search

cbfinn / gps

Guided Policy Search
http://rll.berkeley.edu/gps/
Other
594 stars 239 forks source link

Faster GMM #54

Closed wmontgomery4 closed 7 years ago

wmontgomery4 commented 7 years ago

Rewrote the Estep to be much faster (only one call to solve_triangular per cluster*iteration, while before it had two). This should also scale better wrt N/K.

Code shouldn't be any different theoretically (ran some tests to confirm that log-likelihoods is still the same), and it's working for me on mjc_mdgps_example.

wmontgomery4 commented 7 years ago

I've tested this with the single GMM experiments on the clustering fork, and it works just as well there, so I think this is a pretty safe update at this point.

The only thing I realized is that I'm just getting mu, not mu[:, :D], but I'm not really sure why this was ever used, since the number of dimensions should always be the same for the data and the GMM model.

cbfinn commented 7 years ago

Just tested it out myself, and it looks good -- identical output and faster. Thanks Harley!