https://github.com/nicomignoni/som

PyTorch implementation of Kohonen's Self-Organizing Map.
https://github.com/nicomignoni/som

kohonen self-organizing-map som unsupervised-learning

Last synced: 6 months ago
JSON representation

PyTorch implementation of Kohonen's Self-Organizing Map.

• Host: GitHub
• URL: https://github.com/nicomignoni/som
• Owner: nicomignoni
• License: mit
• Created: 2020-05-03T13:07:15.000Z (over 5 years ago)
• Default Branch: master
• Last Pushed: 2021-02-11T19:06:52.000Z (almost 5 years ago)
• Last Synced: 2024-11-18T17:46:45.228Z (about 1 year ago)
• Topics: kohonen, self-organizing-map, som, unsupervised-learning
• Language: Python
• Homepage:
• Size: 1.2 MB
• Stars: 2
• Watchers: 1
• Forks: 1
• Open Issues: 0
• Metadata Files:
  • Readme: README.md
  • License: LICENSE

Awesome Lists containing this project

README

          
# Kohonen's Self-Organizing Map (SOM)

[![PyPI version shields.io](https://img.shields.io/pypi/v/kohonen-som.svg)](https://pypi.python.org/pypi/kohonen-som/)

![gif](https://github.com/nicomignoni/SOM/blob/master/animation/animation.gif)

## Installation

```

pip install kohonen-som

```

## Background

The original [paper](https://sci2s.ugr.es/keel/pdf/algorithm/articulo/1990-Kohonen-PIEEE.pdf) written by Teuvo Kohonen in 1990 was one of the first neural network model capable of unsupervised learning.

Out of the different implementations of the algorithm, this one follows almost entirely the original paper. The update function is defined as



where



and ![equation](https://latex.codecogs.com/gif.latex?%5Cinline%20t) is the current epoch.

Also, each neuron is connected to all the other ones, hence the map is a ![equation](https://latex.codecogs.com/gif.latex?K_p) complete graph, where ![equation](https://latex.codecogs.com/gif.latex?p) is the number of neurons. 

## Example

```python

from sklearn.datasets import load_iris

from sklearn.decomposition import PCA

import matplotlib.pyplot as plt

import numpy as np

from som.mapping import SOM

dataset = load_iris()

train = dataset.data

# Reducing the dimensionality of the train set from

# 4 to 2 with PCA

pca       = PCA(n_components=2)

train_pca = pca.fit_transform(train)

parameters = {'n_points'  : 500,

              'alpha0'    : 0.5,

              't_alpha'   : 25,

              'sigma0'    : 2,

              't_sigma'   : 25,

              'epochs'    : 300,

              'seed'      : 124,

              'scale'     : True,

              'shuffle'   : True,

              'history'   : True}

# Load and train the model

model = SOM()

model.set_params(parameters)

model.fit(train_pca)

weights = model.get_weights()

# Plot the train dataset and the weights

fig, ax = plt.subplots()

fig.suptitle("Train set (PCA-reduced) and weights")

t = ax.scatter(train_pca[:,0], train_pca[:,1])

w = ax.scatter(weights[:, 0], weights[:, 1])

fig.legend((t, w), ("Train", "Weights"))

plt.show()

```