Adding KeyToValueTransformers before finishing exporting to ONNX

src/Microsoft.ML.OnnxConverter/SaveOnnxCommand.cs

@@ -15,6 +15,7 @@
 using Microsoft.ML.Internal.Utilities;
 using Microsoft.ML.Model.OnnxConverter;
 using Microsoft.ML.Runtime;
+using Microsoft.ML.Transforms;
 using Newtonsoft.Json;
 using static Microsoft.ML.Model.OnnxConverter.OnnxCSharpToProtoWrapper;
 
@@ -241,6 +242,21 @@ private static void AddSlotNames(OnnxContextImpl ctx, DataViewSchema.Column colu
             ctx.AddOutputVariable(NumberDataViewType.Int64, labelEncoderOutput);
         }
 
+        // Checks if a column has KeyValues Annotations of any type,
+        // So to know if it is safe to use KeyToValue Transformer on it.
+        private bool HasKeyValues(DataViewSchema.Column column)
+        {
+            if (column.Type.GetItemType() is KeyDataViewType keyType)
+            {
+                var metaColumn = column.Annotations.Schema.GetColumnOrNull(AnnotationUtils.Kinds.KeyValues);
+                return metaColumn != null &&
+                    metaColumn.Value.Type is VectorDataViewType vectorType &&
+                    keyType.Count == (ulong)vectorType.Size;
+            }
+
+            return false;
+        }
+
         private void Run(IChannel ch)
         {
             ILegacyDataLoader loader = null;
@@ -308,6 +324,20 @@ private void Run(IChannel ch)
                     Host.Assert(scorePipe.Source == end);
                     end = scorePipe;
                     transforms.AddLast(scoreOnnx);
+
+                    if(rawPred.PredictionKind == PredictionKind.BinaryClassification || rawPred.PredictionKind == PredictionKind.MulticlassClassification)
+                    {
+                        // Check if the PredictedLabel Column is a KeyDataViewType and has KeyValue Annotations.
+                        // If it does, add a KeyToValueMappingTransformer, to enable NimbusML to get the values back
+                        // when using an ONNX model, as described in https://github.com/dotnet/machinelearning/pull/4841
+                        var predictedLabelColumn = scorePipe.Schema.GetColumnOrNull(DefaultColumnNames.PredictedLabel);
+                        if (predictedLabelColumn.HasValue && HasKeyValues(predictedLabelColumn.Value))
+                        {
+                            var outputData = new KeyToValueMappingTransformer(Host, DefaultColumnNames.PredictedLabel).Transform(scorePipe);
+                            end = outputData;
+                            transforms.AddLast(outputData as ITransformCanSaveOnnx);
+                        }
+                    }
                 }
                 else
                 {
@@ -322,6 +352,30 @@ private void Run(IChannel ch)
                     nameof(Arguments.LoadPredictor), "We were explicitly told to load the predictor but one was not present.");
             }
 
+            // Convert back to values the KeyDataViewType "pass-through" columns
+            // (i.e those that remained untouched by the model). This is done to enable NimbusML to get these values
+            // as described in https://github.com/dotnet/machinelearning/pull/4841
+
+            var inputKeyDataViewTypeColumnsNames = new HashSet<string>();
+            foreach (var col in source.Schema)
+                if (col.IsHidden == false && HasKeyValues(col))
+                    inputKeyDataViewTypeColumnsNames.Add(col.Name);
+
+            var passThroughColumnNames = new HashSet<string>();
+            var onlyDistinctColumnNames = end.Schema.Select(c => c.Name).Distinct(); //  only check column names that appear once in the output schema
+            foreach (var col in end.Schema)
+                if (col.IsHidden == false && onlyDistinctColumnNames.Contains(col.Name) && HasKeyValues(col))
+                    passThroughColumnNames.Add(col.Name);
+
+            passThroughColumnNames.IntersectWith(inputKeyDataViewTypeColumnsNames); // Only count those columns that were in the input of the pipeline
+
+            foreach (var name in passThroughColumnNames)
+            {
+                var outputData = new KeyToValueMappingTransformer(Host, name).Transform(end);
+                end = outputData;
+                transforms.AddLast(end as ITransformCanSaveOnnx);
+            }
+
             var model = ConvertTransformListToOnnxModel(ctx, ch, source, end, transforms, _inputsToDrop, _outputsToDrop);
 
             using (var file = Host.CreateOutputFile(_outputModelPath))