By templating the Prior over the State, we can avoid having to serialize and de-serialize to and from proto every time we sample with a SemiConjugateUpdater.
In a benchmark one-dimensional example with 2 clusters, this gives a 30% performance improvement. It is likely that the improvement will eventually be irrelevant in more challenging scenarios with multivariate data.
At the moment, only the NNIGHierarchy has been changed.
mberaha
commented
2 years ago
By templating the Prior over the State, we can avoid having to serialize and de-serialize to and from proto every time we sample with a SemiConjugateUpdater.
In a benchmark one-dimensional example with 2 clusters, this gives a 30% performance improvement. It is likely that the improvement will eventually be irrelevant in more challenging scenarios with multivariate data.
At the moment, only the NNIGHierarchy has been changed.