Help with Custom VarVector to Vector mapping (does it have to so hard?)

[acrigney]

System information

• OS version/distro: Windows 10
• .NET Version (eg., dotnet --info): ML.NET v4.0.30319

Issue

• What did you do?
• I have a data class that contains single fields and fields that are arrays
• What happened?
• How to custom map the the varvector generated into a single vector
• similar to https://github.com/dotnet/machinelearning/issues/4977
• What did you expect?
• Easy way to convert the varvectors into a single vector

Source code / logs

// Code abbreviated

MLContext mlContext = new MLContext();

SchemaDefinition definedSchema;            

// For fixed array dimension sizes use
definedSchema = SchemaDefinition.Create(typeof(MLDataForAnalysisFactored));

// If we can use variable array dimension sizes we can use this
int featuresCount = SetFeatureArrayDimensions(mlDataList, out definedSchema);

IDataView trainDataView = mlContext.Data.LoadFromEnumerable<MLDataForAnalysisFactored>(mlDataList, definedSchema);

// The label must not be in the input features
string[] features = typeof(MLDataForAnalysisFactored).GetProperties().ToList().
        Where(p => p.Name != labelColumnName).Select(x => x.Name).ToArray();

var estimatorChain = mlContext.Transforms.Conversion.MapValueToKey(outputColumnName: "Label", inputColumnName: 
    nameof(MLDataForAnalysisFactored.Call))
    .Append(mlContext.Transforms.Concatenate("Features", features));

var transformedTrainData = estimatorChain.Fit(trainDataView);

trainDataView = transformedTrainData.Transform(trainDataView);

// check the data
var rowEnumerable = mlContext.Data
    .CreateEnumerable<MLDataForAnalysisFactored>(trainDataView,
    reuseRowObject: true).ToList();

var schema = trainDataView.Schema;

//var featureColumns = trainDataView.GetColumn<float[]>(trainDataView.Schema["Features"]).Take(4);

//create a custom schema-definition that overrides the type for the Values field...  
var singleVectorSchemaDef = SchemaDefinition.Create(typeof(SingleVector));
singleVectorSchemaDef[nameof(SingleVector.Values)].ColumnType
              = new VectorDataViewType(NumberDataViewType.Single, featuresCount);

// How to do a custom transform to map the varvectors to a single vector?
// otherwise you have to do this.
////use that schema definition when creating the training dataview  
//trainDataView = mlContext.Data.LoadFromEnumerable(mlDataList, singleVectorSchemaDef);

// check the data

//        var someRows = mlContext.Data.  // Convert to an enumerable of user-defined type. 
//                .CreateEnumerable<MLDataForAnalysisFactored>(trainDataView, reuseRowObject: false)
//// Take a couple values as an array.
//.Take(4).ToArray();

// Extract the 'AllFeatures' column.
// This will give the entire dataset: make sure to only take several row
// in case the dataset is huge. The is similar to the static API, except
// you have to specify the column name and type.

//var featureColumns = estimatorChain.GetColumn<float[]>(trainDataView.Schema["Features"]);

// STEP 2: Run AutoML experiment
Console.WriteLine($"Running AutoML multiclass classification experiment for {ExperimentTime} seconds...");            

ExperimentResult<MulticlassClassificationMetrics> experimentResult = mlContext.Auto()
    .CreateMulticlassClassificationExperiment(ExperimentTime)
    .Execute(trainDataView, LabelColumnName, null, estimatorChain);

where SingleVector is

  public class SingleVector
  {
      //it's not required to specify the type here since we will override in our custom schema 
      public float[] Values;
  }
  // Above function SetFeatureArrayDimensions
  // You would expect that once the size of the var vectors has been set then that is all you should have to do but ML.NET needs
  // a single vector.
  private int SetFeatureArrayDimensions(List<MLDataForAnalysisFactored> mlDataList, out SchemaDefinition definedSchema)
      {
          int totalFeatureCount = 0;
          // STEP 1: Load data
  
          // The feature dimension (typically this will be the Count of the array 
          // of the features vector known at runtime).
          int featureArrayDimension = 0;
          definedSchema = SchemaDefinition.Create(typeof(MLDataForAnalysisFactored));
  
          var properties = typeof(MLDataForAnalysisFactored).GetProperties();
  
          foreach (var property in properties)
          {
              if (property.PropertyType.IsArray)
              {
                  Array array = property.GetValue(mlDataList[0]) as Array;
                  featureArrayDimension = array.Length;
                  totalFeatureCount += featureArrayDimension;
  
                  //// Set the column type to be a known-size vector.
                  var vectorItemType = ((VectorDataViewType)definedSchema[property.Name].ColumnType)
                              .ItemType;
  
                  definedSchema[property.Name].ColumnType = new VectorDataViewType(vectorItemType,
                      featureArrayDimension);
  
                  var featureColumn = definedSchema[property.Name]
                  .ColumnType as VectorDataViewType;
  
                  Diag.Debug.WriteLine($"Is the size of the Feature array {property.Name} column known: " +
                      $"{featureColumn.IsKnownSize}.\nSize: {featureColumn.Size}");
              }
              else
              {
                  totalFeatureCount++;
              }
          }
          return (totalFeatureCount);
      }        

Anyway I can get away with using fixed array sizes. Please paste or attach the code or logs or traces that would be helpful to diagnose the issue you are reporting.