Support for stochastic models #9
Conversation
Codecov ReportAll modified and coverable lines are covered by tests ✅
Additional details and impacted files@@ Coverage Diff @@
## main #9 +/- ##
=========================================
Coverage 100.00% 100.00%
=========================================
Files 38 38
Lines 1575 1597 +22
=========================================
+ Hits 1575 1597 +22 ☔ View full report in Codecov by Sentry. |
richfitz
force-pushed
the
mrc-5095
branch
2 times, most recently
from
9f47375 to
17835a7
Compare
richfitz
force-pushed
the
mrc-5095
branch
3 times, most recently
from
98ed591 to
e180d64
Compare
| ## in order to get trajectories out etc, but that's not super | ||
| ## obvious really; that means we're going to need some sort of | ||
| ## special sauce for the observer function really. | ||
| }) |
There was a problem hiding this comment.
I have not understood this comment well yet!
There was a problem hiding this comment.
ah, it's out of date now after #10
There was a problem hiding this comment.
would like to definitely chat through this and that sir model on call!
| i = hint), | ||
| arg = "model", call = call) | ||
| } | ||
| list(set = model$set_rng_state) |
There was a problem hiding this comment.
just for clarification, why is this returned as list(set = model$set_rng_state) instead of returning model$set_rng_state?
There was a problem hiding this comment.
Because I anticipate that later we'll need get = model$get_rng_state too for cases where we have a stochastic model that we want to be able to restart
| if (isTRUE(model$properties$is_stochastic)) { | ||
| model$model$set_rng_state(rng) | ||
| } |
There was a problem hiding this comment.
will this be applied to different samplers? if so perhaps we should put this in mcstate_run_chain right after sampler$initialise so that this is a general functionality instead of tied to this particular initialisation function
There was a problem hiding this comment.
None of the other samplers can cope with stochastic models; they will be throwing errors!
| ## in order to get trajectories out etc, but that's not super | ||
| ## obvious really; that means we're going to need some sort of | ||
| ## special sauce for the observer function really. | ||
| }) |
There was a problem hiding this comment.
would like to definitely chat through this and that sir model on call!
Note that this PR pulls in
dustfor now, to use in the tests. That is probably not what we really want because dust is going to depend on mcstate for its rng calls! Once we work out what we need, I'll hand-write a version of this model that does the job, but this setup should be ok for a while (probably until we know enough to write dust2, really).I've left a little placeholder in around getting rng state from the model; that is going to interact with restartable mcmc in ways that are going to require some care I think.
There's no actual testing in the stochastic model + random walk; however, this constitutes "particle MCMC" which is cool. There are a couple of small features we will likely add there later based the covid work: