Make Multiclass Linear Trainers Typed Based on Output Model Types. (#2976)

* Step 1: create two multi-class linear models

Step 2: Make SDCA trainers typed

Finish version 0.1

Delete commented lines

* Add some doc strings

More document

* Handle static extensions

* Rename several maximum entropy models and trainers

* Fix EP test

Fix two tests and address a comment

Add missing piece

* Address comments

* Improve option of MCSDCA

* Address comments

* Update code sample

* Refactorize saving family

* Rename a class following binary SDCA trainer

Author: wschin

docs/code/MlNetCookBook.md

@@ -244,7 +244,7 @@ We tried to make `Preview` debugger-friendly: our expectation is that, if you en
 Here is the code sample:
 ```csharp
 var estimator = mlContext.Transforms.Categorical.MapValueToKey("Label")
-    .Append(mlContext.MulticlassClassification.Trainers.Sdca())
+    .Append(mlContext.MulticlassClassification.Trainers.SdcaCalibrated())
     .Append(mlContext.Transforms.Conversion.MapKeyToValue("PredictedLabel"));
 
 var data = mlContext.Data.LoadFromTextFile(new TextLoader.Column[] {
@@ -423,7 +423,7 @@ var pipeline =
     // Cache data in memory for steps after the cache check point stage.
     .AppendCacheCheckpoint(mlContext)
     // Use the multi-class SDCA model to predict the label using features.
-    .Append(mlContext.MulticlassClassification.Trainers.Sdca())
+    .Append(mlContext.MulticlassClassification.Trainers.SdcaCalibrated())
     // Apply the inverse conversion from 'PredictedLabel' column back to string value.
     .Append(mlContext.Transforms.Conversion.MapKeyToValue(("PredictedLabel", "Data")));
 
@@ -547,13 +547,13 @@ var pipeline =
     // Cache data in memory for steps after the cache check point stage.
     .AppendCacheCheckpoint(mlContext)
     // Use the multi-class SDCA model to predict the label using features.
-    .Append(mlContext.MulticlassClassification.Trainers.Sdca());
+    .Append(mlContext.MulticlassClassification.Trainers.SdcaCalibrated());
 
 // Train the model.
 var trainedModel = pipeline.Fit(trainData);
 
 // Inspect the model parameters. 
-var modelParameters = trainedModel.LastTransformer.Model as MulticlassLogisticRegressionModelParameters;
+var modelParameters = trainedModel.LastTransformer.Model as MaximumEntropyModelParameters;
 
 // Now we can use 'modelParameters' to look at the weights.
 // 'weights' will be an array of weight vectors, one vector per class.
@@ -822,7 +822,7 @@ var pipeline =
     // Notice that unused part in the data may not be cached.
     .AppendCacheCheckpoint(mlContext)
     // Use the multi-class SDCA model to predict the label using features.
-    .Append(mlContext.MulticlassClassification.Trainers.Sdca());
+    .Append(mlContext.MulticlassClassification.Trainers.SdcaCalibrated());
 
 // Split the data 90:10 into train and test sets, train and evaluate.
 var split = mlContext.Data.TrainTestSplit(data, testFraction: 0.1);

docs/samples/Microsoft.ML.Samples/Dynamic/Trainers/MulticlassClassification/StochasticDualCoordinateAscent.cs

@@ -30,7 +30,7 @@ public static void Example()
                     // Convert the string labels into key types.
                     mlContext.Transforms.Conversion.MapValueToKey("Label")
                     // Apply StochasticDualCoordinateAscent multiclass trainer.
-                    .Append(mlContext.MulticlassClassification.Trainers.Sdca());
+                    .Append(mlContext.MulticlassClassification.Trainers.SdcaCalibrated());
 
             // Split the data into training and test sets. Only training set is used in fitting
             // the created pipeline. Metrics are computed on the test.

docs/samples/Microsoft.ML.Samples/Dynamic/Trainers/MulticlassClassification/StochasticDualCoordinateAscentWithOptions.cs

@@ -1,5 +1,4 @@
-using Microsoft.ML.Data;
-using Microsoft.ML.SamplesUtils;
+using Microsoft.ML.SamplesUtils;
 using Microsoft.ML.Trainers;
 
 namespace Microsoft.ML.Samples.Dynamic.Trainers.MulticlassClassification
@@ -26,10 +25,10 @@ public static void Example()
             // CC       1.216908,1.248052,1.391902,0.4326252,1.099942,0.9262842,1.334019,1.08762,0.9468155,0.4811099
             // DD       0.7871246,1.053327,0.8971719,1.588544,1.242697,1.362964,0.6303943,0.9810045,0.9431419,1.557455
 
-            var options = new SdcaMulticlassTrainer.Options
+            var options = new SdcaNonCalibratedMulticlassTrainer.Options
             {
                 // Add custom loss
-                LossFunction = new HingeLoss(),
+                Loss = new HingeLoss(),
                 // Make the convergence tolerance tighter.
                 ConvergenceTolerance = 0.05f,
                 // Increase the maximum number of passes over training data.
@@ -41,7 +40,7 @@ public static void Example()
                     // Convert the string labels into key types.
                     mlContext.Transforms.Conversion.MapValueToKey("Label")
                     // Apply StochasticDualCoordinateAscent multiclass trainer.
-                    .Append(mlContext.MulticlassClassification.Trainers.Sdca(options));
+                    .Append(mlContext.MulticlassClassification.Trainers.SdcaNonCalibrated(options));
 
             // Split the data into training and test sets. Only training set is used in fitting
             // the created pipeline. Metrics are computed on the test.

src/Microsoft.ML.Ensemble/Trainer/Multiclass/MulticlassDataPartitionEnsembleTrainer.cs

@@ -64,9 +64,9 @@ public Arguments()
                             // non-default column names. Unfortuantely no method of resolving this temporary strikes me as being any
                             // less laborious than the proper fix, which is that this "meta" component should itself be a trainer
                             // estimator, as opposed to a regular trainer.
-                            var trainerEstimator = new LogisticRegressionMulticlassClassificationTrainer(env, LabelColumnName, FeatureColumnName);
-                            return TrainerUtils.MapTrainerEstimatorToTrainer<LogisticRegressionMulticlassClassificationTrainer,
-                                MulticlassLogisticRegressionModelParameters, MulticlassLogisticRegressionModelParameters>(env, trainerEstimator);
+                            var trainerEstimator = new LbfgsMaximumEntropyTrainer(env, LabelColumnName, FeatureColumnName);
+                            return TrainerUtils.MapTrainerEstimatorToTrainer<LbfgsMaximumEntropyTrainer,
+                                MaximumEntropyModelParameters, MaximumEntropyModelParameters>(env, trainerEstimator);
                         })
                 };
             }