Deterministic filter-like device #8
Conversation
richfitz
force-pushed
the
mrc-5336
branch
2 times, most recently
from
b060d22
to
0e27238
Compare
|
|
||
| obj <- dust2_cpu_sir_unfilter_alloc(base, time_start, time, dt, data, 1, 0) | ||
| ptr <- obj[[1]] | ||
| expect_equal(dust2_cpu_sir_unfilter_run(ptr, NULL, FALSE), f(pars1)) |
There was a problem hiding this comment.
So this one with NULL, FALSE uses base - which happens to be the same as pars1. We could have said f(base) on the end.
There was a problem hiding this comment.
General comment - I find it quite easy to forget what the args are with the functions... Lots of looking up
There was a problem hiding this comment.
yeah, I'll sort this out in one push once we get an actual interface to any of this stuff - none of these functions will be the ones we call in the end!
|
|
||
| obj <- dust2_cpu_sir_unfilter_alloc(pars, time_start, time, dt, data, 1, 3) | ||
| ptr <- obj[[1]] | ||
| expect_equal(dust2_cpu_sir_unfilter_run(ptr, NULL, TRUE), f(pars)) |
There was a problem hiding this comment.
Here, pars[[1 .. 3]] is an entry for each group, and f(pars) gives us 3 results, so a summed-log-likelihood over time, for each of the 3 groups.
| time <- c(4, 8, 12, 16) | ||
| data <- lapply(1:4, function(i) { | ||
| lapply(seq_len(n_groups), function(j) list(incidence = 2 * (i - 1) + j)) | ||
| }) |
There was a problem hiding this comment.
So data[[1 .. 4]] each contain 3 incidence points, one for each group.
There was a problem hiding this comment.
and the 1 .. 4 in data refer to the 4 timesteps
| dt <- 1 | ||
|
|
||
| expect_error( | ||
| dust2_cpu_sir_unfilter_alloc(pars, 5, time, dt, data, 1, 0), |
There was a problem hiding this comment.
2nd arg (5) is r_time_start.
We expect (and get) errors if r_time_start >= length(time)
Merge after #2, contains those commitsMerge after #11, contains those commitsThis PR implements a particle-filter-like method for computing likelihoods out of a time series model by comparing to data periodically. It is the limit case of a particle filter to one particle, and has been used as the "deterministic" method in sircovid/dust1/etc.
The name
unfilteris temporary, we'll probably come up with something better later.The core implementation is in
inst/dust2/filter.hpp- you'll see it's pretty straightforward - we loop over all times in the data set, do the comparison to data and accumulate a likelihood. The R wrapper around this then returns that likelihood out of the container back to R.We'll complicate this shortly with things like returning trajectories as the model runs.
The time handling, and in particular the conversion to
steps_is pretty gross and addressed somewhat in #12, which we might use here soonThe current version only allows using initial conditions from the model, this is addressed in #13