Created by: DavidDoukhan

Hi list, I was willing to use class bob.learn.em.GMMMachine. I was following the following documentation, which consist to create a GMMMachine instance, and set means, covariance and weights manually: https://www.idiap.ch/software/bob/docs/latest/bioidiap/bob.learn.misc/master/guide.html

I installed bob with default options using pip. According to the few logs bellow, it seems the GMM weights are not taken into account when computing the log likelihood of a sample.

import bob.learn.em
from bob.learn.em import GMMMachine
import numpy as np
print bob.learn.em.__version__

2.0.8

gmm1 = GMMMachine(1,1)
data = np.ones((1,1))
gmm1(data)

-1.4189385332046727

gmm1.weights[:] = 666
gmm1(data)

-1.4189385332046727

After few more investigation, I had the feeling that this behavior had to do with the caching of the log weights, which is not done when assigning new values to the weights, so I found a workaround consisting to create a new instance from a illed instance:

gmm2 = GMMMachine(gmm1)
gmm2(data)

5.082351137335717

This workaround seems to provide coherent results.

In conclusion, I think that this behavior should either be fixed, or set explicit within the documentation.

Regards,