for opset in (10, 11, 12, 13, 14, get_opset_number_from_onnx()):  # opset=13, 14, ...
             if onnx_opset_version() < opset:
                 continue
             if opset < 13:
@@ -2902,7 +2910,7 @@ def test_onnxt_runtime_reduce_sum(self):
             else:
                 self.assertEqual(name, 'ReduceSum_1')
 
-        for opset in (11, 12, 13):
+        for opset in (11, 12, 13, 14):  # opset=13, 14, ...
             if onnx_opset_version() < opset:
                 continue
             onx = OnnxReduceSumApi11('X', output_names=['Y'], axes=1, keepdims=1,
@@ -2969,7 +2977,7 @@ def test_onnxt_runtime_reduce_sum_square(self):
     def test_onnxt_runtime_reduce_sum_noop(self):
         X = numpy.array([], dtype=float).reshape((2, 0))
 
-        for opset in (13, get_opset_number_from_onnx()):
+        for opset in (13, 14, get_opset_number_from_onnx()):  # opset=13, 14, ...
             if onnx_opset_version() < opset:
                 continue
             cl = OnnxReduceSum
@@ -3183,7 +3191,7 @@ def test_onnxt_runtime_slice(self):
 
     @wraplog()
     def test_onnxt_runtime_slice_step_none(self):
-        for opset in [13, get_opset_number_from_onnx()]:
+        for opset in [13, 14, get_opset_number_from_onnx()]:  # opset=13, 14, ...
             if opset > get_opset_number_from_onnx():
                 continue
             with self.subTest(opset=opset):
@@ -3203,7 +3211,7 @@ def test_onnxt_runtime_slice_step_none(self):
 
     @wraplog()
     def test_onnxt_runtime_split(self):
-        for opset in [10, 11, 12, 13, get_opset_number_from_onnx()]:
+        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:  # opset=13, 14, ...
             if opset > get_opset_number_from_onnx():
                 continue
             with self.subTest(opset=opset):
@@ -3251,7 +3259,7 @@ def test_onnxt_runtime_sqrt(self):
 
     @wraplog()
     def test_onnxt_runtime_squeeze(self):
-        for opset in [10, 11, 12, 13, get_opset_number_from_onnx()]:
+        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:  # opset=13, 14, ...
             if opset > get_opset_number_from_onnx():
                 continue
             with self.subTest(opset=opset):
@@ -3436,7 +3444,7 @@ def test_onnxt_runtime_transpose(self):
 
     @wraplog()
     def test_onnxt_runtime_unsqueeze(self):
-        for opset in [10, 11, 12, 13, get_opset_number_from_onnx()]:
+        for opset in [10, 11, 12, 13, 14, get_opset_number_from_onnx()]:  # opset=13, 14, ...
             if opset > get_opset_number_from_onnx():
                 continue
             with self.subTest(opset=opset):
@@ -3749,7 +3757,7 @@ def test_make_sparse_tensor_12(self):
                         def bprint(*args):
                             rows.append(str(args))  # pylint: disable=W0640
                         try:
-                            oinf.run({'X': X.astype(nty)},
+                            oinf.run({'X': X.astype(nty)},  # opset=13, 14, ...
                                      verbose=13, fLOG=bprint)
                         except Exception:  # pylint: disable=W0703
                             pass
@@ -3813,6 +3821,20 @@ def test_make_constant(self):
                     self.assertEqual(list(sorted(got)), ['Ad_C0'])
                     self.assertEqualArray(exp, got['Ad_C0'])
 
+    def test_op_constant(self):
+        for opv in [9, 10, 11, 12, 13, 14]:  # opset=13, 14, ...
+            for dtype in [numpy.float32, numpy.float64,
+                          numpy.int32, numpy.int64]:
+                with self.subTest(opv=opv, dtype=dtype):
+                    X = numpy.array([1], dtype=dtype)
+                    pX = from_array(X)
+                    op = OnnxAdd('X', OnnxConstant(op_version=opv, value=pX),
+                                 output_names=['Y'], op_version=opv)
+                    onx = op.to_onnx({'X': X})
+                    oinf = OnnxInference(onx)
+                    res = oinf.run({'X': X})
+                    self.assertEqualArray(res['Y'], X + X)
+
 
 if __name__ == "__main__":
     # Working

_unittests/ut_onnxrt/test_onnxrt_switch_types.py

@@ -107,7 +107,7 @@ def test_onnxt_iris_gaussian_process_exp_sine_squared_13(self):
         model_def = to_onnx(
             clr, X_train.astype(numpy.float32),
             options={GaussianProcessRegressor: {'return_std': True}},
-            target_opset=13)
+            target_opset=14)  # opset=13, 14, ...
         oinf = OnnxInference(model_def, runtime='python')
 
         res = oinf.run({'X': X_test.astype(numpy.float32)})

_unittests/ut_tools/test_ort.py

@@ -13,10 +13,11 @@
 class TestOrt(ExtTestCase):
 
     def test_prepare_c_profiling(self):
+        opset = 14  # opset=13, 14, ...
         dtype = numpy.float32
         x = numpy.array([1, 2, 4, 5, 5, 4]).astype(
             numpy.float32).reshape((3, 2))
-        cop = OnnxAdd('X', numpy.array([1], dtype=dtype), op_version=13)
+        cop = OnnxAdd('X', numpy.array([1], dtype=dtype), op_version=opset)
         cop2 = OnnxAdd('X', numpy.array([1], dtype=dtype), op_version=13)
         cop3 = OnnxAdd('X', numpy.array([2], dtype=dtype), op_version=13,
                        output_names=['inter'])