Added onnx export support for SlotsDroppingTransformer

src/Microsoft.ML.Data/Prediction/Calibrator.cs

@@ -1742,12 +1742,17 @@ bool ISingleCanSaveOnnx.SaveAsOnnx(OnnxContext ctx, string[] scoreProbablityColu
             _host.CheckValue(scoreProbablityColumnNames, nameof(scoreProbablityColumnNames));
             _host.Check(Utils.Size(scoreProbablityColumnNames) == 2);
 
-            string opType = "Affine";
-            string linearOutput = ctx.AddIntermediateVariable(null, "linearOutput", true);
-            var node = ctx.CreateNode(opType, new[] { scoreProbablityColumnNames[0] },
-                new[] { linearOutput }, ctx.GetNodeName(opType), "");
-            node.AddAttribute("alpha", Slope * -1);
-            node.AddAttribute("beta", -0.0000001);
+            // The Affine operator is no longer supported in the v11 opset.
+            // So we have to decompose it using Mul and Add
+            string opType = "Mul";
+            var slopVar = ctx.AddInitializer((float)(-Slope), "Slope");
+            var mulNodeOutput = ctx.AddIntermediateVariable(null, "MulNodeOutput", true);
+            var node = ctx.CreateNode(opType, new[] { scoreProbablityColumnNames[0], slopVar }, new[] { mulNodeOutput }, ctx.GetNodeName(opType), "");
+
+            opType = "Add";
+            var betaVar = ctx.AddInitializer(-0.0000001f, "Slope");
+            var linearOutput = ctx.AddIntermediateVariable(null, "linearOutput", true);
+            node = ctx.CreateNode(opType, new[] { mulNodeOutput, betaVar }, new[] { linearOutput }, ctx.GetNodeName(opType), "");
 
             opType = "Sigmoid";
             node = ctx.CreateNode(opType, new[] { linearOutput },

src/Microsoft.ML.Data/Transforms/SlotsDroppingTransformer.cs

@@ -12,6 +12,7 @@
 using Microsoft.ML.Data;
 using Microsoft.ML.Internal.Internallearn;
 using Microsoft.ML.Internal.Utilities;
+using Microsoft.ML.Model.OnnxConverter;
 using Microsoft.ML.Runtime;
 using Microsoft.ML.Transforms;
 
@@ -442,11 +443,12 @@ private static bool AreRangesValid(int[][] slotsMin, int[][] slotsMax)
         private protected override IRowMapper MakeRowMapper(DataViewSchema schema)
             => new Mapper(this, schema);
 
-        private sealed class Mapper : OneToOneMapperBase
+        private sealed class Mapper : OneToOneMapperBase, ISaveAsOnnx
         {
             private readonly SlotsDroppingTransformer _parent;
             private readonly int[] _cols;
             private readonly DataViewType[] _srcTypes;
+            private readonly DataViewType[] _rawTypes;
             private readonly DataViewType[] _dstTypes;
             private readonly SlotDropper[] _slotDropper;
             // Track if all the slots of the column are to be dropped.
@@ -459,6 +461,7 @@ public Mapper(SlotsDroppingTransformer parent, DataViewSchema inputSchema)
                 _parent = parent;
                 _cols = new int[_parent.ColumnPairs.Length];
                 _srcTypes = new DataViewType[_parent.ColumnPairs.Length];
+                _rawTypes = new DataViewType[_parent.ColumnPairs.Length];
                 _dstTypes = new DataViewType[_parent.ColumnPairs.Length];
                 _slotDropper = new SlotDropper[_parent.ColumnPairs.Length];
                 _suppressed = new bool[_parent.ColumnPairs.Length];
@@ -471,8 +474,8 @@ public Mapper(SlotsDroppingTransformer parent, DataViewSchema inputSchema)
                     _srcTypes[i] = inputSchema[_cols[i]].Type;
                     VectorDataViewType srcVectorType = _srcTypes[i] as VectorDataViewType;
 
