davmre
commented
5 years ago Hi Michael, it's certainly possible, though I think the new logic isn't totally trivial. Instead of just doing an update vs not doing an update, I think you'd need to modify the observation model -- essentially project the observation matrix, noise mean and cov into the space of just the unmasked dimensions -- and do a partial update from just those dimensions.
Multivariate masking isn't on our immediate roadmap, but we're certainly happy to review pull requests!
Is it possible to allow mask (in
forward_filter,log_prob, etc) to be multivariate i.e. with dimensions[num_timesteps, observation_size]rather than just[num_timesteps]`? That will allow missing observations within individual variables per time step and will be very helpful. I believe most of the current logic should work with minimal change.