Add support for ONNX model export and execution. Merge to AML branch

build/libs_win.txt

@@ -14,3 +14,4 @@ TensorFlow.NET.dll
 NumSharp.Core.dll
 System.Drawing.Common.dll
 Microsoft.ML.*
+onnxruntime.dll

src/DotNetBridge/Bridge.cs

@@ -298,7 +298,7 @@ private static unsafe int GenericExec(EnvironmentBlock* penv, sbyte* psz, int cd
             //env.ComponentCatalog.RegisterAssembly(typeof(AutoInference).Assembly); // ML.PipelineInference
             env.ComponentCatalog.RegisterAssembly(typeof(DataViewReference).Assembly);
             env.ComponentCatalog.RegisterAssembly(typeof(ImageLoadingTransformer).Assembly);
-            //env.ComponentCatalog.RegisterAssembly(typeof(SaveOnnxCommand).Assembly);
+            env.ComponentCatalog.RegisterAssembly(typeof(OnnxExportExtensions).Assembly);
             //env.ComponentCatalog.RegisterAssembly(typeof(TimeSeriesProcessingEntryPoints).Assembly);
             //env.ComponentCatalog.RegisterAssembly(typeof(ParquetLoader).Assembly);
             env.ComponentCatalog.RegisterAssembly(typeof(SsaChangePointDetector).Assembly);

src/DotNetBridge/DotNetBridge.csproj

@@ -38,6 +38,7 @@
     <PackageReference Include="Microsoft.ML.Mkl.Components" Version="1.5.0-preview*" />
     <PackageReference Include="Microsoft.ML.ImageAnalytics" Version="1.5.0-preview*" />
     <PackageReference Include="Microsoft.ML.LightGBM" Version="1.5.0-preview*" />
+    <PackageReference Include="Microsoft.ML.OnnxConverter" Version="0.17.0-preview*" />
     <PackageReference Include="Microsoft.ML.OnnxTransformer" Version="1.5.0-preview*" />
     <PackageReference Include="Microsoft.ML.TensorFlow" Version="1.5.0-preview*" />
     <PackageReference Include="Microsoft.ML.Vision" Version="1.5.0-preview*" />

src/DotNetBridge/Entrypoints.cs

@@ -178,5 +178,59 @@ public static ScoringTransformOutput Score(IHostEnvironment env, ScoringTransfor
                 };
 
         }
+
+        public sealed class OnnxTransformInput : TransformInputBase
+        {
+            [Argument(ArgumentType.Required, HelpText = "Path to the onnx model file.", ShortName = "model", SortOrder = 0)]
+            public string ModelFile;
+
+            [Argument(ArgumentType.Multiple, HelpText = "Name of the input column.", SortOrder = 1)]
+            public string[] InputColumns;
+
+            [Argument(ArgumentType.Multiple, HelpText = "Name of the output column.", SortOrder = 2)]
+            public string[] OutputColumns;
+
+            [Argument(ArgumentType.AtMostOnce, HelpText = "GPU device id to run on (e.g. 0,1,..). Null for CPU. Requires CUDA 9.1.", SortOrder = 3)]
+            public int? GpuDeviceId = null;
+
+            [Argument(ArgumentType.AtMostOnce, HelpText = "If true, resumes execution on CPU upon GPU error. If false, will raise the GPU execption.", SortOrder = 4)]
+            public bool FallbackToCpu = false;
+        }
+
+        public sealed class OnnxTransformOutput
+        {
+            [TlcModule.Output(Desc = "ONNX transformed dataset", SortOrder = 1)]
+            public IDataView OutputData;
+
+            [TlcModule.Output(Desc = "Transform model", SortOrder = 2)]
+            public TransformModel Model;
+        }
+
+        [TlcModule.EntryPoint(Name = "Models.OnnxTransformer",
+                              Desc = "Applies an ONNX model to a dataset.",
+                              UserName = "Onnx Transformer",
+                              ShortName = "onnx-xf")]
+        public static OnnxTransformOutput ApplyOnnxModel(IHostEnvironment env, OnnxTransformInput input)
+        {
+            var host = EntryPointUtils.CheckArgsAndCreateHost(env, "OnnxTransform", input);
+
+            var inputColumns = input.InputColumns ?? (Array.Empty<string>());
+            var outputColumns = input.OutputColumns ?? (Array.Empty<string>());
+
+            var transformsCatalog = new TransformsCatalog(host);
+            var onnxScoringEstimator = OnnxCatalog.ApplyOnnxModel(transformsCatalog,
+                                                                  outputColumns,
+                                                                  inputColumns,
+                                                                  input.ModelFile,
+                                                                  input.GpuDeviceId,
+                                                                  input.FallbackToCpu);
+
+            var view = onnxScoringEstimator.Fit(input.Data).Transform(input.Data);
+            return new OnnxTransformOutput()
+            {
+                Model = new TransformModelImpl(host, view, input.Data),
+                OutputData = view
+            };
+        }
     }
 }

