These are chat archives for bayespy/bayespy

3rd
Apr 2016
Deebul Nair
@deebuls
Apr 03 2016 03:06 UTC
@jluttine Thanks
@jluttine For the problem statement mentioned above . Hierarchical modelling using Bag of Marbles. I have created the model and created posterior inference from the data . Now I want to do prediction . If I choose a bag I need to predict what will be the next color . How can I do this ?
Jaakko Luttinen
@jluttine
Apr 03 2016 08:36 UTC
@deebuls Create a new node which represents the thing you want to predict. that is, use categorical/multinomial as appropriate and give as probability vector the node you have already learned. then you'll be able to examine that node by plotting, with .show() method or .get_moments() method
@deebuls fyi, i'll modify the notebook example a bit. i'll remove references to hierarchical modelling, because i wouldn't consider that as a hierarchical model because there is no shared parameter that is learned in higher levels. i'd consider that as a basic multinomial/categorical example, and a good one as that
Jaakko Luttinen
@jluttine
Apr 03 2016 08:43 UTC
@deebuls unless you want to learn the hyperparameter (beta) too! then i'd call it a hierarchical model :)
this approach would also enable you to predict new bags, not only colors of next marbles
that's one point of hierarchical modelling: you can make predictions on higher levels
but i need to implement that first: bayespy/bayespy#49
Deebul Nair
@deebuls
Apr 03 2016 08:53 UTC
@jluttine Thanks for the prediction part I will try that .
@jluttine for hierarchical modelling . Today I will try to extend the example as suggested. Will update the pull request
Jaakko Luttinen
@jluttine
Apr 03 2016 14:10 UTC
@deebuls i need to implement the hierarchical stuff myself, becaus of some missing features
Deebul Nair
@deebuls
Apr 03 2016 22:40 UTC
@jluttine
So The following hierarchical models cannot be implemented because of missing features ?
Imgur
Imgur