model.LastTransformer doesn't exist

[philiplai]

I was trying to implement Permutation Feature Importance (PFI) for Binary Classification. But I was stuck on the following line of code. This method simply doesn't exist.
// Extract the predictor.
var linearPredictor = model.LastTransformer;

I was following the example on https://docs.microsoft.com/en-us/dotnet/api/microsoft.ml.permutationfeatureimportanceextensions.permutationfeatureimportance?view=ml-dotnet

Any idea?

---

Document Details

⚠ Do not edit this section. It is required for docs.microsoft.com ➟ GitHub issue linking.

• ID: 5ec45442-8e5f-a544-cc9a-58483a383513
• Version Independent ID: 040169f9-d169-cbbd-099e-f04a7a5a6b3f
• Content: PermutationFeatureImportanceExtensions.PermutationFeatureImportance Method (Microsoft.ML)
• Content Source: dotnet/xml/Microsoft.ML/PermutationFeatureImportanceExtensions.xml
• Product: dotnet-ml-api
• GitHub Login: @natke
• Microsoft Alias: nakersha

[najeeb-kazmi]

@philiplai can you share the code where you load the data, define the pipeline, and fit the model?

LastTransformer is a property of TransformerChain. My best guess without looking at the code is that your model has only one transformer, and is not a TransformerChain. If that is the case, your model will be of type BinaryPredictionTransformer, which you can directly pass to PFI instead of the linearPredictor in the example.

[artemiusgreat]

A duplicate of this issue #3976
The problem is that model.Load by default returns incorrect type that doesn't have property LastTransformer, so you have to cast it to the right data type explicitly.

[antoniovs1029]

For the record, I wouldn't think that issue #3976 is a duplicate of this issue here. @artemiusgreat is right in that mlContext.Model.Load always return an ITransformer (link to docs). Since ITransformer doesn't have a .LastTransformer property, then it needs to be casted to a TransformerChain as shown in the following sample:

machinelearning/docs/samples/Microsoft.ML.Samples/Dynamic/Trainers/BinaryClassification/PermutationFeatureImportanceLoadFromDisk.cs

Lines 36 to 41 in 290da82

	var model = mlContext.Model.Load(modelPath, out var schema);
	// Transform the dataset.
	var transformedData = model.Transform(data);
	var linearPredictor = (model as TransformerChain<ITransformer>).LastTransformer as BinaryPredictionTransformer<CalibratedModelParametersBase<LinearBinaryModelParameters, PlattCalibrator>>;

The need for a cast to access the .LastTransformer is by design, and it isn't a bug.

On the other hand, issue #3976 was actually that attempting to do the cast above caused an exception, and so PRs #4262 and #4306 fixed this bug. The only reason #3976 is still open is because the issue specifically refers to using ML.NET on F#, and the fix made on those PRs haven't been tested on F#, only on C#.

Notice that an alternative cast that would also work would be:

var linearPredictor = (loadedmodel as TransformerChain<ITransformer>).LastTransformer as ISingleFeaturePredictionTransformer<object>;

As for the issue in here, I wouldn't know if the user who created the issue is trying to load a model (in which case, making the cast @artemiusgreat suggested might be enough), or if the user has a model with one transformer (as @najeeb-kazmi has suggested, in which case passing that predictor directly to PFI should be enough).

[najeeb-kazmi]

Closing this issue. Please feel free to reopen if something remains unanswered.