Save ConvertTransform as ONNX Operator and Control the Use of Experimental Features with a Flag

src/Microsoft.ML.Data/DataView/RowToRowMapperTransform.cs

@@ -246,7 +246,7 @@ private static VersionInfo GetVersionInfo()
 
         public override ISchema Schema { get { return _bindings; } }
 
-        public bool CanSaveOnnx => _mapper is ICanSaveOnnx onnxMapper ? onnxMapper.CanSaveOnnx : false;
+        public bool CanSaveOnnx(OnnxContext ctx) => _mapper is ICanSaveOnnx onnxMapper ? onnxMapper.CanSaveOnnx(ctx) : false;
 
         public bool CanSavePfa => _mapper is ICanSavePfa pfaMapper ? pfaMapper.CanSavePfa : false;
 
@@ -338,7 +338,7 @@ public void SaveAsOnnx(OnnxContext ctx)
             Host.CheckValue(ctx, nameof(ctx));
             if (_mapper is ISaveAsOnnx onnx)
             {
-                Host.Check(onnx.CanSaveOnnx, "Cannot be saved as ONNX.");
+                Host.Check(onnx.CanSaveOnnx(ctx), "Cannot be saved as ONNX.");
                 onnx.SaveAsOnnx(ctx);
             }
         }

src/Microsoft.ML.Data/Model/Onnx/ICanSaveOnnx.cs

@@ -15,7 +15,7 @@ public interface ICanSaveOnnx
         /// only detectable during runtime that would prevent its being savable. (E.g.,
         /// it may wrap some other object that may or may not be savable.)
         /// </summary>
-        bool CanSaveOnnx { get; }
+        bool CanSaveOnnx(OnnxContext ctx);
     }
 
     /// <summary>

src/Microsoft.ML.Data/Model/Onnx/OnnxContext.cs

@@ -7,6 +7,8 @@
 
 namespace Microsoft.ML.Runtime.Model.Onnx
 {
+    public enum OnnxVersion { Latest, Experimental }
+
     /// <summary>
     /// A context for defining a ONNX output. The context internally contains the model-in-progress being built. This
     /// same context object is iteratively given to exportable components via the <see cref="ICanSaveOnnx"/> interface
@@ -98,5 +100,11 @@ public abstract OnnxNode CreateNode(string opType, IEnumerable<string> inputs,
         /// <returns>A node added to the in-progress ONNX graph, that attributes can be set on</returns>
         public OnnxNode CreateNode(string opType, string input, string output, string name, string domain = null)
             => CreateNode(opType, new[] { input }, new[] { output }, name, domain);
+
+        /// <summary>
+        /// Get the targeted ONNX version string. Only two values are allowed now: "latest" and "experimental".
+        /// </summary>
+        /// <returns></returns>
+        public abstract OnnxVersion GetOnnxVersion();
     }
 }

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

@@ -225,7 +225,7 @@ public abstract class ValueMapperCalibratedPredictorBase : CalibratedPredictorBa
         public ColumnType OutputType => _mapper.OutputType;
         public ColumnType DistType => NumberType.Float;
         public bool CanSavePfa => (_mapper as ICanSavePfa)?.CanSavePfa == true;
-        public bool CanSaveOnnx => (_mapper as ICanSaveOnnx)?.CanSaveOnnx == true;
+        public bool CanSaveOnnx(OnnxContext ctx) => (_mapper as ICanSaveOnnx)?.CanSaveOnnx(ctx) == true;
 
         protected ValueMapperCalibratedPredictorBase(IHostEnvironment env, string name, IPredictorProducing<Float> predictor, ICalibrator calibrator)
             : base(env, name, predictor, calibrator)
@@ -308,7 +308,7 @@ public bool SaveAsOnnx(OnnxContext ctx, string[] outputNames, string featureColu
                 return false;
 
             var calibrator = Calibrator as ISingleCanSaveOnnx;
-            if (!(calibrator?.CanSaveOnnx == true && calibrator.SaveAsOnnx(ctx, new[] { outputNames[1], outputNames[2] }, featureColumnName)))
+            if (!(calibrator?.CanSaveOnnx(ctx) == true && calibrator.SaveAsOnnx(ctx, new[] { outputNames[1], outputNames[2] }, featureColumnName)))
                 ctx.RemoveVariable(outputNames[1], true);
 
             return true;
@@ -617,7 +617,7 @@ private static VersionInfo GetVersionInfo()
         /// </summary>
         public bool CanSavePfa => (_bindable as ICanSavePfa)?.CanSavePfa == true;
 
