Fixes for onnx==1.11.0 (#527)

* add onnx==1.11.0
* Fixes one unit test involving sklearn-onnx.
* use CPUExecutionProvider
* remove one unnecessary initializer

.azure-pipelines/linux-conda-CI.yml

@@ -14,9 +14,15 @@ jobs:
     vmImage: 'ubuntu-latest'
   strategy:
     matrix:
+      Python39-1110-RT1100:
+        python.version: '3.9'
+        ONNX_PATH: onnx==1.11.0  # '-i https://test.pypi.org/simple/ onnx==1.9.101'
+        ONNXRT_PATH: onnxruntime==1.10.0
+        COREML_PATH: git+https://github.com/apple/coremltools@3.1
+        xgboost.version: '>=1.2'
       Python39-1101-RT190:
         python.version: '3.9'
-        ONNX_PATH: onnx==1.10.1  # '-i https://test.pypi.org/simple/ onnx==1.9.101'
+        ONNX_PATH: onnx==1.10.1
         ONNXRT_PATH: onnxruntime==1.9.0
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
         xgboost.version: '>=1.2'

.azure-pipelines/win32-conda-CI.yml

@@ -14,9 +14,15 @@ jobs:
     vmImage: 'windows-latest'
   strategy:
     matrix:
-      Python39-1101-RT190:
+      Python39-1110-RT190:
         python.version: '3.9'
-        ONNX_PATH: 'onnx==1.10.1'  # '-i https://test.pypi.org/simple/ onnx==1.9.101'
+        ONNX_PATH: 'onnx==1.11.0'  # '-i https://test.pypi.org/simple/ onnx==1.9.101'
+        ONNXRT_PATH: onnxruntime==1.10.0
+        COREML_PATH: git+https://github.com/apple/coremltools@3.1
+
+      Python39-1102-RT190:
+        python.version: '3.9'
+        ONNX_PATH: 'onnx==1.10.2'  # '-i https://test.pypi.org/simple/ onnx==1.9.101'
         ONNXRT_PATH: onnxruntime==1.9.0
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
 

onnxmltools/__init__.py

@@ -5,7 +5,7 @@
 This framework converts any machine learned model into onnx format
 which is a common language to describe any machine learned model.
 """
-__version__ = "1.10.0"
+__version__ = "1.11.0"
 __author__ = "Microsoft"
 __producer__ = "OnnxMLTools"
 __producer_version__ = __version__

onnxmltools/convert/sparkml/operator_converters/word2vec.py

@@ -28,8 +28,6 @@ def convert_word2vec(scope, operator, container):
     container.add_initializer(one, onnx_proto.TensorProto.INT64, [1], [1])
     zero = scope.get_unique_variable_name('zero_tensor')
     container.add_initializer(zero, onnx_proto.TensorProto.INT64, [1], [0])
-    word_count_tensor = scope.get_unique_variable_name('word_count_tensor')
-    container.add_initializer(word_count_tensor, onnx_proto.TensorProto.INT64, [1], [word_count])
 
     sliced_outputs = []
     for i in range(0, word_count):

tests/lightgbm/test_LightGbmTreeEnsembleConvertersPkl.py

@@ -17,6 +17,8 @@
 
 TARGET_OPSET = min(DEFAULT_OPSET_NUMBER, onnx_opset_version())
 
+ort_version = ".".join(onnxruntime.__version__.split('.')[:2])
+
 
 class TestLightGbmTreeEnsembleModelsPkl(unittest.TestCase):
 
@@ -37,7 +39,7 @@ def test_root_leave(self):
     @unittest.skipIf(sys.platform.startswith('lin'), reason="recover linux CI build, needs to be fixed")
     @unittest.skipIf(not hummingbird_installed(), reason="Hummingbird is not installed")
     @unittest.skipIf(
-        StrictVersion(onnxruntime.__version__) < StrictVersion('1.0.0'), reason="Hummingbird supports only latest versions of ORT"
+        StrictVersion(ort_version) < StrictVersion('1.0.0'), reason="Hummingbird supports only latest versions of ORT"
     )
     def test_root_leave_onnx_only(self):
         this = os.path.abspath(os.path.dirname(__file__))

tests/lightgbm/test_LightGbmTreeEnsembleConverters_split.py

@@ -22,6 +22,7 @@
 
 
 TARGET_OPSET = min(DEFAULT_OPSET_NUMBER, onnx_opset_version())
+ort_version = ".".join(ort_version.split('.')[:2])
 
 
 class TestLightGbmTreeEnsembleModelsSplit(unittest.TestCase):

tests/lightgbm/test_objective_functions.py

@@ -71,7 +71,7 @@ def _assert_almost_equal(actual: np.array, desired: np.array, decimal: int=7, fr
         success_rel = success_abs / len(actual)
         assert success_rel >= frac, f"Only {success_abs} out of {len(actual)} rows are almost equal to {decimal} decimals."
 
-    @unittest.skipIf(tuple(int(ver) for ver in onnxruntime.__version__.split(".")) < (1, 3), "not supported in this library version")
+    @unittest.skipIf(tuple(int(ver) for ver in onnxruntime.__version__.split(".")[:2]) < (1, 3), "not supported in this library version")
     def test_objective(self):
         """
         Test if a LGBMRegressor a with certain objective (e.g. 'poisson') can be converted to ONNX

tests/sparkml/sparkml_test_utils.py

@@ -71,7 +71,7 @@ def save_data_models(input, expected, model, onnx_model, basename="model", folde
 
 
 def run_onnx_model(output_names, input, onnx_model):
-    sess = onnxruntime.InferenceSession(onnx_model)
+    sess = onnxruntime.InferenceSession(onnx_model, providers=['CPUExecutionProvider'])
     if isinstance(input, dict):
         inputs = input
     elif isinstance(input, list):

tests/xgboost/test_xgboost_pipeline.py

@@ -34,6 +34,7 @@
 
 
 TARGET_OPSET = min(DEFAULT_OPSET_NUMBER, onnx_opset_version())
+TARGET_OPSET_ML = 2
 
 
 @unittest.skipIf(sys.version_info[:2] <= (3, 5), reason="not available")
@@ -134,7 +135,7 @@ def common_test_xgboost_10_skl(self, missing, replace=False):
             input_xgb[input_xgb[:, :] == missing] = np.nan
         onnx_last = convert_sklearn(model.steps[1][-1],
                                     initial_types=[('X', FloatTensorType(shape=[None, input_xgb.shape[1]]))],
-                                    target_opset=TARGET_OPSET)
+                                    target_opset={'': TARGET_OPSET, 'ai.onnx.ml': TARGET_OPSET_ML})
         session = rt.InferenceSession(onnx_last.SerializeToString())
         pred_skl = model.steps[1][-1].predict(input_xgb).ravel()
         pred_onx = session.run(None, {'X': input_xgb})[0].ravel()