Previous topic


Next topic

Classification Learners

This Page

Generic Learners

The learners.generic module contains Learners that are not designed for a specific type of problem or data. They mostly serve as interfaces to derive new Learners.

This module contains the following classes:

  • Learner: Root class for all learning algorithms.
  • OnlineLearner: Interface for Learners that can be traiend “online”.
class learners.generic.Learner[source]

Root class or interface for a learning algorithm.

All Learner objects inherit from this class. It is meant to standardize the behavior of all learning algorithms.


Runs the learning algorithm on trainset


Resets the Learner to its original state.


Computes and returns the output of the Learner for dataset. The method should return an iterator over these outputs.


Computes and returns the outputs of the Learner as well as the cost of those outputs for dataset. The method should return a pair of two iterators, the first being over the outputs and the second over the costs.

class learners.generic.OnlineLearner(n_stages, **kw)[source]

Class (interface) for Learners that can be trained “online”.

This class interface makes it easier to construct a learner. All that must be defined are four following methods:

  • initialize_learner(self,metadata)
  • update_learner(self,example)
  • use_learner(self,example)
  • cost(self,output,example)

Method initialize_learner() simply initializes the learner. The training set’s ‘metadata’ is also available.

Method update_learner() updates the learner’s parameters using the given ‘example’.

Method use_learner() should return the output for the given ‘example’. The output should be a sequence (even if it has just one element in it), to allow for multiple outputs. Make sure not to return an object that is referenced internally and is still being used by the class object.

Method cost() should return the cost associated to some ‘output’ for the given ‘example’. The returned cost should be a sequence (even if it has just one element in it), to allow for multiple costs.

Option n_stages specifies how many iterations over the training set is done at every call of train().

All other hyper-parameters for the learner supplied through the constructor will be assigned as attributes to the object, and hence will be accessible by all methods.

Example of methods for a linear perceptron.

import numpy as np
import mlpython

class Perceptron(mlpython.learners.OnlineLearner):

   def initialize_learner(self,metadata):
      self.w = np.zeros((metadata['input_size']))
      self.b = 0.
   def update_learner(self,example):
      input = example[0]
      target = 2*example[1]-1 # Targets are 0/1
      output =,input)+self.b
      if np.sign(output) != target:
         self.w += * target * input
         self.b += * target
   def use_learner(self,example):
      return [np.sign(,example[0])+self.b)]
   def cost(self,output,example):
      return [int(output != 2*example[1]-1)]

When creating an instance, must provide the value of the hyper-parameter lr:

perceptron = Perceptron(1,lr=0.01)

Alternatively, one could override the constructor to specify some default hyper-parameters:

class Perceptron(mlpython.learners.OnlineLearner):
   def __init__(self, n_stages, 
                lr = 0.01):
       self.n_stages = n_stages
       self.stage = 0 = lr