Models.md

Trained Model Interfaces

• Trained Model Interfaces
  • For classification
  • For regression
  • Load and save a model
    • Save example
    • Load example

A trained model for classification or regression will implement one of the common interfaces: IModel, IRegressionModel, or IMultivariateRegressionModel. These interfaces provide the methods that allow you to use the model for classification or prediction, save and load it.

For classification

All classification models implement the IModel interface. It provides the Classify method for classifying data and the Serialize method for saving and loading the model to external storage.

class NEOML_API IModel : virtual public IObject {
public:
	virtual ~IModel() = 0;

	// The number of classes
	virtual int GetClassCount() const = 0;

	// Classifies the input vector and returns true if successful, false otherwise
	virtual bool Classify( const CSparseFloatVectorDesc& data, CClassificationResult& result ) const = 0;

	// Serializes the model
	virtual void Serialize( CArchive& archive ) = 0;
};

For regression

The regression models implement the IRegressionModel (for functions that return a number) and the IMultivariateRegressionModel (for functions that return a vector) interfaces. They provide the Predict method for predicting the function value on a given vector and the Serialize method for saving and loading the model to external storage.

// Regression model for a function that returns a number
class IRegressionModel : virtual public IObject {
public:
	// Predict the function value on a vector
	virtual double Predict( const CSparseFloatVector& data ) const = 0;
	virtual double Predict( const CFloatVector& data ) const = 0;
	virtual double Predict( const CSparseFloatVectorDesc& desc ) const = 0;

	// Serialize the model
	virtual void Serialize( CArchive& archive ) = 0;
};

// Regression model for a function that returns a vector
class IMultivariateRegressionModel : virtual public IObject {
public:
	// Predict the function value on a vector
	virtual CFloatVector MultivariatePredict( const CSparseFloatVector& data ) const = 0;
	virtual CFloatVector MultivariatePredict( const CFloatVector& data ) const = 0;

	// Serialize the model
	virtual void Serialize( CArchive& archive ) = 0;
};

Load and save a model

Use the Serialize method and the CArchive class to save and load a model.

Save example

void StoreModel( CArchive& archive, IModel& model )
{
	CString modelName = GetModelName( &model );
	archive << modelName;
	model.Serialize( archive );
}

Load example

CPtr<IModel> LoadModel( CArchive& archive )
{
	CString name;
	archive >> name;
	CPtr<IModel> result = CreateModel<IModel>( name );
	result->Serialize( archive );
	return result;
}