src/DotNetBridge/ManifestUtils.cs

@@ -43,6 +43,7 @@ public static class ManifestUtils
             typeof(ImageLoadingTransformer),
             typeof(SymbolicSgdLogisticRegressionBinaryTrainer),
             typeof(OnnxContext),
+            typeof(OnnxExportExtensions),
             typeof(SsaForecastingTransformer),
             typeof(VariableColumnTransform),
             typeof(DateTimeTransformer)

src/Platforms/build.csproj

@@ -17,6 +17,7 @@
     <PackageReference Include="Microsoft.ML.Mkl.Components" Version="1.5.0-preview*" />
     <PackageReference Include="Microsoft.ML.ImageAnalytics" Version="1.5.0-preview*" />
     <PackageReference Include="Microsoft.ML.LightGBM" Version="1.5.0-preview*" />
+    <PackageReference Include="Microsoft.ML.OnnxConverter" Version="0.17.0-preview*" />
     <PackageReference Include="Microsoft.ML.OnnxTransformer" Version="1.5.0-preview*" />
     <PackageReference Include="Microsoft.ML.TensorFlow" Version="1.5.0-preview*" />
     <PackageReference Include="Microsoft.ML.Vision" Version="1.5.0-preview*" />

src/python/nimbusml.pyproj

@@ -322,6 +322,7 @@
     <Compile Include="nimbusml\internal\core\preprocessing\filter\skipfilter.py" />
     <Compile Include="nimbusml\internal\core\preprocessing\filter\takefilter.py" />
     <Compile Include="nimbusml\internal\core\preprocessing\normalization\lpscaler.py" />
+    <Compile Include="nimbusml\internal\core\preprocessing\onnxrunner.py" />
     <Compile Include="nimbusml\internal\core\preprocessing\normalization\robustscaler.py" />
     <Compile Include="nimbusml\internal\core\preprocessing\schema\prefixcolumnconcatenator.py" />
     <Compile Include="nimbusml\internal\core\preprocessing\schema\columnduplicator.py" />
@@ -359,6 +360,8 @@
     <Compile Include="nimbusml\internal\entrypoints\models_multioutputregressionevaluator.py" />
     <Compile Include="nimbusml\internal\entrypoints\models_naivecalibrator.py" />
     <Compile Include="nimbusml\internal\entrypoints\models_oneversusall.py" />
+    <Compile Include="nimbusml\internal\entrypoints\models_onnxconverter.py" />
+    <Compile Include="nimbusml\internal\entrypoints\models_onnxtransformer.py" />
     <Compile Include="nimbusml\internal\entrypoints\models_ovamodelcombiner.py" />
     <Compile Include="nimbusml\internal\entrypoints\models_pavcalibrator.py" />
     <Compile Include="nimbusml\internal\entrypoints\models_plattcalibrator.py" />
@@ -672,6 +675,7 @@
     <Compile Include="nimbusml\preprocessing\normalization\minmaxscaler.py" />
     <Compile Include="nimbusml\preprocessing\normalization\robustscaler.py" />
     <Compile Include="nimbusml\preprocessing\normalization\__init__.py" />
+    <Compile Include="nimbusml\preprocessing\onnxrunner.py" />
     <Compile Include="nimbusml\preprocessing\schema\prefixcolumnconcatenator.py" />
     <Compile Include="nimbusml\preprocessing\schema\columnconcatenator.py" />
     <Compile Include="nimbusml\preprocessing\schema\columndropper.py" />
@@ -747,6 +751,7 @@
     <Compile Include="nimbusml\timeseries\ssaspikedetector.py" />
     <Compile Include="nimbusml\timeseries\timeseriesimputer.py" />
     <Compile Include="nimbusml\timeseries\__init__.py" />
+    <Compile Include="tests_extended\test_export_to_onnx.py" />
     <Compile Include="tests\test_estimator_checks.py" />
     <Compile Include="nimbusml\tests\feature_extraction\text\test_lightlda.py" />
     <Compile Include="nimbusml\tests\idv\test_idv.py" />

