Indexing errors with multi-level GPs

[Reece844]

When specifying a model with something like gp(x,y, by=factor) in the formula, the code generated in the partial log likelihood function will result in a run-time error.

For example, the error occurs due to the way indexing is done without threads, using Igp_1_* and the slicing.

mu[Igp_1_1[start:end]] += Cgp_1_1[start:end] .* gp_pred_1_1[Jgp_1_1[start:end]];
mu[Igp_1_2[start:end]] += Cgp_1_2[start:end] .* gp_pred_1_2[Jgp_1_2[start:end]];
mu[Igp_1_3[start:end]] += Cgp_1_3[start:end] .* gp_pred_1_3[Jgp_1_3[start:end]];
mu[Igp_1_4[start:end]] += Cgp_1_4[start:end] .* gp_pred_1_4[Jgp_1_4[start:end]];

There are a couple different workarounds I can think of, the easiest being to have separate reduce_sum calls for each factor, but I don't know how that would affect speed.

[paul-buerkner]

Thanks! Multiple reduce_sum may not be a good option at the moment because it has too many implications on too much other stuff. I will think of something that works in the current structure. Ideas are very welcome! Reece Willoughby <[email protected]> schrieb am Fr., 29. Jan. 2021, 22:17:

…

When specifying a model with something like gp(x,y, by=factor) in the formula, the code generated in the partial log likelihood function will result in a run-time error. For example, the error occurs due to the way indexing is done without threads, using Igp_1_* and the slicing. mu[Igp_1_1[start:end]] += Cgp_1_1[start:end] .* gp_pred_1_1[Jgp_1_1[start:end]]; mu[Igp_1_2[start:end]] += Cgp_1_2[start:end] .* gp_pred_1_2[Jgp_1_2[start:end]]; mu[Igp_1_3[start:end]] += Cgp_1_3[start:end] .* gp_pred_1_3[Jgp_1_3[start:end]]; mu[Igp_1_4[start:end]] += Cgp_1_4[start:end] .* gp_pred_1_4[Jgp_1_4[start:end]]; There are a couple different workarounds I can think of, the easiest being to have separate reduce_sum calls for each factor, but I don't know how that would affect speed. — You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub <#1081>, or unsubscribe <https://github.com/notifications/unsubscribe-auth/ADCW2AD42S6SKTS2JDD6CWTS4MQXZANCNFSM4WZRT55Q> .

[paul-buerkner]

This should now be fixed on github. Thanks again for opening this issue.