Do you have any suggestions on setting the max number of iteration in training som?

[zhenzonglei]

Hi,
I just found there is a inconsistency statement for the verbose output of the train method.
Line 347 states that if verbose is true, the status of the training will be printed at each iteration. But in line 361, the status is only printed after all iterations. I guess the code between 361-363 should be indented.
Thanks

[JustGlowing]

Hi, thanks for checking that! However, the status is printed at every iteration via the generator defined outside the class. Only there errors are printed in there end. It's unpractical to recompute there errors at each iteration.

…

On Wed, Apr 15, 2020, 16:24 Zonglei Zhen ***@***.***> wrote: Hi, I just found there is a inconsistency statement for the verbose output of the train method. Line 347 states that if verbose is true, the status of the training will be printed at each iteration. But in line 361, the status is only printed after all iterations. I guess the code between 361-363 should be indented. Thanks — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#66>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ABFTNGOZG4DR5KAOM47KPTLRMXGTLANCNFSM4MIVLB7A> .

[zhenzonglei]

Got it! Thanks very much.
BTW，do you have any suggestions on setting the max number of iteration in training som? For example, if I have 10,000 samples, what is a reasonable number for the iteration?
Thanks again

[JustGlowing]

Hi again, the number of iterations required for convergence depends on many factors. The main ones are size of the som and shape of the data. The only way to know if you reached convergence is to look at the learning curve and check if it reached a plateau (see the Iris example).

If you have a som 100-by-100, start with 10000 iterations so that each sample is observed at least once and check the results. Increase the the number of iterations if you think that the error is on a downward trajectory.

[V-for-Vaggelis]

Let me extend this question a little further with some emphasis on the topographic error. I have a dataset with around 360 rows, and small correlations between features. After plotting the learning curves like in the "Iris" example I noticed the quantization error indeed shows a decreasing behavior that reaches a plateau and the topographic error shows a fluctuating behavior that tends to become stable as well. The problem is it fluctuates around 0.8 which is too large. Since the t.e is an indication of how representative the SOM is I believe it is an important issue.

The question is whether there is a parameter that if properly tuned can decrease the t.e, or if it is inevitable to get a non-representative SOM for low-correlated data ?

[JustGlowing]

hi @V-for-Vaggelis,

Have you tried inspecting the results visually? You want to check that the u-matrix (that you can get with the method distance_map()) is smooth.

You can obtain a smooth mapping no matter how the data is correlated.

[JustGlowing]

@V-for-Vaggelis also, to really understand if the som has converged you can check the weights step by step and stop when the they don't change anymore (||W_i - W_i-1|| < epsilon).

[V-for-Vaggelis]

@JustGlowing A weird thing happened. I updated minisom and would not print t.e anymore. So I print it myself and got 0.09 for the same data. Could it be a bug you had fixed? I also got the distance map as you advised. In general it has a smooth behavior, but there is a small red area (large distances). I guess it means this small area of the grid can't be trusted to draw conclusions.

Also another thing, is there a paper I can refer to in my thesis for minisom or should I just link to the repo?

[JustGlowing]

@V-for-Vaggelis there was a bug fix released in December related to the quantization error.

Can you please cite MiniSom as follows:

G. Vettigli, "MiniSom: minimalistic and NumPy-based implementation of the Self Organizing Map,". Available:
whttps://github.com/JustGlowing/minisom.