theta doc index don't make sense

[bhomass]

I read out Theta for each layer T, and use that to extract topic weights for each document. In particular I worked with the first layer T = 1. The dimensions are correct (K, N). But when looking at the dominant topic for each document, sorted based on the original document index, the assignment is clearly wrong. Is there some kind of shuffling of the index of these Theta values?

[bhomass]

the inconsistency on the topic weight per doc shows up over multiple experiments and source text. There is definitely a bug somewhere.

[bhomass]

look like the topic indices are good for s=3, but fails miserably for s=1. A very strange problem to have.