-                    DataViewType itemType = srcVectorType?.ItemType ?? _srcTypes[i];
-                    if (!IsValidColumnType(itemType))
+                    _rawTypes[i] = srcVectorType?.ItemType ?? _srcTypes[i];
+                    if (!IsValidColumnType(_rawTypes[i]))
                         throw Host.ExceptSchemaMismatch(nameof(inputSchema), "input", _parent.ColumnPairs[i].inputColumnName);
 
                     int valueCount = srcVectorType?.Size ?? 1;
@@ -868,6 +871,57 @@ protected override DataViewSchema.DetachedColumn[] GetOutputColumnsCore()
                 }
                 return result;
             }
+
+            public bool CanSaveOnnx(OnnxContext ctx) => true;
+
+            public void SaveAsOnnx(OnnxContext ctx)
+            {
+                Host.CheckValue(ctx, nameof(ctx));
+
+                for (int iinfo = 0; iinfo < _cols.Length; ++iinfo)
+                {
+                    string inputColumnName = _parent.ColumnPairs[iinfo].inputColumnName;
+                    if (!ctx.ContainsColumn(inputColumnName))
+                        continue;
+
+                    string srcVariableName = ctx.GetVariableName(inputColumnName);
+                    string dstVariableName = ctx.AddIntermediateVariable(_dstTypes[iinfo], _parent.ColumnPairs[iinfo].outputColumnName);
+                    if (!SaveAsOnnxCore(ctx, iinfo, srcVariableName, dstVariableName))
+                        ctx.RemoveColumn(dstVariableName);
+                }
+            }
+
+            public bool SaveAsOnnxCore(OnnxContext ctx, int iinfo, string srcVariableName, string dstVariableName)
+            {
+                string opType;
+                if (_srcTypes[iinfo] is VectorDataViewType)
+                {
+                    opType = "GatherElements";
+                    IEnumerable<long> slots = _slotDropper[iinfo].GetPreservedSlots();
+                    var slotsVar = ctx.AddInitializer(slots, new long[] { 1, slots.Count() }, "PreservedSlots");
+                    var node = ctx.CreateNode(opType, new[] { srcVariableName, slotsVar }, new[] { dstVariableName }, ctx.GetNodeName(opType), "");
+                    node.AddAttribute("axis", 1);
+                }
+                else
+                {
+                    string constVal;
+                    long[] dims = { 1, 1 };
+                    float[] floatVals = { 0.0f };
+                    long[] keyVals = { 0 };
+                    string[] stringVals = { "" };
+                    if (_rawTypes[iinfo] is TextDataViewType)
+                        constVal = ctx.AddInitializer(stringVals, dims);
+                    else if (_rawTypes[iinfo] is KeyDataViewType)
+                        constVal = ctx.AddInitializer(keyVals, dims);
+                    else
+                        constVal = ctx.AddInitializer(floatVals, dims);
+
+                    opType = "Identity";
+                    ctx.CreateNode(opType, constVal, dstVariableName, ctx.GetNodeName(opType), "");
+                }
+                return true;
+            }
+
         }
     }
 }

src/Microsoft.ML.Data/Utilities/SlotDropper.cs

@@ -3,6 +3,8 @@
 // See the LICENSE file in the project root for more information.
 
 using System;
+using System.Collections.Generic;
+using System.Linq;
 using Microsoft.ML.Data;
 using Microsoft.ML.Internal.Utilities;
 using Microsoft.ML.Runtime;
@@ -16,6 +18,7 @@ namespace Microsoft.ML.Internal.Internallearn
     internal sealed class SlotDropper
     {
         private readonly int[] _lengthReduction;
+        private readonly int _srcLength;
 
         /// <summary>
         /// Returns -1 for non vector and unknown length vectors.
@@ -43,6 +46,7 @@ public SlotDropper(int srcLength, int[] slotsMin, int[] slotsMax)
 
             SlotsMin = slotsMin;
             SlotsMax = slotsMax;
+            _srcLength = srcLength;
             _lengthReduction = ComputeLengthReduction();
 
