Implement option split to reduce discrepancies for lightgbm regressors

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

@@ -47,7 +47,6 @@ jobs:
       python -m pip install $(ONNX_PATH)
       python -m pip install hummingbird-ml --no-deps
       python -m pip install -r requirements.txt
-      pip install torch==1.8.1+cpu torchvision==0.9.1+cpu torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
       python -m pip install -r requirements-dev.txt
       python -m pip install $(ORT_PATH)
       python -m pip install pytest

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

@@ -9,6 +9,7 @@ trigger:
 jobs:
 
 - job: 'Test'
+  timeoutInMinutes: 25
   pool:
     vmImage: 'ubuntu-latest'
   strategy:
@@ -70,31 +71,28 @@ jobs:
     maxParallel: 3
 
   steps:
-  - script: sudo install -d -m 0777 /home/vsts/.conda/envs
-    displayName: Fix Conda permissions
-
-  - task: CondaEnvironment@1
+  - task: UsePythonVersion@0
     inputs:
-      createCustomEnvironment: true
-      environmentName: 'py$(python.version)'
-      packageSpecs: 'python=$(python.version)'
+      versionSpec: '$(python.version)'
+      architecture: 'x64'
 
   - script: |
       python -m pip install --upgrade pip
-      conda config --set always_yes yes --set changeps1 no
-      conda install -c conda-forge protobuf
-      conda install -c conda-forge numpy
-      conda install -c conda-forge cmake
-      pip install $(COREML_PATH)
-      pip install $(ONNX_PATH)
-      pip install hummingbird-ml --no-deps
+      pip install $(ONNX_PATH) $(ONNXRT_PATH) cython
       pip install -r requirements.txt
-      pip install torch==1.8.1+cpu torchvision==0.9.1+cpu torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
+    displayName: 'Install dependencies'
+
+  - script: |
       pip install -r requirements-dev.txt
+    displayName: 'Install dependencies-dev'
+
+  - script: |
+      python -m pip install --upgrade pip
       pip install xgboost$(xgboost.version)
+      pip install $(ONNX_PATH)
       pip install $(ONNXRT_PATH)
-      pip install pytest
-    displayName: 'Install dependencies'
+      pip install $(COREML_PATH)
+    displayName: 'Install xgboost, onnxruntime'
 
   - script: |
       pip install flake8
@@ -109,8 +107,63 @@ jobs:
       export PYTHONPATH=.
       python -c "import onnxconverter_common;print(onnxconverter_common.__version__)"
       python -c "import onnxruntime;print(onnxruntime.__version__)"
-      pytest tests --doctest-modules --junitxml=junit/test-results.xml
-    displayName: 'pytest - onnxmltools'
+    displayName: 'version'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/baseline --durations=0
+    displayName: 'pytest - baseline'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/catboost --durations=0
+    displayName: 'pytest - catboost'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/coreml --durations=0
+    displayName: 'pytest - coreml'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/lightgbm --durations=0
+    displayName: 'pytest - lightgbm'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/sparkml --durations=0
+    displayName: 'pytest - sparkml'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/svmlib --durations=0
+    displayName: 'pytest - svmlib'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/utils --durations=0
+    displayName: 'pytest - utils'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/xgboost --durations=0
+    displayName: 'pytest - xgboost'
+
+  - script: |
+      export PYTHONPATH=.
+      pip install h2o
+      pytest tests/h2o --durations=0
+    displayName: 'pytest - h2o'
+
+  - script: |
+      pip install torch==1.8.1+cpu torchvision==0.9.1+cpu torchaudio==0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
+      pip install hummingbird-ml --no-deps
+    displayName: 'Install hummingbird-ml'
+
+  - script: |
+      export PYTHONPATH=.
+      pytest tests/hummingbirdml --durations=0
+    displayName: 'pytest - hummingbirdml'
 
   - task: PublishTestResults@2
     inputs:

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

@@ -45,7 +45,6 @@ jobs:
       pip install %COREML_PATH% %ONNX_PATH%
       pip install humming-bird-ml --no-deps
       pip install -r requirements.txt
-      python -m pip install torch==1.8.1+cpu torchvision==0.9.1+cpu torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
       pip install -r requirements-dev.txt
       pip install %ONNXRT_PATH%
     displayName: 'Install dependencies'

