Trainers offer one of the simplest interfaces in Shark, but also form a very diverse group of algorithms for a range of different settings and applications. Each trainer represents a one-step solver for one certain machine learning problem. Usually, it will adapt the parameters of a model such that it represents the solution to some machine learning or data processing problem given a certain data set.

Some trainers are very simple – for example those training a linear model such that it normalizes the components of all examples in a data set to unit variance. Other trainers are quite complex, for example for some multi-class support vector machines. But in most cases, trainers are used to reach analytical solutions to relatively simple problems which are not stated as an iterative optimization problem underneath.

The base class ‘AbstractTrainer<ModelT, LabelTypeT>’

AbstractTrainer is the base interface for trainers for supervised learning problems. It is templatized with respect to the type of model it trains and the type of label in the data set. The trainer then defines the following types in its public interface:

Types Description
ModelType The type of model the trainer optimizes
InputType The type of inputs the model takes
LabelType The type of the labels in the data set

A trainer offers the following methods:

Method Description
train(ModelType&, LabeledData<InputType, LabelType>) Solves the problem and sets the model parameters
std::string name() Returns the trainer’s name

Usage of trainers is equally straightforward:

MyModel model;
MyTrainer trainer;
MyDataset data;
trainer.train(model, data);  //model now represents the solution to the problem.

The base class ‘AbstractUnsupervisedTrainer<ModelT>’

AbstractUnsupervisedTrainer is the base interface for trainers for unsupervised learning poblems. It only needs to know about the model type, and offers a typedef for the data format:

Types Description
ModelType Type of model which the trainer optimizes
InputType Type of inputs the model takes

These trainers also offer the following methods:

Method Description
train(ModelType&, UnlabeledData<InputType>) Solves the problem and stores it in the model.
std::string name() Returns the name of the trainer

List of trainers

We first list the unsupervised trainers in Shark. Many of these operate on models for data normalization.

Trainer Model Description
NormalizeComponentsUnitInterval Normalizer Trains a linear model to normalize the components of data to the unit interval.
NormalizeComponentsUnitVariance Normalizer Trains a linear model to normalize the components of data to unit variance.
NormalizeComponentsWhitening LinearModel Trains a linear model to whiten the data (uncorrelate all components, and normalize to unit variance).
PCA LinearModel Trains a linear model for a principal component analysis, see the PCA tutorial.
NormalizeKernelUnitVariance ScaledKernel Trains the scaling factor of a ScaledKernel such that the data has unit variance in its induced feature space. Note how this trainer operates on a kernel rather than a (linear) model.
OneClassSvmTrainer KernelExpansion Trains a one-class SVM which can be used for outlier detection

List of some supervised trainers:

Trainer Model Description
FisherLDA LinearModel Performs Fisher Linear Discriminant.
KernelMeanClassifier KernelExpansion Computes the class means in the kernel induced feature space and generates a classifier which assigns the points to the class of the nearest mean.
LDA LinearClassifier Performs Linear Discriminant Analysis, see the LDA tutorial.
LinearRegression LinearModel Finds the best linear regression model for the labels.
OptimizationTrainer all Combines the elements of a given learning problem – optimizer, model, error function and stopping criterion – into a trainer.
Perceptron KernelExpansion Kernelized perceptron – tries to find a separating hyperplane of the data in the feature space induced by the kernel.
RFTrainer RFClassifier Implements a random forest of decision trees, see the random forest tutorial.
AbstractSvmTrainer KernelExpansion Base class for all support vector machine trainers.
MissingFeatureSvmTrainer MissingFeaturesKernelExpansion Trainer for binary SVMs supporting missing features.
CSvmTrainer KernelExpansion Trainer for binary and multiclass SVMs, with one-norm regularization, see the SVM introduction.
EpsilonSvmTrainer KernelExpansion Trains an epsilon-SVM for regression.
RegularizationNetworkTrainer KernelExpansion Trains a Gaussian Process model / regularization network.
AbstractLinearSvmTrainer LinearModel Base class for all linear-SVM trainers