Support ExpandoObject dynamic members for PredictionEngine<TIn,Out> where TIn is an ExpandoObject

[IvoTops]

The PredictionEngine raises 'Type should contain a member named xyz' when using dynamic properties from an expandoobject that actually does have the member xyz.

var expando = new ExpandoObject();
// generic code to fill object with lots of properties removed
var predictionEngine = mlContext.Model.CreatePredictionEngine<ExpandoObject, ModelPrediction>(mlModel, mlSchema);
var prediction = predictionEngine.Predict(expando);

Please fix this. I train lots of models with each a different and large set of features. Trying to do this with ML.NET now. For training I have a generic solution with my own IDataView where I create dynamic getter functions and use the schemabuilder. That works to train a model and create a schema dynamically supporting thousands of features.

For prediction I have to specify my input as a type like this;

var predictionEngine = mlContext.Model.CreatePredictionEngine<REQUIREDTYPE, ModelPrediction>(mlModel, mlSchema);

I do not have type definitions with thousands of fields. It has to be generic. I create an expandoobject and add all the members and values that the engine needs, But it complains that the fields are not there -> 'Type should contain a member named xx'. So that check does not take dynamic properties into account. Support for dynamic properties (or a boolean flag to skip the type check) would open a lot of usecases for flexible featuresets.

Hope someone can fix this test to support also dynamic properties....

[masilver99]

@IvoTops, what are you doing as a work around? I'm looking at generated code, but I'm not sure there is another viable work around.

[IvoTops]

@masilver99 I have no workaround. Until this issue is resolved I will not be using ML.NET. Downloaded ML.NET codebase but complains about missing artifacts\bin\Native\x64.Debug\LdaNative.dll when I try to build so I also cannot easily fix it myself.

[masilver99]

@IvoTops, I do the same thing with IDataView. I thought it strange that the rest of ml.net doesn't support a large number of dynamic fields. I'm going to spend more time on a work around and I'll post back here if I find one. Sadly, I don't think it'll be pretty. I'm trying to resist the pull towards Python as AutoML is pretty fantastic.

[masilver99]

@IvoTops, I've found a work around. I am generating a class at runtime based on the DataViewSchema. This gives me a usable class/object with properties for every field in the schema and thus a valid PredictionEngine to use. I haven't tried populating the object, but I'm sure than can be automated. Contact me at my email address and I'll share it with you.

EDIT:
In case someone else needs a work around, here's an example: https://stackoverflow.com/questions/66893993/ml-net-create-prediction-engine-using-dynamic-class/66913705#66913705