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

@@ -9,6 +9,7 @@ trigger:
 jobs:
 
 - job: 'Test'
+  timeoutInMinutes: 30
   pool:
     vmImage: 'windows-latest'
   strategy:
@@ -18,79 +19,63 @@ jobs:
         ONNX_PATH: 'onnx==1.10.1'  # '-i https://test.pypi.org/simple/ onnx==1.9.101'
         ONNXRT_PATH: onnxruntime==1.8.1
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
-        sklearn.version: ''
 
       Python39-190-RT181:
         python.version: '3.9'
         ONNX_PATH: 'onnx==1.9.0'
         ONNXRT_PATH: onnxruntime==1.8.1
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
-        sklearn.version: ''
 
       Python39-190-RT180:
         python.version: '3.9'
         ONNX_PATH: onnx==1.9.0
         ONNXRT_PATH: onnxruntime==1.8.0
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
-        sklearn.version: ''
 
       Python38-181-RT170:
         python.version: '3.8'
         ONNX_PATH: onnx==1.8.1
         ONNXRT_PATH: onnxruntime==1.7.0
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
-        sklearn.version: ''
 
       Python37-180-RT160:
         python.version: '3.7'
         ONNX_PATH: onnx==1.8.0
         ONNXRT_PATH: onnxruntime==1.6.0
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
-        sklearn.version: ''
 
       Python37-160-RT111:
         python.version: '3.7'
         ONNX_PATH: onnx==1.6.0
         ONNXRT_PATH: onnxruntime==1.1.1
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
-        sklearn.version: ''
 
       Python37-170-RT130:
         python.version: '3.7'
         ONNX_PATH: onnx==1.7.0
         ONNXRT_PATH: onnxruntime==1.3.0
         COREML_PATH: git+https://github.com/apple/coremltools@3.1
-        sklearn.version: ''
 
     maxParallel: 3
 
   steps:
-  - task: UsePythonVersion@0
-    inputs:
-      versionSpec: '$(python.version)'
-      architecture: 'x64'
-
   - powershell: Write-Host "##vso[task.prependpath]$env:CONDA\Scripts"
     displayName: Add conda to PATH
 
-  - script: conda create --yes --quiet --name py$(python.version) -c conda-forge python=$(python.version) numpy protobuf
+  - script: conda create --yes --quiet --name py$(python.version) -c conda-forge python=$(python.version) numpy protobuf scikit-learn scipy cython
     displayName: Create Anaconda environment
 
   - script: |
       call activate py$(python.version)
       python -m pip install --upgrade pip numpy
       echo Test numpy installation... && python -c "import numpy"
-      python -m pip install scikit-learn
-      python -m pip install %ONNX_PATH%
-      python -m pip install humming-bird-ml --no-deps
       python -m pip install -r requirements.txt
-      python -m pip install torch==1.8.1+cpu torchvision==0.9.1+cpu torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
     displayName: 'Install dependencies (1)'
 
   - script: |
       call activate py$(python.version)
       python -m pip install -r requirements-dev.txt
-    displayName: 'Install dependencies (2)'
+    displayName: 'Install dependencies-dev'
 
   - script: |
       call activate py$(python.version)
@@ -99,14 +84,10 @@ jobs:
 
   - script: |
       call activate py$(python.version)
+      python -m pip install %ONNX_PATH%
       python -m pip install %ONNXRT_PATH%
     displayName: 'Install onnxruntime'
 
