search

probml / dynamax

State Space Models library in JAX
https://probml.github.io/dynamax/
MIT License
667 stars 76 forks source link

Optionally extracting intermediate states for `nonlinear_gaussian_ssm` methods #288

Closed gerdm closed 1 year ago

gerdm commented 1 year ago

Hi,

Is there support (or plan to support) to decide which variables to store in the output history for EKF and UKF inside nonlinear_gaussian_ssm?

For example, if the latent space is high-dimensional and T is large, I would like to not have the filtered_cov history as I would run into memory issues. Also, in the case for online training and testing, having h(prior_mean, u) would be useful to asses the performance of the method.

murphyk commented 1 year ago

These are both useful features. Can you please write them, test them, and make PR?

On Fri, Nov 25, 2022 at 1:30 AM Gerardo Durán-Martín < @.***> wrote:

Hi,

Is there support (or plan to support) to decide which variables to store in the output history for EKF and UKF inside nonlinear_gaussian_ssm?

For example, if the latent space is high-dimensional and T is large, I would like to not have the filtered_cov history as I would run into memory issues. Also, in the case for online training and testing, having h(prior_mean, u) would be useful to asses the performance of the method.

— Reply to this email directly, view it on GitHub https://github.com/probml/dynamax/issues/288, or unsubscribe https://github.com/notifications/unsubscribe-auth/ABDK6EGELUET364WN5NDDM3WKCBLBANCNFSM6AAAAAASLD7HVI . You are receiving this because you are subscribed to this thread.Message ID: @.***>

-- Sent from Gmail Mobile