             Contracts.Check(SlotsMin.Length == _lengthReduction.Length);
@@ -212,5 +216,16 @@ public void DropSlots<TDst>(ref VBuffer<TDst> src, ref VBuffer<TDst> dst)
 
             dst = editor.CommitTruncated(iiDst);
         }
+
+        public IEnumerable<long> GetPreservedSlots()
+        {
+            var slots = Enumerable.Range(0, _srcLength);
+            var droppedSlots = Enumerable.Range(SlotsMin[0], SlotsMax[0] - SlotsMin[0] + 1);
+            for (int i = 1; i < SlotsMin.Length; i++)
+            {
+                droppedSlots = droppedSlots.Concat(Enumerable.Range(SlotsMin[i], SlotsMax[i] - SlotsMin[i] + 1));
+            }
+            return slots.Except(droppedSlots).Select(i=>(long)i);
+        }
     }
 }

src/Microsoft.ML.OnnxConverter/OnnxUtils.cs

@@ -305,7 +305,7 @@ public static ModelProto MakeModel(List<NodeProto> nodes, string producerName, s
             model.IrVersion = (long)OnnxCSharpToProtoWrapper.Version.IrVersion;
             model.ModelVersion = modelVersion;
             model.OpsetImport.Add(new OperatorSetIdProto() { Domain = "ai.onnx.ml", Version = 2 });
-            model.OpsetImport.Add(new OperatorSetIdProto() { Domain = "", Version = 9 });
+            model.OpsetImport.Add(new OperatorSetIdProto() { Domain = "", Version = 11 });
             model.Graph = new GraphProto();
             var graph = model.Graph;
             graph.Node.Add(nodes);

src/Microsoft.ML.StandardTrainers/Standard/LogisticRegression/MulticlassLogisticRegression.cs

@@ -1004,10 +1004,18 @@ private bool SaveAsOnnxCore(OnnxContext ctx, string[] outputs, string featureCol
 
             // Onnx outputs an Int64, but ML.NET outputs UInt32. So cast the Onnx output here
             opType = "Cast";
-            var castNode = ctx.CreateNode(opType, predictedLabelInt64, predictedLabelUint32, ctx.GetNodeName(opType), "");
+            var castNodeOutput = ctx.AddIntermediateVariable(NumberDataViewType.UInt32, "CastNodeOutput", true);
+            var castNode = ctx.CreateNode(opType, predictedLabelInt64, castNodeOutput, ctx.GetNodeName(opType), "");
             var t = InternalDataKindExtensions.ToInternalDataKind(DataKind.UInt32).ToType();
             castNode.AddAttribute("to", t);
 
+            // The predictedLabel is a scalar. But the onnx output of ML.NET output expects a [1x1] tensor for output. So reshape it here
+            opType = "Reshape";
+            long[] shape = { 1, 1 };
+            long[] shapeDim = { 2 };
+            var shapeVar = ctx.AddInitializer(shape, shapeDim, "ShapeVar");
+            var reshapeNode = ctx.CreateNode(opType, new[] { castNodeOutput, shapeVar }, new[] { predictedLabelUint32 }, ctx.GetNodeName(opType), "");
+
             return true;
         }
 

src/Microsoft.ML.StandardTrainers/Standard/MulticlassClassification/OneVersusAllTrainer.cs

@@ -559,8 +559,8 @@ public string[] SaveAsOnnxPreProcess(OnnxContext ctx, string featureColumn, bool
                         outputs[i] = clipOutput;
 
                         string opType = "Clip";
-                        var clipNode = ctx.CreateNode(opType, clipInput, outputs[i], ctx.GetNodeName(opType), "");
-                        clipNode.AddAttribute("min", 0.0);
+                        var zeroVar = ctx.AddInitializer(0.0f, "Zero");
+                        var clipNode = ctx.CreateNode(opType, new[] { clipInput, zeroVar }, new[] { outputs[i] }, ctx.GetNodeName(opType), "");
                     }
                     else
                         outputs[i] = predictorOutputNames[2];

src/Microsoft.ML.Transforms/CountFeatureSelection.cs

@@ -30,7 +30,7 @@ namespace Microsoft.ML.Transforms
     /// | Does this estimator need to look at the data to train its parameters? | Yes |
     /// | Input column data type | Vector or scalar of numeric, [text](xref:Microsoft.ML.Data.TextDataViewType) or [key](xref:Microsoft.ML.Data.KeyDataViewType) data types|
     /// | Output column data type | Same as the input column|
-    /// | Exportable to ONNX | No |
+    /// | Exportable to ONNX | Yes |
     ///
     /// This transform uses a set of aggregators to count the number of values for each slot (vector element)
     /// that are non-default and non-missing (for the definitions of default and missing, refer to the remarks section

