A Python Library for Self Organizing Map (SOM)

Hi,

I have a question about the HitMapView. I do clustering as follows and the output is perfectly fine:

cl = som.cluster(n_clusters=3)

But the HitMapView always uses the default number of clusters which is 8 and I could not find a way to change it:

h = hitmap.HitMapView(10, 10, 'hitmap', text_size=8, show_text=True)

How can I specify number of clusters in a HitMapView to a different number?

Regards Amin

If labels == True labels are generated using build_data_labels(), if labels is a list of strings then those are used instead.

So all I added the following within the show() method of UMatrixView:

if labels:
    if labels == True:
      labels = som.build_data_labels()
    for label, x, y in zip(labels, coord[:, 1], coord[:, 0]):
        plt.annotate(str(label), xy = (x, y), horizontalalignment = 'center', verticalalignment = 'center')

I add a parameter to the show function inside the View2D class for desnormalize the components scale. This way it is far easier to interpret the results. Example of usage:

from sompy.visualization.mapview import View2D view2D = View2D(5,1,"rand data",text_size=7) view2D.show(sm, col_sz=3, desnormalize=True)

Hello,

I have installed sompy without any errors (https://github.com/sevamoo/SOMPY). However when i try:

import sompy

I receive an error code:

File "C:\Anaconda3\lib\site-packages\sompy-1.0-py3.5.egg\sompy__init__.py", line 30, in from sompy import SOMFactory ImportError: cannot import name 'SOMFactory'

I am using Python 3.5.1 (Anaconda 4.1.1 for Windows) Any help or suggestions would be greatly appreciated.

Stefan

hello, I'm trying to use some of the visualization forms, however, I always run into the same error.

Model mapsize = [10,10] som = sompy.SOMFactory.build(diabetes, mapsize, mask=None, mapshape='planar', lattice='rect', normalization='var', initialization='pca', neighborhood='gaussian', training='batch', name='sompy') som.train(n_job=1, verbose='info')

MapView `from sompy.visualization.mapview import View2D

view2D = View2D(10,10,"", text_size=7) view2D.show(som) plt.show()`

Good day,

There seems to be another bug, in particular, the using rect mode in train and later k-means and visualizing. This is the end product of the rect visualization using hitmapview.

The color seems to be out of place with the labels:

This seems to be the color mapping associated with the hexa plot: The following is training on hexa mode and using k-means

I was using the an older version initially (rect visual works) and i believe there might be a chance that there is some bug associated with it. Below is the rect mod plotted using the older version.

Hi there,

After the map has been trained by the algorithm, is there any way to acess the neuron weight values? I would like to check what has been stored in specific indexes to see which clusters have been identified from my training set.

Thanks, Jack

Can you make any of the examples worked please? None of it can be runned because of changes in the sompy library (parameters/methods names, etc)

In the sample http://nbviewer.jupyter.org/urls/gist.githubusercontent.com/sevamoo/8f26d64470e00960684a/raw/SOMPY_example

the constructor call has the wrong parameters:

sm = SOM.SOM('sm', Data, mapsize = [msz0, msz1],norm_method = 'var',initmethod='pca')

and so on..

Hi there! First of all thank you for developing SOMPY! I have been using SOMPY for some time now and one of the features I think it's missing is the hexagonal grid for the u-matrix plot so I decided to implement it myself and share it here. Let me know of what you think and if this is a feature you want to add to the package or not. Also, I am planning to build an interactive u-matrix visualization in the future with Bokeh, let me know if you would be interested in adding it to the package.

Changed generate_hex_lattice() from sompy/codebook.py such as the coordinates correspond to a regular hexagon grid (which wasn't the case previously).

Changed sompy/visualization/umatrix.py such that if the som object has a hexagonal lattice, then the u-matrix is plotted on a hexagonal grid. All additional plot features (e.g. contour, blob, etc.) are compatible with the hexagonal grid. Also added colorbar to the plot.

Hi another time,

I just added something that I need! I wrote a function for calculating the topographic error, which in my opinion is equal or more important than the quantization error. It controls the proportion of units which model similar prototypes are contiguous. It is calculated as the proportion of all data vectors for which first and second BMUs are not adjacent units.

