read_csv_header() can match more than one line

Summary:

If the path to the output file in the Stan CSV matches one of the regexes in read_csv_header(), then the function throws an error.

Description:

read_csv_header() uses regexes to extract information like the number of warmup iterations from the Stan CSV. Because this is done in a while() loop, if multiple lines (prior to the header) match a regex, then the last match will be used. I happened to include the term within in the name one of my models (I was experimenting with within-chain parallelization), and noticed that read_stan_csv() was throwing an error. It turns out that the output file name was matching the thin regex, and because the result was a character, integer coercion resulted in an NA value that propagated to iter.count.

It seems that this could possibly be avoided by replacing the regex "#.*thin" with "#[[:blank:]]*thin". Another solution would be to avoid including "reserved words" in the model name.

Reproducible Steps:

The following reprex demonstrates the issue, and shows the solution proposed above.

csv_lines <- c(
  "# model = redcard_within_model",
  "# method = sample (Default)",
  "#   sample",
  "#     num_samples = 1000 (Default)",
  "#     num_warmup = 1000 (Default)",
  "#     save_warmup = 0 (Default)",
  "#     thin = 1 (Default)",
  "# output",
  "#   file = redcard_within-202103251619-1-16b40e.csv",
  "lp__,accept_stat__,stepsize__,treedepth__,n_leapfrog__,divergent__,energy__,beta.1,beta.2"
)
tmp <- tempfile(fileext = ".csv")
writeLines(csv_lines, tmp)
rstan:::read_csv_header(tmp)
#> Warning in rstan:::read_csv_header(tmp): NAs introduced by coercion to integer
#> range
#> [1] "lp__,accept_stat__,stepsize__,treedepth__,n_leapfrog__,divergent__,energy__,beta.1,beta.2"
#> attr(,"iter.count")
#> [1] NA
#> attr(,"lineno")
#> [1] 10

g <- function(f, comment.char = '#') {
  con <- file(f, 'r')
  niter <- 0
  iter.count <- NA
  save.warmup <- FALSE
  sample.count <- NA_integer_
  thin <- NULL
  while (length(input <- readLines(con, n = 1)) > 0) {
    niter <- niter + 1
    if (!grepl(comment.char, input)) break;
    if (grepl("# iter=",input))
      iter.count <- as.integer(gsub("# iter=","",input))
    if (grepl("#.*num_samples",input)){
      sample.count <- as.integer(gsub("[^0-9]*([0-9]*).*","\\1",input))
    }
    if (grepl("#.*num_warmup",input)){
      warmup.count <- as.integer(gsub("[^0-9]*([0-9]*).*","\\1",input))
    } else {
      warmup.count <- 0L
    }
    if (grepl("#[[:blank:]]*thin", input)){
      thin <- as.integer(gsub("[^0-9]*([0-9]*).*","\\1",input))
    }
    if (grepl("#.*save_warmup",input)){
      save.warmup <- !grepl("0",input)
    }
    if (grepl("#.*output_sample",input)){
      iter.count <- as.numeric(gsub("[^0-9]*([0-9]*).*","\\1",input))
    }
    
  }
  header <- input
  if(is.na(iter.count)){
    if(save.warmup)
      iter.count <- warmup.count + sample.count
    else
      iter.count <- sample.count
  }
  if(!is.null(thin)){
    iter.count <- iter.count %/% thin
  }
  attr(header, "iter.count") <- iter.count
  attr(header, "lineno") <- niter
  close(con)
  header
}

g(tmp)
#> [1] "lp__,accept_stat__,stepsize__,treedepth__,n_leapfrog__,divergent__,energy__,beta.1,beta.2"
#> attr(,"iter.count")
#> [1] 1000
#> attr(,"lineno")
#> [1] 10

RStan Version:

2.21.2

R Version:

R version 4.0.3 (2020-10-10)

Operating System:

Ubuntu 18.04

Start at the read_csv_header() entry point and reproduce the issue with the supplied csv_lines example, focusing on matching the thin metadata line. Done means a model or output filename containing "thin" no longer produces an NA iter.count, while the actual # thin header still yields the expected iteration count.