A Python Library for Self Organizing Map (SOM)

1. The library implement fast online training. The implementation of the algorithm is carefully done through matrix calculations, such as scipy sparse matrix and numexpr for calculation of Euclidean distance.
2. PCA (or RandomPCA (default)) initialization, using sklearn or random initialization.
3. component plane visualization (different modes).
4. Hitmap.
5. U-Matrix visualization.
6. 1-d or 2-d SOM with only rectangular, planar grid. (works well in comparison with hexagonal shape, when I was checking in Matlab with somtoolbox).

After the SOM has been trained, the map needs to be evaluated to find out if it has been optimally trained, or if further training is required. The SOM quality is usually measured with two criteria: quantization error (QE) and topographic error (TE). The QE is the average distance between each data point and its BMU, and TE represents the proportion of all data for which the first and second BMU are not adjacent with respect to the measurement of topology preservation (Kohonen, 2001).

SOMPY has the following dependencies:

• numpy
• scipy
• scikit-learn
• matplotlib
• pandas

there are various notebooks in the examples directory

pip install lightSOM