src/python/nimbusml/base_predictor.py

@@ -178,7 +178,6 @@ def summary(self):
         self.model_summary_ = pipeline.summary()
         return self.model_summary_
 
-    @trace
     def _get_implicit_transforms(
             self,
             features,
@@ -354,3 +353,20 @@ def _get_graph_nodes(
             row_group_column_name=group_id_column)
         graph_nodes['learner_node'] = [learner_node]
         return graph_nodes, learner_features
+
+    @trace
+    def export_to_onnx(self, *args, **kwargs):
+        """
+        Export the model to the ONNX format.
+
+        See :py:meth:`nimbusml.Pipeline.export_to_onnx` for accepted arguments.
+        """
+        if not hasattr(self, 'model_') \
+            or self.model_ is None \
+            or not os.path.isfile(self.model_):
+
+            raise ValueError("Model is not fitted. Train or load a model before "
+                             "export_to_onnx().")
+
+        pipeline = Pipeline([self], model=self.model_)
+        pipeline.export_to_onnx(*args, **kwargs)

src/python/nimbusml/base_transform.py

@@ -124,3 +124,20 @@ def transform(self, X, as_binary_data_stream=False, **params):
         data = pipeline.transform(
             X, as_binary_data_stream=as_binary_data_stream, **params)
         return data
+
+    @trace
+    def export_to_onnx(self, *args, **kwargs):
+        """
+        Export the model to the ONNX format.
+
+        See :py:meth:`nimbusml.Pipeline.export_to_onnx` for accepted arguments.
+        """
+        if not hasattr(self, 'model_') \
+            or self.model_ is None \
+            or not os.path.isfile(self.model_):
+
+            raise ValueError("Model is not fitted. Train or load a model before "
+                             "export_to_onnx().")
+
+        pipeline = Pipeline([self], model=self.model_)
+        pipeline.export_to_onnx(*args, **kwargs)

src/python/nimbusml/internal/core/preprocessing/onnxrunner.py

@@ -0,0 +1,71 @@
+# --------------------------------------------------------------------------------------------
+# Copyright (c) Microsoft Corporation. All rights reserved.
+# Licensed under the MIT License.
+# --------------------------------------------------------------------------------------------
+# - Generated by tools/entrypoint_compiler.py: do not edit by hand
+"""
+OnnxRunner
+"""
+
+__all__ = ["OnnxRunner"]
+
+
+from ...entrypoints.models_onnxtransformer import models_onnxtransformer
+from ...utils.utils import trace
+from ..base_pipeline_item import BasePipelineItem, DefaultSignature
+
+
+class OnnxRunner(BasePipelineItem, DefaultSignature):
+    """
+    **Description**
+        Applies an ONNX model to a dataset.
+
+    :param model_file: Path to the onnx model file.
+
+    :param input_columns: Name of the input column.
+
+    :param output_columns: Name of the output column.
+
+    :param gpu_device_id: GPU device id to run on (e.g. 0,1,..). Null for CPU.
+        Requires CUDA 9.1.
+
+    :param fallback_to_cpu: If true, resumes execution on CPU upon GPU error.
+        If false, will raise the GPU execption.
+
+    :param params: Additional arguments sent to compute engine.
+
+    """
+
+    @trace
+    def __init__(
+            self,
+            model_file,
+            input_columns=None,
+            output_columns=None,
+            gpu_device_id=None,
+            fallback_to_cpu=False,
+            **params):
+        BasePipelineItem.__init__(
+            self, type='transform', **params)
+
+        self.model_file = model_file
+        self.input_columns = input_columns
+        self.output_columns = output_columns
+        self.gpu_device_id = gpu_device_id
+        self.fallback_to_cpu = fallback_to_cpu
+
+    @property
+    def _entrypoint(self):
+        return models_onnxtransformer
+
+    @trace
+    def _get_node(self, **all_args):
+        algo_args = dict(
+            model_file=self.model_file,
+            input_columns=self.input_columns,
+            output_columns=self.output_columns,
+            gpu_device_id=self.gpu_device_id,
+            fallback_to_cpu=self.fallback_to_cpu)
+
+        all_args.update(algo_args)
+        return self._entrypoint(**all_args)

