Infer.NET user guide : Learners : Bayes Point Machine classifiers : The Learner API

Creation & serialization

Creating a Bayes Point Machine classifier

Once you have an instance of a mapping , creating a Bayes Point Machine classifier is as easy as calling one of the following two factory methods, depending on the type of classification problem:

• BayesPointMachineClassifier.CreateBinaryClassifier (for problems with two classes)
• BayesPointMachineClassifier.CreateMulticlassClassifier (for problems with three or more classes)

Moreover, there are two versions for both methods, one which takes as input a mapping to the native data format (an implementation of IBayesPointMachineClassifierMapping) and one which takes as input a mapping to the standard data format (an implementation of IClassifierMapping). The factory methods return classifiers of type IBayesPointMachineClassifier, providing appropriate settings and means for training and prediction .

For instance, let’s imagine we have implemented a mapping from some SQL database to the native format of the Bayes Point Machine classifier, called SqlNativeMapping. To create a classifier for a multi-class problem, we then simply write

var mapping = new SqlNativeMapping();  
var classifier = BayesPointMachineClassifier.CreateMulticlassClassifier(mapping);

Creating a Bayes Point Machine classifier is a lightweight operation - no data are accessed and no computations are performed.

Serialization

Serialization is achieved through the Save extension method on the ILearner interface. There are two overloads of this method - one that takes in a file, and one that takes in a stream and a formatter. Both trained and untrained classifiers can be serialized (and deserialized).

To save a Bayes Point Machine classifier to file, for instance, we write

classifier.Save("bpm.bin");

This serializes the parameters of the classifier and the user-defined mapping. It does not serialize any training data nor any attached event handler.

Deserialization

By calling one of the static Load methods of BayesPointMachineClassifier, it is possible to load a previously serialized classifier. Just like the Save method, Load can deserialize a classifier from a file or a stream (and a formatter). Here is how to load a classifier from a file:

var classifier = BayesPointMachineClassifier.Load  
        <SqlInstanceSource,   
        Instance,   
        SqlLabelSource,   
        int,   
        Discrete,     
        BayesPointMachineClassifierTrainingSettings,  
        MulticlassBayesPointMachineClassifierPredictionSettings<int>>("bpm.bin");

The generic types that this method requires are

• TInstanceSource - the type of the instance source;
• TInstance - the type of a single instance;
• TLabelSource - the type of the label source;
• TLabel - the type of a label (in native data format bool for a binary Bayes Point Machine classifier and int for a multi-class Bayes Point Machine classifier);
• TLabelDistribution - the type of the label distribution (Bernoulli or Discrete if a native data format mapping is used);
• TTrainingSettings - the type of the classifier’s training settings (BayesPointMachineClassifierTrainingSettings in case of the Bayes Point Machine classifier);
• TPredictionSettings - the type of the classifier’s prediction settings (BinaryBayesPointMachineClassifierPredictionSettings or MulticlassBayesPointMachineClassifierPredictionSettings for the respective Bayes Point Machine classifiers).

When a classifier is deserialized the generic parameters have to exactly match the ones used to serialize the classifier. Also, during deserialization a version check will be performed. Should the version of the serialized learner not match the version of the target learner, an exception will be thrown.

Then, for convenience, a number of methods exist which are more specific than Load, and which decrease the number of generic types that need to be specified. There is, for instance,

var classifier = BayesPointMachineClassifier.LoadMulticlassClassifier  
        <SqlInstanceSource, Instance, SqlLabelSource, int, Discrete>("bpm.bin");

which equally deserializes the previously saved multi-class Bayes Point Machine classifier from file.