src/Microsoft.ML.Transforms/MutualInformationFeatureSelection.cs

@@ -32,7 +32,7 @@ namespace Microsoft.ML.Transforms
     /// | Does this estimator need to look at the data to train its parameters? | Yes |
     /// | Input column data type | Vector or scalar of numeric, [text](xref:Microsoft.ML.Data.TextDataViewType) or [key](xref:Microsoft.ML.Data.KeyDataViewType) data types|
     /// | Output column data type | Same as the input column|
-    /// | Exportable to ONNX | No |
+    /// | Exportable to ONNX | Yes |
     ///
     /// Formally, the mutual information can be written as:
     ///

test/BaselineOutput/Common/Onnx/BinaryClassification/BreastCancer/ExcludeVariablesInOnnxConversion.txt

@@ -381,25 +381,25 @@
       },
       {
         "input": [
-          "Score"
+          "Score",
+          "Slope"
+        ],
+        "output": [
+          "MulNodeOutput"
+        ],
+        "name": "Mul",
+        "opType": "Mul"
+      },
+      {
+        "input": [
+          "MulNodeOutput",
+          "Slope0"
         ],
         "output": [
           "linearOutput"
         ],
-        "name": "Affine",
-        "opType": "Affine",
-        "attribute": [
-          {
-            "name": "alpha",
-            "f": 0.4,
-            "type": "FLOAT"
-          },
-          {
-            "name": "beta",
-            "f": -1E-07,
-            "type": "FLOAT"
-          }
-        ]
+        "name": "Add",
+        "opType": "Add"
       },
       {
         "input": [
@@ -478,6 +478,22 @@
       }
     ],
     "name": "A Simple Pipeline",
+    "initializer": [
+      {
+        "dataType": 1,
+        "floatData": [
+          0.4
+        ],
+        "name": "Slope"
+      },
+      {
+        "dataType": 1,
+        "floatData": [
+          -1E-07
+        ],
+        "name": "Slope0"
+      }
+    ],
     "input": [
       {
         "name": "F1",
@@ -671,7 +687,7 @@
       "version": "2"
     },
     {
-      "version": "9"
+      "version": "11"
     }
   ]
 }

test/BaselineOutput/Common/Onnx/BinaryClassification/BreastCancer/LightGbmBinaryClassificationOnnxConversionTest.txt

@@ -526,7 +526,7 @@
       "version": "2"
     },
     {
-      "version": "9"
+      "version": "11"
     }
   ]
 }

test/BaselineOutput/Common/Onnx/BinaryClassification/BreastCancer/LogisticRegressionSaveModelToOnnxTest.txt

@@ -270,7 +270,7 @@
       "version": "2"
     },
     {
-      "version": "9"
+      "version": "11"
     }
   ]
 }

test/BaselineOutput/Common/Onnx/BinaryClassification/BreastCancer/ModelWithLessIO.txt

@@ -707,25 +707,25 @@
       },
       {
         "input": [
-          "Score"
+          "Score",
+          "Slope"
+        ],
+        "output": [
+          "MulNodeOutput"
+        ],
+        "name": "Mul",
+        "opType": "Mul"
+      },
+      {
+        "input": [
+          "MulNodeOutput",
+          "Slope0"
         ],
         "output": [
           "linearOutput"
         ],
-        "name": "Affine",
-        "opType": "Affine",
-        "attribute": [
-          {
-            "name": "alpha",
-            "f": 0.4,
-            "type": "FLOAT"
-          },
-          {
-            "name": "beta",
-            "f": -1E-07,
-            "type": "FLOAT"
-          }
-        ]
+        "name": "Add",
+        "opType": "Add"
       },
       {
         "input": [
@@ -804,6 +804,22 @@
       }
     ],
     "name": "modelWithLessIO",
+    "initializer": [
+      {
+        "dataType": 1,
+        "floatData": [
+          0.4
+        ],
+        "name": "Slope"
+      },
+      {
+        "dataType": 1,
+        "floatData": [
+          -1E-07
+        ],
+        "name": "Slope0"
+      }
+    ],
     "input": [
       {
         "name": "F1",
@@ -961,7 +977,7 @@
       "version": "2"
     },
     {
-      "version": "9"
+      "version": "11"
     }
   ]
 }