在 ML.NET 中执行 ITransformer.Transform 后从 IDataView 中提取 MultiClass 结果

Question

i am trying to use ML.NET generically without having to create a class as input and output of a model. To do that, after creating a model with this:

        public static (ITransformer model, double accuracy) TrainMultiClassModel(MulticlassExperimentSettings experimentSettings, MLContext mlContext, IDataView myview, string LabelName)
    {
        ITransformer trainedModel;
        MulticlassClassificationExperiment experiment = mlContext.Auto().CreateMulticlassClassificationExperiment(experimentSettings);

        ExperimentResult<MulticlassClassificationMetrics> experimentResult = experiment.Execute(myview, LabelName);
        RunDetail<MulticlassClassificationMetrics> best = experimentResult.BestRun;

        trainedModel = best.Model;

        return (trainedModel, best.ValidationMetrics.MacroAccuracy);
    }

Where the myView contains a CSV file with correctly set DataKinds.

Example of the Data:

Then i execute that model by running something like this:

            MemoryStream modelStream = new MemoryStream(ModelData);
            ITransformer trainedModel = mlContext.Model.Load(modelStream, out var modelInputSchema);
            var predictions = trainedModel.Transform(myview);

Again, the myView contains Data from a CSV file, just with the predicted column empty.

Now we have the "predictions", which is of type IDataView.

For regression results, thats easy. Look for the Schema named "Score" and load it as float:

float[] scoreColumn = predictions.GetColumn<float>("Score").ToArray();

But how does it work for MultiClass experiments? There is a Schema called "PredictedLabel" of type "String", but it contains numbers between 0 and 1 when read like this:

var labelColumn = predictions.Schema.FirstOrDefault(s => s.Name == "PredictedLabel" && s.IsHidden == false);
string[] scoreColumn = predictions.GetColumn<string>(labelColumn).ToArray();

How do i get the actual names of the (in this case) Species? Or do i have to map the numbers to the name somehow? Which mapping table do i use for that?

Thank you in advance.

edit: the code by Eric gave this list:

1.4
1.9
0.2
0.4
 0.3
0.1
0.5
0.6
1.5
1.3
1.6
1.0
1.1
1.8
1.2
1.7
2.5
2.1
2.2
2.0
2.4
2.3

Those are 22, which is weired: none of the correct species does have 22 characters (in case thats the chars from a name), and i did input just 4 rows of data to solve. The "PredictedLabel" does meanwhile output 4 values, but which are still numbers:

But now i am wondering: how do i read this field? Maybe it contains the answer:

Answer 1

What you want to use is a method called GetKeyValues . This will give you a VBuffer<ReadOnlyMemory<char>> , where each string in the buffer is the "value" for the corresponding index into the "keys" or "classes" in your multi-class classification model.

var predictions = trainedModel.Transform(myview);

var labelColumn = predictions.Schema[labelName]; // this is "Species" in your example above

VBuffer<ReadOnlyMemory<char>> keys = default;
labelColumn.GetKeyValues(ref keys);

foreach (var key in keys.DenseValues())
{
    Console.WriteLine(key);
}

Answer 2

For the sake of completeness, and to resolve the confusion in the initial question, here is the answer.

First, why did "PredictedLabel" and "GetKeyValues" (from Erics answer) not provide usable results? The problem here was the way i used the IDataView. When training the data i loaded the whole CSV, including the "ID" column, which i did not provide while executing the model (as that column is not valuable to use the model). after switching to always omit the "ID" column and have the same CSV layout in training and execution, both the approach from Eric and mine started working.

So when you want to interpret your results into the correct formats, first see if a "PredictedLabel" exists in the resulting DataView Schema.

if (predictions.Schema.Any(s => s.Name == "PredictedLabel"))

If it does, check its DataType. That is how you can differentiate between MultiClass and Binary results:

    var labelColumn = predictions.Schema.FirstOrDefault(s => s.Name == "PredictedLabel" && s.IsHidden == false);            
    if (labelColumn.Type.ToString() == "Boolean")
{
    bool[] binaryResults = predictions.GetColumn<bool>(labelColumn).ToArray();
}

(or)

    if (labelColumn.Type.ToString() == "String")
{
    string[] multiclassResults = predictions.GetColumn<string>(labelColumn).ToArray();
}

The multiclassResults will now contain your written down MultiClass results as string.

In case there is no PredictedLabel, there should be a "Score" Schema, which contains your regression results:

float[] regressionResults = predictions.GetColumn<float>("Score").ToArray();