The general strategy I follow for training reliable SOM is:

• Generate several maps varying the seed and the accessible parameters like training_length, and radius.
• Calculate the quantization and topographic errors
• Choose, between the maps which have very low or 0 topographic errors, the one which have less quantization error. This way I am choosing a map which will be very interpretable... hint: it is a good idea, once all the maps have been trained, to draw a scatterplot representing the quantization error vs the topographic error.

I modified the find_som function and the related ones for allowing the calculation of the 2nd BMU. If not specified, it works in a normal way

Hi!

I have just implemented hex lattice and its visualization tools. Please find attached a teaser!

Apart from that, my contribution contains

• Global minor refactoring
• Add function to correctly calculate the quantization error

In the next days, I will probably be adding more support to the hex lattice and cleaning some parts of the library! I am also planning to write a tutorial jupyter notebook for with an hex lattice example

Best, Iván Vallés

    Hello  Vahid Moosavi,

Thank you for inventing this great sompy package. It was very useful to me. I have been using it for my article publication and it will be referenced duly.

However I have one problem understanding the component planes. My input variables are many, hence deciphering the patterns to associations just from the component planes is very cumbersome. I need help on the sompy command which I can use to get the results from the component planes so I can do statistical inferences on it.

I want to know if there is a way out to get this problem resolved. I will be glad if my requisition is given the due clarification.

Counting on your cooperation. Thank you

Best Regards, John Owusu

Originally posted by @nanakayhacker in https://github.com/sevamoo/SOMPY/issues/128#issuecomment-1366372339

When running:

sm = SOMFactory().build(data, mapsize=[10,10], normalization = 'var', initialization='random', component_names=names, lattice="hexa")
sm.train(n_job=1, verbose=False, train_rough_len=2, train_finetune_len=5)

from sompy.visualization.mapview import View2D
view2D  = View2D(10,10,"",text_size=10)
view2D.show(sm, col_sz=5, which_dim="all", denormalize=True)
plt.tight_layout()
plt.show()

I get overlapping hexagons. I'd like to adjust their spacing so that they just touch or even have a little space between them.

I have reviewed the source codes and can't seem to find where to make that adjustment.

When installing requirements using setup file python setup.py install

An exception has occurred, use %tb to see the full traceback.

Traceback (most recent call last):

File "C:\Users\IPCS\anaconda3\lib\distutils\core.py", line 134, in setup ok = dist.parse_command_line()

File "C:\Users\IPCS\anaconda3\lib\site-packages\setuptools\dist.py", line 707, in parse_command_line result = _Distribution.parse_command_line(self)

File "C:\Users\IPCS\anaconda3\lib\distutils\dist.py", line 501, in parse_command_line raise DistutilsArgError("no commands supplied")

DistutilsArgError: no commands supplied

During handling of the above exception, another exception occurred:

SystemExit: usage: setup.py [global_opts] cmd1 [cmd1_opts] [cmd2 [cmd2_opts] ...] or: setup.py --help [cmd1 cmd2 ...] or: setup.py --help-commands or: setup.py cmd --help

error: no commands supplied

I've dataset ok 645K rows and 25 features, I've trained using sompy with this:

# train some models and check the best one
for i in range(7):
    sm = SOMFactory().build(data, mapsize=[random.choice(list(range(15, 25))),
                                           random.choice(list(range(10, 15)))],
                            normalization = 'var', initialization='random',
                            component_names=names, lattice="hexa")
    
    sm.train(n_job=4, verbose=False, train_rough_len=16, train_finetune_len=16)
    joblib.dump(sm, path_out+"Models/model_{}.joblib".format(i))

the joblib file has about 750Mo how this can be possible?

Hi,

I was wondering if there is a SAMMON Map feature in sompy like that in the matlab SOMtoolbox? I was looking through the visualization folder but couldn't find something similar where I could plot the relative location of each node to each other. I tried checking the sompy output to see if there was something that could be used to make a SAMMON map but the closest thing I could find would be the "lattice_distances" however it was turning out to be a really convoluted process.

Thanks in advance!

I was using somoclu's library but It isn't working for python 3 and I just started working with SOMPY. The idea is try to do the same thing, I would like to plot the class's point for every input. This's with somoclu

And this's with SOMPY

Is that possible?

Thank you in advance!