-        public bool CanSaveOnnx => (_bindable as ICanSaveOnnx)?.CanSaveOnnx == true;
+        public bool CanSaveOnnx(OnnxContext ctx) => (_bindable as ICanSaveOnnx)?.CanSaveOnnx(ctx) == true;
 
         public SchemaBindableCalibratedPredictor(IHostEnvironment env, IPredictorProducing<Single> predictor, ICalibrator calibrator)
             : base(env, LoaderSignature, predictor, calibrator)
@@ -663,7 +663,7 @@ public bool SaveAsOnnx(OnnxContext ctx, RoleMappedSchema schema, string[] output
             Host.CheckValue(ctx, nameof(ctx));
             Host.CheckParam(Utils.Size(outputs) == 2, nameof(outputs), "Expected this to have two outputs");
             Host.CheckValue(schema, nameof(schema));
-            Host.Check(CanSaveOnnx, "Called despite not being savable");
+            Host.Check(CanSaveOnnx(ctx), "Called despite not being savable");
             return false;
         }
 
@@ -1342,7 +1342,7 @@ private static VersionInfo GetVersionInfo()
         public Double ParamA { get; }
         public Double ParamB { get; }
         public bool CanSavePfa => true;
-        public bool CanSaveOnnx => true;
+        public bool CanSaveOnnx(OnnxContext ctx) => true;
 
         public PlattCalibrator(IHostEnvironment env, Double paramA, Double paramB)
         {

src/Microsoft.ML.Data/Scorers/GenericScorer.cs

@@ -140,7 +140,7 @@ private static VersionInfo GetVersionInfo()
 
         public bool CanSavePfa => (Bindable as ICanSavePfa)?.CanSavePfa == true;
 
-        public bool CanSaveOnnx => (Bindable as ICanSaveOnnx)?.CanSaveOnnx == true;
+        public bool CanSaveOnnx(OnnxContext ctx) => (Bindable as ICanSaveOnnx)?.CanSaveOnnx(ctx) == true;
 
         /// <summary>
         /// The <see cref="SignatureDataScorer"/> entry point for creating a <see cref="GenericScorer"/>.

src/Microsoft.ML.Data/Scorers/MultiClassClassifierScorer.cs

@@ -77,7 +77,7 @@ public sealed class LabelNameBindableMapper : ISchemaBindableMapper, ICanSaveMod
 
             public VectorType Type => _type;
             public bool CanSavePfa => (_bindable as ICanSavePfa)?.CanSavePfa == true;
-            public bool CanSaveOnnx => (_bindable as ICanSaveOnnx)?.CanSaveOnnx == true;
+            public bool CanSaveOnnx(OnnxContext ctx) => (_bindable as ICanSaveOnnx)?.CanSaveOnnx(ctx) == true;
             public ISchemaBindableMapper InnerBindable => _bindable;
 
             private static VersionInfo GetVersionInfo()
@@ -207,7 +207,7 @@ public bool SaveAsOnnx(OnnxContext ctx, RoleMappedSchema schema, string[] output
             {
                 Contracts.CheckValue(ctx, nameof(ctx));
                 Contracts.CheckValue(schema, nameof(schema));
-                Contracts.Check(CanSaveOnnx, "Cannot be saved as ONNX.");
+                Contracts.Check(CanSaveOnnx(ctx), "Cannot be saved as ONNX.");
                 Contracts.Assert(_bindable is IBindableCanSaveOnnx);
                 return ((IBindableCanSaveOnnx)_bindable).SaveAsOnnx(ctx, schema, outputNames);
             }

src/Microsoft.ML.Data/Scorers/PredictedLabelScorerBase.cs

@@ -284,7 +284,7 @@ protected override BindingsBase GetBindings()
 
         public bool CanSavePfa => (Bindable as ICanSavePfa)?.CanSavePfa == true;
 
