[Rcpp-devel] Using formula

[Denis Haine]

Hi,

Quick follow-up and additional question.

I got it working with fastglm:

#include <Rcpp.h>
  using namespace Rcpp;
  // [[Rcpp::export]]
  NumericVector misclass(NumericMatrix obs_mat) {
    // Obtain environment containing function
    Rcpp::Environment base("package:stats");
   // Obtaining namespace of fastglm package
   Environment pkg = Environment::namespace_env("fastglm");
   Function f = pkg["fastglm"];
  ...
  mod_log = f(Named("x", ematrix), Named("y", d), Named("family", 
"binomial"));

which works fine, but I cannot figure out how to get another link than 
the default 'logit' (here for binomial()). I tried:

Named("family", "binomial(link = \"log\")")

Named("family", "binomial(link = log)")

Both gave me this error: "object of node 'function' not found.

Named("link", "log")

Named("make.link", "log")

These last two do not cause an error, but they give the default logit 
link (as if the Named was not evaluated?).

Thanks,

Denis

Le 2024-09-04 16:24, Dirk Eddelbuettel a écrit :

> On 4 September 2024 at 15:37, Denis Haine wrote:
> | Hi,
> |
> | Sorry for a beginner's question. I'm trying to call an R function 
> (glm())
> | inside my cpp code. The code compiles with no problem, but when I'm 
> running it,
> | it cannot find the second element of the formula, i.e. the x in y~x. 
> The error
> | message is: Error in eval (predvars; data, env) : object 'e' not 
> found.
> |
> | However, if I return 'e', it is correctly calculated. I guess the 
> formula is
> | not correctly evaluated, but I haven't found any examples that could 
> point me
> | in one direction or another.
> |
> | #include <Rcpp.h>
> | using namespace Rcpp;
> | // [[Rcpp::export]]
> | NumericVector misclass(NumericMatrix obs_mat) {
> |   // Obtain environment containing function
> |   Rcpp::Environment base("package:stats");
> |
> |   // Picking up glm() and summary() function from base stats
> |   Rcpp::Function glm_r = base["glm"];
> |   Rcpp::Function sum_r = base["summary.glm"];
> 
> You are running glm() from base here. At the speed of glm() from base.
> 
> That you call it from C++ does not make it faster, sadly, it just 
> creates
> more work for you. Sorry to be a bearer of bad news.
> 
> |   Rcpp::NumericVector d = obs_mat(_, 1);
> |   Rcpp::NumericVector e = no_init(n);
> |   Rcpp::NumericVector mod_coef = no_init(n);
> |   // e is calculated in other section of the code
> |   e = as<IntegerVector>(e);
> |   Rcpp::List mod_pois = glm_r(_["formula"] = "d ~ e",
> |                     _["family"] = "poisson");
> 
> You need to 'work out' (ie expand) the formula on the R side. I have 
> done
> that for the lm() case of model matrices and whatnot in the fastLm()
> example(s) inside eg RcppArmadillo.  Once you have done the R-level 
> unpacking
> you can call the C++ function to fit.
> 
> Big takeaway from that exercise there: the time you spent dealing with 
> the
> (convenient) formula dominates the gain from using a different, 
> simpler,
> slightly faster 'fitter' by many orders of magnitude.
> 
> [ I think there are some packages dealing with glm() fits using Rcpp 
> and
> friends among the by now over 2800 CRAN packages using Rcpp -- but I 
> can't
> right now recall their names. You may find some help in those if you 
> find
> them, and if my recollection was correct in the first place ... ]
> 
> Dirk
> 
> |   Rcpp::List mod_sum = sum_r(mod_pois);
> |   Rcpp::NumericMatrix M_coef = mod_sum[12];
> |   mod_coef = M_coef(2, 1);
> |
> |   return mod_coef;
> | }
> |
> |
> | I also tried providing the formula in the call, i.e. NumericVector 
> misclass
> | (NumericMatrix obs_mat, Formula f) and using it in glm_r, i.e. 
> glm_r(_
> | ["formula"] = f, etc. but with the same outcome.
> |
> | Thanks,
> |
> | Denis
> |
> | _______________________________________________
> | Rcpp-devel mailing list
> | Rcpp-devel at lists.r-forge.r-project.org
> | 
> https://lists.r-forge.r-project.org/cgi-bin/mailman/listinfo/rcpp-devel
-------------- next part --------------
An HTML attachment was scrubbed...
URL: <http://lists.r-forge.r-project.org/pipermail/rcpp-devel/attachments/20240926/1f748e20/attachment.htm>