Creating a Map

In order to create and train a Self-Organizing Map on GEMA we will import the Map module from the library:

from GEMA import Map

To initialize a map we will call the constructor and save it on a variable:

my_map = Map(data, size, period, initial_lr, initial_neighbourhood, distance, use_decay, reinforcement, extension, compression, normalization, presentation, weights)

Parameters:

• data : Required

  Numpy array of (N*M) where N is the number of data entries and M is the number of characteristics of the data.

• size: Required

  Integer to indicates the lateral size of the map. It allows the creation of a squared map with size x size neurons.

• period: Optional - Default = 10

  Integer that represents the number of iterations of the training process. Each iteration is a full presentation of all of the patterns.

• initial_lr: Optional - Default = 0.1

  Float value. Indicates the learning rate for the first iteration. It will decrease to 0 in the last iteration according to a lineal function.

• initial_neighbourhood: Optional - Default = 0

  Integer that indicates the distance from where a neuron will be considered inside the neighbourhood and actually learn in the first iteration. It will decrease to 0 (only the BMU will learn) in the last iteration according to a lineal function.

• distance Optional - Default = 'euclidean'

  euclidean considers the neighbourhood as a diamond.

  chebyshev considers the neighbourhood as a square.

• use_decay Optional - Default = False

  Experimental feature to allow the use of decay on the training process.

• reinforcement Optional - Default = 0

  Integer that indicates extra reinforcement period after normal training.

• extension Optional - Default = 1

  Float that indicates the rate at which the period will increase during reinforcements iterations.

• compression Optional - Default = 1

  Float that indicates the rate at which the initial learning rate will decrease during reinforcements iterations.

• normalization Optional - Default = 'none'

  • 'none': No normalization apply to the training data
  • 'euclidean': Euclidean normalization.
  • 'fwn': Feature Wise Normalization

• presentation Optional - Default = 'random'

  Allows to choose if the patterns will be presented in a random ('random') or a sequential ('sequential') way.

• weights Optional - Default = 'random'

  Allows for different weights initialization:

  • 'random': values between 0 and 1
  • 'random_negative': values between -1 and 1
  • 'sample': random values based on the training data
  • 'PCA': values given by a linear combination of the two first principal components of the training data.