-        public bool CanSaveOnnx => (Bindable as ICanSaveOnnx)?.CanSaveOnnx == true;
+        public bool CanSaveOnnx(OnnxContext ctx) => (Bindable as ICanSaveOnnx)?.CanSaveOnnx(ctx) == true;
 
         protected PredictedLabelScorerBase(ScorerArgumentsBase args, IHostEnvironment env, IDataView data,
             ISchemaBoundMapper mapper, RoleMappedSchema trainSchema, string registrationName, string scoreColKind, string scoreColName,

src/Microsoft.ML.Data/Scorers/SchemaBindablePredictorWrapper.cs

@@ -46,7 +46,7 @@ public abstract class SchemaBindablePredictorWrapperBase : ISchemaBindableMapper
 
         public bool CanSavePfa => (ValueMapper as ICanSavePfa)?.CanSavePfa == true;
 
-        public bool CanSaveOnnx => (ValueMapper as ICanSaveOnnx)?.CanSaveOnnx == true;
+        public bool CanSaveOnnx(OnnxContext ctx) => (ValueMapper as ICanSaveOnnx)?.CanSaveOnnx(ctx) == true;
 
         public SchemaBindablePredictorWrapperBase(IPredictor predictor)
         {

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

@@ -422,7 +422,7 @@ private sealed class Mapper : IRowMapper, ISaveAsOnnx, ISaveAsPfa
             private readonly ConcatTransform _parent;
             private readonly BoundColumn[] _columns;
 
-            public bool CanSaveOnnx => true;
+            public bool CanSaveOnnx(OnnxContext ctx) => true;
             public bool CanSavePfa => true;
 
             public Mapper(ConcatTransform parent, ISchema inputSchema)
@@ -893,7 +893,7 @@ public void SaveAsPfa(BoundPfaContext ctx)
             public void SaveAsOnnx(OnnxContext ctx)
             {
                 _host.CheckValue(ctx, nameof(ctx));
-                Contracts.Assert(CanSaveOnnx);
+                Contracts.Assert(CanSaveOnnx(ctx));
 
                 string opType = "FeatureVectorizer";
                 for (int iinfo = 0; iinfo < _columns.Length; ++iinfo)

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

@@ -16,6 +16,7 @@
 using Microsoft.ML.Runtime.Data.Conversion;
 using Microsoft.ML.Runtime.Internal.Utilities;
 using Microsoft.ML.Runtime.Model;
+using Microsoft.ML.Runtime.Model.Onnx;
 using Microsoft.ML.Runtime.Command;
 using Microsoft.ML.Runtime.EntryPoints;
 
@@ -374,6 +375,30 @@ public override void Save(ModelSaveContext ctx)
             }
         }
 
+        public override bool CanSaveOnnx(OnnxContext ctx) => ctx.GetOnnxVersion() == OnnxVersion.Experimental;
+
+        protected override bool SaveAsOnnxCore(OnnxContext ctx, int iinfo, ColInfo info, string srcVariableName, string dstVariableName)
+        {
+            var opType = "CSharp";
+
+            for (int i = 0; i < _exes.Length; i++)
+            {
+                var ex = _exes[i];
+                var node = ctx.CreateNode(opType, srcVariableName, dstVariableName, ctx.GetNodeName(opType));
+                node.AddAttribute("type", LoaderSignature);
+                node.AddAttribute("to", (byte)ex.Kind);
+                if (ex.HasKeyRange)
+                {
+                    var key = ex.TypeDst.ItemType.AsKey;
+                    node.AddAttribute("min", key.Min);
+                    node.AddAttribute("max", key.Count);
+                    node.AddAttribute("contiguous", key.Contiguous);
+                }
+            }
+
+            return true;
+        }
+
         private static bool TryCreateEx(IExceptionContext ectx, ColInfo info, DataKind kind, KeyRange range, out PrimitiveType itemType, out ColInfoEx ex)
         {
             ectx.AssertValue(info);

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

@@ -609,7 +609,7 @@ private ValueGetter<VBuffer<float>> MakeGetterInd(IRow input, int iinfo)
                     };
             }
 
-            public bool CanSaveOnnx => true;
+            public bool CanSaveOnnx(OnnxContext ctx) => true;
 
             public bool CanSavePfa => true;
 