-  - script: |
-      call activate py$(python.version)
-      python -m pip install scikit-learn$(sklearn.version)
-    displayName: 'Install scikit-learn'
-
   - script: |
       call activate py$(python.version)
       python -m flake8 ./onnxmltools
@@ -118,8 +99,67 @@ jobs:
       export PYTHONPATH=.
       python -c "import onnxconverter_common;print(onnxconverter_common.__version__)"
       python -c "import onnxruntime;print(onnxruntime.__version__)"
-      python -m pytest tests --doctest-modules --junitxml=junit/test-results.xml
-    displayName: 'pytest - onnxmltools'
+    displayName: 'version'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/baseline --durations=0
+    displayName: 'pytest baseline'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/catboost --durations=0
+    displayName: 'pytest catboost'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/coreml --durations=0
+    displayName: 'pytest coreml'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/lightgbm --durations=0
+    displayName: 'pytest lightgbm'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/sparkml --durations=0
+    displayName: 'pytest sparkml'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/svmlib --durations=0
+    displayName: 'pytest svmlib'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/utils --durations=0
+    displayName: 'pytest utils'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/xgboost --durations=0
+    displayName: 'pytest xgboost'
+
+  - script: |
+      call activate py$(python.version)
+      python -m pip install torch==1.8.1+cpu torchvision==0.9.1+cpu torchaudio===0.8.1 -f https://download.pytorch.org/whl/torch_stable.html
+      python -m pip install hummingbird-ml --no-deps
+    displayName: 'Install hummingbird-ml'
+
+  - script: |
+      call activate py$(python.version)
+      export PYTHONPATH=.
+      python -m pytest tests/hummingbirdml --durations=0
+    displayName: 'pytest hummingbirdml'
 
   - task: PublishTestResults@2
     inputs:

onnxmltools/convert/lightgbm/_parse.py

@@ -81,17 +81,20 @@ def _get_lightgbm_operator_name(model):
     return lightgbm_operator_name_map[model_type]
 
 
-def _parse_lightgbm_simple_model(scope, model, inputs):
+def _parse_lightgbm_simple_model(scope, model, inputs, split=None):
     '''
     This function handles all non-pipeline models.
 
     :param scope: Scope object
     :param model: A lightgbm object
     :param inputs: A list of variables
+    :param split: split TreeEnsembleRegressor into multiple node to reduce
+        discrepancies
     :return: A list of output variables which will be passed to next stage
     '''
     operator_name = _get_lightgbm_operator_name(model)
     this_operator = scope.declare_local_operator(operator_name, model)
+    this_operator.split = split
     this_operator.inputs = inputs
 
     if operator_name == 'LgbmClassifier':
@@ -151,27 +154,29 @@ def _parse_sklearn_classifier(scope, model, inputs, zipmap=True):
     return this_operator.outputs
 
 
-def _parse_lightgbm(scope, model, inputs, zipmap=True):
+def _parse_lightgbm(scope, model, inputs, zipmap=True, split=None):
     '''
     This is a delegate function. It doesn't nothing but invoke the correct parsing function according to the input
     model's type.
     :param scope: Scope object
     :param model: A lightgbm object
     :param inputs: A list of variables
     :param zipmap: add operator ZipMap after operator TreeEnsembleClassifier
+    :param split: split TreeEnsembleRegressor into multiple node to reduce
+        discrepancies
     :return: The output variables produced by the input model
     '''
     if isinstance(model, LGBMClassifier):
         return _parse_sklearn_classifier(scope, model, inputs, zipmap=zipmap)
     if (isinstance(model, WrappedBooster) and
             model.operator_name == 'LgbmClassifier'):
         return _parse_sklearn_classifier(scope, model, inputs, zipmap=zipmap)
-    return _parse_lightgbm_simple_model(scope, model, inputs)
+    return _parse_lightgbm_simple_model(scope, model, inputs, split=split)
 
 
 def parse_lightgbm(model, initial_types=None, target_opset=None,
                    custom_conversion_functions=None, custom_shape_calculators=None,
-                   zipmap=True):
+                   zipmap=True, split=None):
     raw_model_container = LightGbmModelContainer(model)
     topology = Topology(raw_model_container, default_batch_size='None',
                         initial_types=initial_types, target_opset=target_opset,
@@ -186,7 +191,7 @@ def parse_lightgbm(model, initial_types=None, target_opset=None,
     for variable in inputs:
         raw_model_container.add_input(variable)
 
-    outputs = _parse_lightgbm(scope, model, inputs, zipmap=zipmap)
+    outputs = _parse_lightgbm(scope, model, inputs, zipmap=zipmap, split=split)
 
     for variable in outputs:
         raw_model_container.add_output(variable)