[Rd] R 4.1.x make check fails, stats-Ex.R, step factor reduced below minFactor
Andrew Piskorski
@tp @end|ng |rom p|@kor@k|@com
Fri Oct 1 18:05:44 CEST 2021
On Fri, Oct 01, 2021 at 03:45:48PM +0200, Martin Maechler wrote:
> Is there anything special (system libraries, compilers, ..)
> on your platform?
No. As far as I know this is an ordinary SuperMicro x86-64 server,
nothing strange or unusual. /proc/cpuinfo says "Intel(R) Xeon(R) CPU
E5-2670 0 @ 2.60GHz".
> o2 <- options(digits = 10) # more accuracy for 'trace'
> ## central differencing works here typically (PR#18165: not converging on *some*):
> ctr2 <- nls.control(nDcentral=TRUE, tol = 8e-8, # <- even smaller than above
> warnOnly = (grepl("^aarch64.*linux", R.version$platform) && grepl("^NixOS", osVersion)
> ))
> (nlm2 <- update(nlmod, control = ctr2, trace = TRUE)); options(o2)
>
> ... now would that run w/o error on your Ubuntu-installed R ?
Interactively, the code above runs fine. In fact, the original code
ALSO seems to run fine, no errors at all! See output below. I get
the error when running the tests via either "make check" or
tools::testInstalledPackages(scope="base"), but outside of that
testing framework it runs fine.
Ah, interactively, if I ALSO run the code for the immediately prior
test in stats-Ex.R, THEN the nlm2 code fails the same way as with
"make check". That prior test does set.seed(27), which seems to
trigger the downstream failures. Simply skipping the set.seed(27)
(interactively) makes the failure go away for me. But if the
set.seed(27) is necessary, maybe the second test should be doing its
own set.seed() of some sort?
I don't know how/where to comment out that set.seed(27) to try running
tests without it. Editing "src/library/stats/man/nls.Rd" and re-running
"make check" still does the set.seed(27).
Just run code from the single failing test, it works fine:
------------------------------------------------------------
## R --vanilla
R version 4.1.1 Patched (2021-09-21 r80946) -- "Kick Things"
Copyright (C) 2021 The R Foundation for Statistical Computing
Platform: x86_64-pc-linux-gnu/x86_64 (64-bit)
> Sys.setenv(LC_COLLATE = "C", LC_TIME = "C", LANGUAGE = "en")
> options(digits = 10)
> x <- -(1:100)/10
> y <- 100 + 10 * exp(x / 2) + rnorm(x)/10
> nlmod <- nls(y ~ Const + A * exp(B * x))
Warning message:
In nls(y ~ Const + A * exp(B * x)) :
No starting values specified for some parameters.
Initializing 'Const', 'A', 'B' to '1.'.
Consider specifying 'start' or using a selfStart model
> nlm1 <- update(nlmod, control = list(tol = 1e-7))
Warning message:
In nls(formula = y ~ Const + A * exp(B * x), algorithm = "default", :
No starting values specified for some parameters.
Initializing 'Const', 'A', 'B' to '1.'.
Consider specifying 'start' or using a selfStart model
> nlm2 <- update(nlmod, control = list(tol = 8e-8, nDcentral=TRUE), trace=TRUE)
1017400.445 (4.11e+02): par = (1 1 1)
752239.9094 (1.96e+02): par = (13.41553998 1.959746504 0.01471383253)
668978.9926 (1.65e+02): par = (189.3774772 -162.3882591 1.397507535)
375910.4745 (1.20e+02): par = (167.1787529 -119.9960435 1.42386803)
93230.26788 (5.49e+01): par = (133.8879258 -56.45697809 1.498055399)
382.9221937 (2.42e+00): par = (100.6364489 6.806405333 1.84811172)
138.7915397 (9.68e+00): par = (100.6763251 6.489793899 0.7564107501)
24.47843640 (5.42e+00): par = (100.4024547 8.003646622 0.4918079622)
0.8056918383 (4.49e-03): par = (99.9629562 10.01549373 0.4913706525)
0.8056755692 (4.09e-06): par = (99.96295295 10.01549135 0.4914577719)
0.8056755692 (7.83e-09): par = (99.96295344 10.01549217 0.4914579487)
Warning message:
In nls(formula = y ~ Const + A * exp(B * x), algorithm = "default", :
No starting values specified for some parameters.
Initializing 'Const', 'A', 'B' to '1.'.