src/python/nimbusml/internal/entrypoints/models_onnxconverter.py

@@ -0,0 +1,116 @@
+# - Generated by tools/entrypoint_compiler.py: do not edit by hand
+"""
+Models.OnnxConverter
+"""
+
+
+from ..utils.entrypoints import EntryPoint
+from ..utils.utils import try_set, unlist
+
+
+def models_onnxconverter(
+        onnx,
+        data_file=None,
+        json=None,
+        name=None,
+        domain=None,
+        inputs_to_drop=None,
+        outputs_to_drop=None,
+        model=None,
+        onnx_version='Stable',
+        predictive_model=None,
+        **params):
+    """
+    **Description**
+        Converts the model to ONNX format.
+
+    :param data_file: The data file (inputs).
+    :param onnx: The path to write the output ONNX to. (inputs).
+    :param json: The path to write the output JSON to. (inputs).
+    :param name: The 'name' property in the output ONNX. By default
+        this will be the ONNX extension-less name. (inputs).
+    :param domain: The 'domain' property in the output ONNX.
+        (inputs).
+    :param inputs_to_drop: Array of input column names to drop
+        (inputs).
+    :param outputs_to_drop: Array of output column names to drop
+        (inputs).
+    :param model: Model that needs to be converted to ONNX format.
+        (inputs).
+    :param onnx_version: The targeted ONNX version. It can be either
+        "Stable" or "Experimental". If "Experimental" is used,
+        produced model can contain components that is not officially
+        supported in ONNX standard. (inputs).
+    :param predictive_model: Predictor model that needs to be
+        converted to ONNX format. (inputs).
+    """
+
+    entrypoint_name = 'Models.OnnxConverter'
+    inputs = {}
+    outputs = {}
+
+    if data_file is not None:
+        inputs['DataFile'] = try_set(
+            obj=data_file,
+            none_acceptable=True,
+            is_of_type=str)
+    if onnx is not None:
+        inputs['Onnx'] = try_set(
+            obj=onnx,
+            none_acceptable=False,
+            is_of_type=str)
+    if json is not None:
+        inputs['Json'] = try_set(
+            obj=json,
+            none_acceptable=True,
+            is_of_type=str)
+    if name is not None:
+        inputs['Name'] = try_set(
+            obj=name,
+            none_acceptable=True,
+            is_of_type=str,
+            is_column=True)
+    if domain is not None:
+        inputs['Domain'] = try_set(
+            obj=domain,
+            none_acceptable=True,
+            is_of_type=str)
+    if inputs_to_drop is not None:
+        inputs['InputsToDrop'] = try_set(
+            obj=inputs_to_drop,
+            none_acceptable=True,
+            is_of_type=list)
+    if outputs_to_drop is not None:
+        inputs['OutputsToDrop'] = try_set(
+            obj=outputs_to_drop,
+            none_acceptable=True,
+            is_of_type=list)
+    if model is not None:
+        inputs['Model'] = try_set(
+            obj=model,
+            none_acceptable=True,
+            is_of_type=str)
+    if onnx_version is not None:
+        inputs['OnnxVersion'] = try_set(
+            obj=onnx_version,
+            none_acceptable=True,
+            is_of_type=str,
+            values=[
+                'Stable',
+                'Experimental'])
+    if predictive_model is not None:
+        inputs['PredictiveModel'] = try_set(
+            obj=predictive_model, none_acceptable=True, is_of_type=str)
+
+    input_variables = {
+        x for x in unlist(inputs.values())
+        if isinstance(x, str) and x.startswith("$")}
+    output_variables = {
+        x for x in unlist(outputs.values())
+        if isinstance(x, str) and x.startswith("$")}
+
+    entrypoint = EntryPoint(
+        name=entrypoint_name, inputs=inputs, outputs=outputs,
+        input_variables=input_variables,
+        output_variables=output_variables)
+    return entrypoint