Cannot fit an Onnx Transform as part of a pipeline

[rogancarr]

It is possible to fit an ONNX model by itself:

var onnxEstimator = mlContext.Transforms.ApplyOnnxModel(modelPath)
var onnxModel = onnxEstimator.Fit(data);

But it throws when it is part of a pipeline:

var onnxEstimator = mlContext.Transforms.ApplyOnnxModel(modelPath)
    // TODO #2980: ONNX outputs don't match the outputs of the model, so we must hand-correct this for now.
    .Append(mlContext.Transforms.CopyColumns("Score", "Score0"));
    .Append(mlContext.Transforms.CopyColumns("Label", "Fable"));
    .Append(mlContext.Transforms.NormalizeLpNorm("Features2", "Features", LpNormNormalizingEstimatorBase.NormFunction.L2));
var onnxModel = onnxEstimator.Fit(data);

Any of these Append statements cause a throw. They do not affect the ONNX model at all, and use either the result of the calculation (Score0), rows unused by the model (Label) or rows also used by the transform (Features).

In this case, the error message is:
System.ArgumentOutOfRangeException : Schema mismatch for input column 'Label': expected vector, got R4 Parameter name: inputSchema

[rogancarr]

Must be related to #2982, that the type changes when passed through ONNX.

[singlis]

Based on #2982 and #2980, I was able to get this to work by specifying a custom score class:

    internal sealed class OnnxScoreColumn
    {
        [ColumnName("Score0")]
        public float[] Score { get; set; }
    }
...

var onnxPredictionEngine = mlContext.Model.CreatePredictionEngine<HousingRegression, OnnxScoreColumn>(onnxModel);

@rogancarr - I will post a PR with the changes and then close these three issues.

[singlis]

@rogancarr - sorry I misunderstood the bug. I see what you are describing and will post a fix.