How to evaluate the log probability of fitted model?

[juliohm]

I am trying to guess by reading the source code, but I am not 100% sure. Could you please clarify two basic questions. Assume I have fitted a 4D KDE.

1. How to evaluate the log-probability at a new point?
2. How to sample from the KDE?

I think for the latter, there is sample(kdeobj, npts), correct? Could you please comment on the output? I remember it returning a tuple? What is the meaning of the second entry?

For the former, I found a set of eval* methods with very complicated arguments, I appreciate if you can clarify those.

Thanks,

[dehann]

Hi @juliohm , there is a cleaner interface for evaluating the probability in the works:

p = kde!(randn(2,100))
p([0.0;0.0])

The existing method is using evalDualTree as mentioned here #2 . The work is being done by evaluate(...). You are probably looking for evalAvgLogL(...), but a convenient function call has not been implemented yet. Maybe just follow evalDualTree(...) as an example?

EDIT: evaluations are done using the BallTreeDensity class, but there is a convenience wrapper to convert test points into a BallTreeDensity. The new kde!(pts)(evallocations) approach is meant to make use of eval* easier.

[dehann]

Regarding sampling, you can just do

p = kde!(randn(100))
pts = rand(p, 1000)

EDIT: randn -> rand

[juliohm]

Thank you @dehann , I will keep it in mind next time I need KDE in my work, this time I had to rely on the Python's sklearn implementation due to an urgent deadline.