mrgsolve/drake Incompatibility?

[billdenney]

When using drake with mrgsolve, I just got the following error:

There was a problem accessing the model shared object.
  Either the model object is corrupted or  
  the model was not properly compiled and/or loaded.
Check mrgsolve:::funset(mod) for more information.
> mrgsolve::funset(readd(model))
Error: 'funset' is not an exported object from 'namespace:mrgsolve'
> mrgsolve:::funset(readd(model))
$`symbols`
  name loaded
1 <NA>  FALSE
2 <NA>  FALSE
3 <NA>  FALSE
4 <NA>  FALSE

$shlib
            package version compiled
1 pku-so-3549ba2a85   0.9.0     TRUE

> 

Both mrgsolve and drake do magical things with R, and it appears their magic is incompatible in at least this instance.

To reproduce this requires more complexity than a standard reproducible example.

1. Start R
2. source("run1.txt")
3. Quit R
4. Start R
5. source("run2.txt")

run1.txt
run2.txt

[kylebaron]

Try this:

library(mrgsolve)
options(mrgsolve.soloc = "i_love_drake")
mod <- mread_cache("my_model_file.cpp")

I don't know how to use drake; but set the option and use mread_cache rather than mread ... or mcode_cache rather than mcode. I use this in my everyday (drake-less) workflow now.

If that still doesn't work try adding loadso(mod) after the mread_cache call; that will force the loading of the shared object.

[billdenney]

loadso() fixes it, if I manually run it. drake works to only re-run code that has changed while tracking dependencies, so mread() or mcode() isn't run if the model is already current (i.e. the model code hasn't changed). But, I changed my dataset and resimulated, so mrgsim was called in a session where mcode() wasn't called.

Would it be reasonable to add an attempt to load the shared object, if not already loaded, before throwing the error here:

mrgsolve/R/funset.R

Lines 71 to 76 in ea18191

	pointers <- function(x) {
	if(!funs_loaded(x)) stop(FUNSET_ERROR__)
	what <- funs(x)
	ans <- getNativeSymbolInfo(what,PACKAGE=dllname(x))
	setNames(lapply(ans, "[[","address"),names(what))
	}

(And anywhere else that may be an issue)

This code appeared to work:

library(mrgsolve)
library(drake)

model_text <- "
$PARAM @annotated
CL   :  1 : Clearance (volume/time)
V    : 20 : Central volume (volume)
KA   :  1 : Absorption rate constant (1/time)

$CMT  @annotated
EV   : Extravascular compartment
CENT : Central compartment

$GLOBAL
#define CP (CENT/V)

$PKMODEL ncmt = 1, depot = TRUE

$CAPTURE @annotated
CP : Plasma concentration (mass/volume)
"

mrgsim_load <- function(x, ...) {
  loadso(x)
  mrgsim(x, ...)
}

my_plan <-
  drake_plan(
    model=mcode_cache(model_text, model="pk1"),
    my_idata=data.frame(ID=2),
    sim_result=mrgsim_load(model, idata=my_idata),
    strings_in_dots="literals"
  )

make(my_plan)
as.data.frame(readd(sim_result))

[billdenney]

Also, drake is one of my favorite new (to me) R tools. Especially for long-running projects (such as model fitting or larger mrgsolve simulations), it lets me iterate on the reporting side without rebuilding everything from scratch or worrying about dependencies or stale outputs (like saved files).

https://ropensci.org/technotes/2019/03/18/drake-700/

[kylebaron]

Thanks for the report. I'll open tickets to fix the funset output (not sure why the names are not showing up there). Also, will add pointer to try loadso when the model functions are not loaded.