NaN loss after a few epochs about pycave HOT 4 CLOSED

If you want it to debug within PyCave, I would recommend that you add some print statements to pycave/bayes/gmm/engine.py in the train_batch method. Otherwise, I can also have a look if you can tell me how I can replicate your data (I suppose every feature is distributed according to the Standard Normal then?).

from pycave.

Great that you found it! I will make the regularization factor configurable :)

from pycave.

Hi @borchero ,
It seems that my problem was the singularity GMM issue.
I solved the problem by increasing the covariance regularisation factor from 1e-6 to 1e-5.

I found that when I apply StandardScaler to my data (audio MFCCs) and select more than 20 components, one component ended with a covariance very close to the regularization factor, i.e., all its values close to 1e-6.
Using the same logic as in Sklearn, I increase that factor (https://github.com/scikit-learn/scikit-learn/blob/b3ea3ed6a09fe774dfc5160a65172b1bacbb2a82/sklearn/mixture/_gaussian_mixture.py#L306).

Is there a reason why the regularisation factor is not a model parameter?

from pycave.

Fixed in 2.0.4 ;)

from pycave.