symbolicwisardpkg 0.1.4

This project is an extension of the wisardpkg library and provides additional machine learning models based on the WiSARD architecture, including symbolic capabilities. It offers implementations of WiSARD-based models with high performance, easy installation, and a unified usage pattern.

Symbolicwisardpkg

Description:

This project is an extension of the wisardpkg library that expands the WiSARD ecosystem by introducing new models and functionalities while preserving the original design principles of simplicity, performance, and ease of use.

The library provides machine learning models based on the WiSARD architecture, supporting supervised, unsupervised, and semi-supervised learning paradigms. In addition to the classical WiSARD model, this project introduces neuro-symbolic extensions that enable the direct incorporation of symbolic knowledge into weightless neural networks.

The following WiSARD-based models are made available:

• WiSARD: the classical weightless neural network model based on RAM memories.

• SWiSARD: a neuro-symbolic extension of WiSARD that allows the direct insertion of logical rules into discriminators, enabling hybrid learning from rules and examples.

• ClusWiSARD: a clustering-based WiSARD model for unsupervised learning.

to install:

python:

pip install Symbolicwisardpkg

Works to python2 and pyhton3.

If you are on Linux and not in a virtual environment, you may need to run as superuser.

obs:

To install on windows platform you can use anaconda and do:

python -m pip install Symbolicwisardpkg

c++:

copy the file Symbolicwisardpkg.hpp inside your project

include/Symbolicwisardpkg.hpp

to uninstall:

pip uninstall Symbolicwisardpkg

to import:

python:

import Symbolicwisardpkg as wp

c++:

# include "Symbolicwisardpkg.hpp"



namespace wp = Symbolicwisardpkg;

to use:

WiSARD

WiSARD with bleaching by default:

python:

# load input data, just zeros and ones  

X = [

      [1,1,1,0,0,0,0,0],

      [1,1,1,1,0,0,0,0],

      [0,0,0,0,1,1,1,1],

      [0,0,0,0,0,1,1,1]

    ]



# load label data, which must be a string array

y = [

      "cold",

      "cold",

      "hot",

      "hot"

    ]





addressSize = 3     # number of addressing bits in the ram

ignoreZero  = False # optional; causes the rams to ignore the address 0



# False by default for performance reasons,

# when True, WiSARD prints the progress of train() and classify()

verbose = True



wsd = wp.Wisard(addressSize, ignoreZero=ignoreZero, verbose=verbose)







# train using the input data

wsd.train(X,y)



# classify some data

out = wsd.classify(X)



# the output of classify is a string list in the same sequence as the input

for i,d in enumerate(X):

    print(out[i],d)

SWiSARD

SWiSARD with bleaching by default:

python:

# load input data, just zeros and ones  

X = [

      [1,1,1,0,0,0,0,0],

      [1,1,1,1,0,0,0,0],

      [0,0,0,0,1,1,1,1],

      [0,0,0,0,0,1,1,1]

    ]



# load label data, which must be a string array

y = [

      "cold",

      "cold",

      "hot",

      "hot"

    ]



addressSize = 3     # number of addressing bits in the ram

ignoreZero  = False # optional; causes the rams to ignore the address 0

verbose = True      # optional; prints the progress of train() and classify()



wsd = wp.Wisard(addressSize, ignoreZero=ignoreZero, verbose=verbose)



# Add symbolic rules to discriminators

# variableIndexes maps variable names to indices in the input vector

# Example: if the vector has 8 positions [0,1,2,3,4,5,6,7], we can name:

# - x0 for index 0, x1 for index 1, etc.



# Add rule for class "cold": (x0 * x1) + (x0 * x2)

# This means: (position 0 AND position 1) OR (position 0 AND position 2)

variableIndexes_cold = {

    "x0": 0,  # first position of the vector

    "x1": 1,  # second position of the vector

    "x2": 2   # third position of the vector

}

rule_cold = "(x0 * x1) + (x0 * x2)"  # boolean expression: * = AND, + = OR, ! = NOT

alpha = 10  # rule weight (the higher, the more influence the rule has)



wsd.addRule("cold", variableIndexes_cold, rule_cold, alpha)



# Add rule for class "hot": (x4 * x5) + (x5 * x6)

variableIndexes_hot = {

    "x4": 4,

    "x5": 5,

    "x6": 6

}

rule_hot = "(x4 * x5) + (x5 * x6)"

wsd.addRule("hot", variableIndexes_hot, rule_hot, alpha)



# Train using trainWithRules (considers rules during training)

wsd.trainWithRules(X, y)



# Classify using classifyWithRules (considers rules during classification)

out = wsd.classifyWithRules(X)



# the output of classifyWithRules is a string list in the same sequence as the input

for i, d in enumerate(X):

    print(out[i], d)

ClusWiSARD

ClusWiSARD with bleaching by default:

addressSize        = 3    # number of addressing bits in the ram.

minScore           = 0.1  # min score of training process

threshold          = 10   # limit of training cycles by discriminator

discriminatorLimit = 5    # limit of discriminators by clusters



# False by default for performance reasons

# when enabled,e ClusWiSARD prints the progress of train() and classify()

verbose = True



clus = ClusWisard(addressSize, minScore, threshold, discriminatorLimit, verbose=True)







# train using the input data

clus.train(X,y)



# optionally you can train using arbitrary labels for the data

# input some labels in a dict,

# the keys must be integer indices indicating which input array the entry is associated to,

# the values are the labels which must be strings

y2 = {

  1: "cold",

  3: "hot"

}



clus.train(X,y2)



# classify some data

out = clus.classify(X)



# the output of classify is a string list in the same sequence as the input

for i,d in enumerate(X):

    print(out[i], d)