Inconsistent Results using Somoclu about somoclu

[xgdgsc]

Have you tried with the command line interface? Some information about what (single-threaded) library and parameter you use would also help.

[pcamelo]

I'm making a production implementation of a paper I will publish soon. The first prototype was done with Neuroph, a ML library for Java that have a SOM implementation, I've got the 700+ clusters for a learning rate of 0.35 and 2000 iterations.

The current implementation is being done in Python and the library that gave me similar results is minisom (https://github.com/JustGlowing/minisom) with 10000 iterations, a sigma of 0.8 and a learning rate of 0.45. For comparison purposes 2000 iterations with minisom gives me ~250 clusters (sigma=0.8, learning_rate=0.45 also).

I will try with the CLI and see what I get.

[pcamelo]

So, I have ran Minisom and Somoclu's CLI. With the parameters above i've got 723 clusters for Minisom and 23 clusters for Somoclu.

I saved the interim files. If it helps I can send you a sample of the dataset I'm using and the the interim files as well.

[xgdgsc]

I feel the reason is the 'Epoch' in somoclu is different from the 'iteration' in minisom. One minisom iteration trains one sample while one epoch in somoclu goes through all samples.

[kryptpt]

I'm sorry @xgdgsc but what is that means? Should @pcamelo reduce the number of epochs?

[xgdgsc]

I think so. He can calculate the equivalent parameter in both and see the result.

[pcamelo]

I will do that and give feedback once I reach a conclusion.

Thank you @xgdgsc and @kryptpt .