Consider specifying 'start' or using a selfStart model
> nlm1
Nonlinear regression model
model: y ~ Const + A * exp(B * x)
data: parent.frame()
Const A B
99.9629534 10.0154922 0.4914579
residual sum-of-squares: 0.8056756
Number of iterations to convergence: 10
Achieved convergence tolerance: 1.586349e-08
> nlm2
Nonlinear regression model
model: y ~ Const + A * exp(B * x)
data: parent.frame()
Const A B
99.9629534 10.0154922 0.4914579
residual sum-of-squares: 0.8056756
Number of iterations to convergence: 10
Achieved convergence tolerance: 7.832984e-09
Instead run BOTH these tests, now the last one fails:
------------------------------------------------------------
## R --vanilla
R version 4.1.1 Patched (2021-09-21 r80946) -- "Kick Things"
Copyright (C) 2021 The R Foundation for Statistical Computing
Platform: x86_64-pc-linux-gnu/x86_64 (64-bit)
## The two examples below show that you can fit a model to
## artificial data with noise but not to artificial data
## without noise.
> x <- 1:10
> y <- 2*x + 3 # perfect fit
## terminates in an error, because convergence cannot be confirmed:
> try(nls(y ~ a + b*x, start = list(a = 0.12345, b = 0.54321)))
> Error in nls(y ~ a + b * x, start = list(a = 0.12345, b = 0.54321)) :
number of iterations exceeded maximum of 50
## adjusting the convergence test by adding 'scaleOffset' to its denominator RSS:
> nls(y ~ a + b*x, start = list(a = 0.12345, b = 0.54321),
control = list(scaleOffset = 1, printEval=TRUE))
> + It. 1, fac= 1, eval (no.,total): ( 1, 1): new dev = 1.05935e-12
Nonlinear regression model
model: y ~ a + b * x
data: parent.frame()
a b
3 2
residual sum-of-squares: 1.059e-12
Number of iterations to convergence: 1
Achieved convergence tolerance: 3.639e-07
> ## Alternatively jittering the "too exact" values, slightly:
set.seed(27)
> yeps <- y + rnorm(length(y), sd = 0.01) # added noise
> nls(yeps ~ a + b*x, start = list(a = 0.12345, b = 0.54321))
Nonlinear regression model
model: yeps ~ a + b * x
data: parent.frame()
a b
3.001 2.000
residual sum-of-squares: 0.001346
Number of iterations to convergence: 2
Achieved convergence tolerance: 8.658e-09
>
> ## the nls() internal cheap guess for starting values can be sufficient:
> x <- -(1:100)/10
> y <- 100 + 10 * exp(x / 2) + rnorm(x)/10
> nlmod <- nls(y ~ Const + A * exp(B * x))
Warning message:
In nls(y ~ Const + A * exp(B * x)) :
No starting values specified for some parameters.
Initializing 'Const', 'A', 'B' to '1.'.
Consider specifying 'start' or using a selfStart model
> options(digits = 10) # more accuracy for 'trace'
> try(nlm1 <- update(nlmod, control = list(tol = 1e-7))) # where central diff. work here:
Warning message:
In nls(formula = y ~ Const + A * exp(B * x), algorithm = "default", :
No starting values specified for some parameters.
Initializing 'Const', 'A', 'B' to '1.'.
Consider specifying 'start' or using a selfStart model
> (nlm2 <- update(nlmod, control = list(tol = 8e-8, nDcentral=TRUE), trace=TRUE))
1017460.306 (4.15e+02): par = (1 1 1)
758164.7503 (2.34e+02): par = (13.42031396 1.961485 0.05947543745)
269506.3538 (3.23e+02): par = (51.75719816 -13.09155957 0.8428607709)
68969.21893 (1.03e+02): par = (76.0006985 -1.935226745 1.0190858)
633.3672230 (1.29e+00): par = (100.3761515 8.624648402 5.104490259)
151.4400218 (9.39e+00): par = (100.6344391 4.913490985 0.2849209569)
53.08739850 (7.24e+00): par = (100.6830407 6.899303317 0.4637755074)
1.344478640 (5.97e-01): par = (100.0368306 9.897714142 0.5169294939)
0.9908415909 (1.55e-02): par = (100.0300625 9.9144191 0.5023516843)
0.9906046057 (1.84e-05): par = (100.0288724 9.916224018 0.5025207336)
0.9906046054 (9.95e-08): par = (100.028875 9.916228366 0.50252165)
0.9906046054 (9.93e-08): par = (100.028875 9.916228366 0.50252165)
Error in nls(formula = y ~ Const + A * exp(B * x), algorithm = "default", :
step factor 0.000488281 reduced below 'minFactor' of 0.000976562
In addition: Warning message:
In nls(formula = y ~ Const + A * exp(B * x), algorithm = "default", :
No starting values specified for some parameters.
Initializing 'Const', 'A', 'B' to '1.'.
Consider specifying 'start' or using a selfStart model
------------------------------------------------------------
--
Andrew Piskorski <atp using piskorski.com>
More information about the R-devel
mailing list