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

@@ -384,7 +384,7 @@ private AffineColumnFunction(IHost host)
             public abstract void Save(ModelSaveContext ctx);
 
             public abstract JToken PfaInfo(BoundPfaContext ctx, JToken srcToken);
-            public bool CanSaveOnnx => true;
+            public bool CanSaveOnnx(OnnxContext ctx) => true;
             public abstract bool OnnxInfo(OnnxContext ctx, OnnxNode nodeProtoWrapper, int featureCount);
 
             public abstract Delegate GetGetter(IRow input, int icol);
@@ -503,7 +503,7 @@ public JToken PfaInfo(BoundPfaContext ctx, JToken srcToken)
                 return null;
             }
 
-            public bool CanSaveOnnx => false;
+            public bool CanSaveOnnx(OnnxContext ctx) => false;
 
             public bool OnnxInfo(OnnxContext ctx, OnnxNode nodeProtoWrapper, int featureCount)
                 => throw Host.ExceptNotSupp();
@@ -636,7 +636,7 @@ public JToken PfaInfo(BoundPfaContext ctx, JToken srcToken)
                 return null;
             }
 
-            public bool CanSaveOnnx => false;
+            public bool CanSaveOnnx(OnnxContext ctx) => false;
 
             public bool OnnxInfo(OnnxContext ctx, OnnxNode nodeProtoWrapper, int featureCount)
                 => throw Host.ExceptNotSupp();

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

@@ -60,7 +60,7 @@ internal interface IColumnFunction : ICanSaveModel
 
         JToken PfaInfo(BoundPfaContext ctx, JToken srcToken);
 
-        bool CanSaveOnnx { get; }
+        bool CanSaveOnnx(OnnxContext ctx);
 
         bool OnnxInfo(OnnxContext ctx, OnnxNode nodeProtoWrapper, int featureCount);
     }

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

@@ -452,7 +452,7 @@ private sealed class Mapper : MapperBase, ISaveAsOnnx, ISaveAsPfa
         {
             private NormalizerTransformer _parent;
 
-            public bool CanSaveOnnx => true;
+            public bool CanSaveOnnx(OnnxContext ctx) => true;
             public bool CanSavePfa => true;
 
             public Mapper(NormalizerTransformer parent, ISchema schema)
@@ -562,12 +562,12 @@ private bool SaveAsOnnxCore(OnnxContext ctx, int iinfo, ColumnInfo info, string
                 Contracts.AssertValue(ctx);
                 Contracts.Assert(0 <= iinfo && iinfo < _parent._columns.Length);
                 Contracts.Assert(_parent._columns[iinfo] == info);
-                Contracts.Assert(CanSaveOnnx);
+                Contracts.Assert(CanSaveOnnx(ctx));
 
                 if (info.InputType.ValueCount == 0)
                     return false;
 
-                if (info.ColumnFunction.CanSaveOnnx)
+                if (info.ColumnFunction.CanSaveOnnx(ctx))
                 {
                     string opType = "Scaler";
                     var node = ctx.CreateNode(opType, srcVariableName, dstVariableName, ctx.GetNodeName(opType));

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

@@ -721,7 +721,7 @@ private sealed class Mapper : MapperBase, ISaveAsOnnx, ISaveAsPfa
 
             private readonly BoundTermMap[] _termMap;
 
-            public bool CanSaveOnnx => true;
+            public bool CanSaveOnnx(OnnxContext ctx) => true;
 
             public bool CanSavePfa => true;
 

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

@@ -469,7 +469,7 @@ private sealed class ColumnTmp : OneToOneColumn
 
         public virtual bool CanSavePfa => false;
 
-        public virtual bool CanSaveOnnx => false;
+        public virtual bool CanSaveOnnx(OnnxContext ctx) => false;
 
         protected OneToOneTransformBase(IHostEnvironment env, string name, OneToOneColumn[] column,
             IDataView input, Func<ColumnType, string> testType)
@@ -574,7 +574,7 @@ public void SaveAsPfa(BoundPfaContext ctx)
         public void SaveAsOnnx(OnnxContext ctx)
         {
             Host.CheckValue(ctx, nameof(ctx));
-            Host.Assert(CanSaveOnnx);
+            Host.Assert(CanSaveOnnx(ctx));
 
             for (int iinfo = 0; iinfo < Infos.Length; ++iinfo)
             {

src/Microsoft.ML.FastTree/FastTree.cs

@@ -2860,7 +2860,7 @@ public abstract class FastTreePredictionWrapper :
         public ColumnType InputType { get; }
         public ColumnType OutputType => NumberType.Float;
         public bool CanSavePfa => true;
-        public bool CanSaveOnnx => true;
+        public bool CanSaveOnnx(OnnxContext ctx) => true;
 
