Procla QSOM 1.4

This is Procla, frontend for QSOM classifier. QSOM is general classifier based on the Ideas of Teuvo Kohonen SOM Self Orgaizing Memory. Q here stands for quick. QSOM has three modes, train, diagnose and analyze. OSOM uses text files as input and output. Text file are either tab separated columns of of lines or JSON formatted.


Train needs only one file, the training data to make the SOM map. The file is named using the train file name replacing its extension with .som. Train file format is sample lines where column of measurements are separated by tab. '.' dot is used as decimal separator. The first line begins with # and has the tab separated coun of sample lines and count of measurements per sample. The rest of line is considered as the name of sample. Train may also get some options, like --layers=3 which will make a classification map for 4x4 square.


Analyze reads test file which has same format as train file and uses previously generated SOM map to classify the test data samples to known neurons, i.e. Numbered with the neuron number whose distance to the sample is smallest. Also if the sample is outside of the known min and max and error mark is produced.


Analyze combines train and diagnose, where the train data is used as testdata.


This selection show the SOM file in text. The file is divided in header section and neurons section. Header contain information about the structure of SOM map and min,max,dev and ave of the vector values. min and max are the limit of values used in training, ave is the average value of vectors smples and dev is deviation. These can be used to scale back the neuron values to real values as measured. Neurons always has the same values normalized i.e. (real value - ave) dev. Neuron dev value is the deviation of the allocated samples for that neuron.