Thanks for the quick (detailed) response.
On 1/31/2025 12:23 PM, Martin Maechler wrote:
Kevin R Coombes
on Fri, 31 Jan 2025 11:48:33 -0500 writes:
> Hi,
> I have a package that has been in CRAN for years and is now failing
> checks because some of the output of a test script is differing on some
> machines in the fifth or sixth decimal place. I have managed to fix most
> of these issues (by using the "digits" argument in calls to "summary" to
> hide the differences). the only one that remains yields this R CMD check
> report:
> Comparing ‘testDiff.Rout’ to ‘testDiff.Rout.save’ ...52c52
> < 2.600e-06 1.328e-01 4.666e-01 1.060e+00 1.369e+00 1.091e+01
> ---
> > 0.000003 0.132800 0.466600 1.060000 1.369000 10.910000
> Here the digit-limited output is the same (to a human mathematician,
though not to a string-matching computer), but one machine has decided to report
the output in scientific notation.
I'm guessing you are slightly off here:
Almost surely it's *not* a difference in machine/platform but only in versions
of R-devel.
The original issues *were* platform dependent, where the fifth digit was
off-by-one on some platforms. You are correct that what is left is
because of your change in R-devel.
My guess comes from the fact that I've been the R core member
who committed this change to R-devel :
------------------------------------------------------------------------
r87625 | maechler | 2025-01-24 16:58:25 +0100 (Fri, 24 Jan 2025)
parametrize & improve accuracy in print(summary(<numbers>))
------------------------------------------------------------------------
a week ago.
... and BTW, if you look carefully, for the first entry, the new
output *is* slightly more accurate also in your example.
Certainly, it is more accurate. But less accurate (in the form of fewer
digits) was what I needed to solve the cross-platform issues in the
current release.
I agree that the switch from fixed point to
exponential/scientific format is "unlucky" in this case
[and even unnecessary: in this case, keeping fixed format and
showing one digit more (using the same amount of characters),
one could also have shown the 0.0000026 ...]
A workaround for you may be to set something like
options(scipen = 2) # default is 0
in your testDiff.R script before printing
All this is only in R-devel, the development version of R...
and I have contemplated to add more tweaks to
print.summaryDefault() for that upcoming version of R.
The fear stopping me to do more tweaking was that the
consequence could be even *more* (still small) changes in such
summary() printing output.
> Both versions were produced by a command equivalent to
> print(summary(x, digits = 4))
As you may (or may not ..;-) guess from the above commit message ("parametrize")
is that print(summary(<numbers>)) got a new argument zdigits.
Yes. But I can't use the new devel-version argument to fix the problem
in the current release...
So, instead of print(summary(x, digits = 4))
you can, from R version 4.5 (currently only in the development
version of R) on use
print(summary(x, digits = 4), digits = .., zdigits = ..)
but you could also --- already in current versions of R ---
tweak the output using
print(summary(x, digits = 4), digits = <n>)
where you can play to see if n=3 , n=4, or n=5
help you getting better results ..
... actually *not* 'digits = 4' for summary() at all,
but only the digits argument to print(.) {where you'll get a
the new 'zdigits' argument *additionally* in R-devel and future
R version's}.
> What is the best cross-platform way to ensure that the output gets printed in the
same format? Set "options(scipen=999)"?
That's clearly too extreme to generally,
(I mentioned `scipen = 2` earlier).
Alternatively, pass some argument to "print" as well as to "summary"?
yes, see above.
(The other alternative I am considering is to just delete the script from the
"tests" directory.)
(I don't think that would be a good idea, .. but rather a very b.. one)
Ah, but it may have helped someone who knew what they were talking about
to answer more rapidly to avoid the bad solution. Thanks again.
I will try the options(scipen) fix for now, and keep the new zdigts
argument in mind going forward.
> Thanks,
> Kevin
Best,
Martin
______________________________________________
R-package-devel@r-project.org mailing list
https://stat.ethz.ch/mailman/listinfo/r-package-devel