         protected FastTreePredictionWrapper(IHostEnvironment env, string name, Ensemble trainedEnsemble, int numFeatures, string innerArgs)
             : base(env, name)

src/Microsoft.ML.Legacy/CSharpApi.cs

@@ -3266,6 +3266,12 @@ public OneVersusAllPipelineStep(Output output)
 
     namespace Legacy.Models
     {
+        public enum OnnxVersion
+        {
+            Latest = 0,
+            Experimental = 1
+        }
+
 
         /// <summary>
         /// Converts the model to ONNX format.
@@ -3309,6 +3315,11 @@ public sealed partial class OnnxConverter
             /// </summary>
             public Var<Microsoft.ML.Runtime.EntryPoints.ITransformModel> Model { get; set; } = new Var<Microsoft.ML.Runtime.EntryPoints.ITransformModel>();
 
+            /// <summary>
+            /// The targeted ONNX version. It can be either "latest" or "experimental"
+            /// </summary>
+            public OnnxVersion OnnxVersion { get; set; } = OnnxVersion.Latest;
+
             /// <summary>
             /// The data file
             /// </summary>

src/Microsoft.ML.Onnx/OnnxContextImpl.cs

@@ -31,9 +31,10 @@ internal sealed class OnnxContextImpl : OnnxContext
         private readonly string _domain;
         private readonly string _producerVersion;
         private readonly long _modelVersion;
+        private readonly OnnxVersion _onnxVersion;
 
         public OnnxContextImpl(IHostEnvironment env, string name, string producerName,
-            string producerVersion, long modelVersion, string domain)
+            string producerVersion, long modelVersion, string domain, OnnxVersion onnxVersion)
         {
             Contracts.CheckValue(env, nameof(env));
             _host = env.Register(nameof(OnnxContext));
@@ -52,6 +53,7 @@ public OnnxContextImpl(IHostEnvironment env, string name, string producerName,
             _producerVersion = producerVersion;
             _modelVersion = modelVersion;
             _domain = domain;
+            _onnxVersion = onnxVersion;
         }
 
         public override bool ContainsColumn(string colName) => _columnNameMap.ContainsKey(colName);
@@ -251,5 +253,7 @@ public void AddInputVariable(ColumnType type, string colName)
         /// </summary>
         public ModelProto MakeModel()
             => OnnxUtils.MakeModel(_nodes, _producerName, _name, _domain, _producerVersion, _modelVersion, _inputs, _outputs, _intermediateValues);
+
+        public override OnnxVersion GetOnnxVersion() => _onnxVersion;
     }
 }

src/Microsoft.ML.Onnx/OnnxUtils.cs

@@ -308,14 +308,38 @@ public static ModelArgs GetModelArgs(ColumnType type, string colName,
                 case DataKind.TX:
                     dataType = TensorProto.Types.DataType.String;
                     break;
+                case DataKind.I1:
+                    dataType = TensorProto.Types.DataType.Int8;
+                    break;
+                case DataKind.U1:
+                    dataType = TensorProto.Types.DataType.Uint8;
+                    break;
+                case DataKind.I2:
+                    dataType = TensorProto.Types.DataType.Int16;
+                    break;
+                case DataKind.U2:
+                    dataType = TensorProto.Types.DataType.Uint16;
+                    break;
+                case DataKind.I4:
+                    dataType = TensorProto.Types.DataType.Int32;
+                    break;
                 case DataKind.U4:
                     dataType = TensorProto.Types.DataType.Int64;
                     break;
+                case DataKind.I8:
+                    dataType = TensorProto.Types.DataType.Int64;
+                    break;
+                case DataKind.U8:
+                    dataType = TensorProto.Types.DataType.Uint64;
+                    break;
                 case DataKind.R4:
                     dataType = TensorProto.Types.DataType.Float;
                     break;
+                case DataKind.R8:
+                    dataType = TensorProto.Types.DataType.Double;
+                    break;
                 default:
-                    Contracts.Assert(false, "Unknown type.");
+                    Contracts.Assert(false, "Unsupported type: DataKind " + rawKind.ToString());
                     break;
             }