Trainer Entrypoints should allow validation set/preinitialization of model states for learners that support them

[TomFinley]

The underlying trainer implementations can implement IIncrementalTrainer (for preinitializing models, or online training) or IValidatingTrainer (for models that have validation sets). As far as I see, however, no trainer actually puts optional inputs for a validation set or an incremental training in their entry-point inputs.

For example: we have this learner here.

machinelearning/src/Microsoft.ML.StandardLearners/Standard/LinearClassificationTrainer.cs

Lines 1451 to 1453 in fb06f38

	public sealed class StochasticGradientDescentClassificationTrainer :
	LinearTrainerBase<IPredictor>,
	IIncrementalTrainer<RoleMappedData, IPredictor>,

Yet, the input uses the same general type of input used by practically all the typical trainers, without any sort of initial predictor.

machinelearning/src/Microsoft.ML.StandardLearners/Standard/LinearClassificationTrainer.cs

Lines 1756 to 1757 in fb06f38

	[TlcModule.EntryPoint(Name = "Trainers.StochasticGradientDescentBinaryClassifier", Desc = "Train an Hogwild SGD binary model.", UserName = UserNameValue, ShortName = ShortName)]
	public static CommonOutputs.BinaryClassificationOutput TrainBinary(IHostEnvironment env, Arguments input)

This results in the unfortunate situation that while the underlying runtime code does implement code to enable some form of online learning, the new public API has does not actually expose that to users. See e.g., #257 for a request for this.

[TomFinley]

Since the API is no longer entry-point based this has considerably less urgency from ML.NET's perspective. Nonetheless, it may be important for entry-point based interfaces to ML.NET, e.g., NimbusML, so this is not quite ready to close.

[michaelgsharp]

The reason the leave this open was because of "entry-point based interfaces to ML.NET". Since NimbusML is no longer being supported, and the code has changed a lot since this was opened (IIncrementalTrainer doesn't exist anymore for example), I think we are good to close this issue. We can create a new one if the issue arises again, but its not an issue for ML.NET itself.

Thoughts @briacht

[briacht]

I think it's ok to close this issue, though we may want to look more at online training in the future