I think the problem is that, even if the user were to set.seed(), it will have different consequences depending on whether DelayedArray is already loaded, or not yet loaded. DelayedArray gets loaded in some way that is not transparent to the user, as a dependency-of-a-dependency-of-an annotation package.
I guess an acceptable solution would be for DelayedArray to remember and restore the random number seed before creating a BiocParallel param, with an edge case being that .Random.seed is NULL in a new R session. Martin On 5/22/19, 9:57 AM, "Kasper Daniel Hansen" <[email protected]> wrote: Why don't you let this be under the user's control and not do this at all. People should know that reproducibility of random numbers requires setting the seed, but that is best done by the user and not a package author. On Wed, May 22, 2019 at 9:30 AM Steffi Grote <[email protected]> wrote: Hi all, I tried to circumvent the problem by adding an optional seed as parameter like this: my_fun = function(..., seed = NULL){ code that might change the RNG if (!is.null(seed)){ set.seed(seed) } code that runs permutations } which solves the reproducibility issue, but gives me a Warning in BiocCheck: * WARNING: Remove set.seed usage in R code Found in R/ directory functions: my_fun() What is the best way to deal with this? Thanks in advance, Steffi > On April 12, 2019 at 1:10 AM Martin Morgan <[email protected]> wrote: > > > That easy strategy wouldn't work, for instance two successive calls to MulticoreParam() would get the same port assigned, rather than the contract of a 'random' port in a specific range; the port can be assigned by the manager.port= argument if the user wants to avoid random assignment. I could maintain a separate random number stream in BiocParallel for what amounts to a pretty trivial and probably dubious strategy [choosing random ports in hopes that one is not in use], but that starts to sound like a more substantial feature. > > Martin > > On 4/11/19, 7:06 PM, "Pages, Herve" <[email protected]> wrote: > > Hi Steffi, > > Any code that gets called between your calls to set.seed() and runif() > could potentially use the random number generator. So the sequence > set.seed(123); runif(1) is only guaranteed to be deterministic if no > other code is called in between, or if the code called in between does > not use the random number generator (but if that code is not under your > control it could do anything). > > @Martin: I'll look at your suggestion for DelayedArray. An easy > workaround would be to avoid changing the RNG state in BiocParallel by > having .snowPort() make a copy of .Random.seed (if it exists) before > calling runif() and restoring it on exit. > > H. > > On 4/11/19 15:25, Martin Morgan wrote: > > This is actually from a dependency DelayedArray which, on load, calls DelayedArray::setAutoBPPARAM, which calls BiocParallel::MulticoreParam(), which uses the random number generator to select a random port for connection. > > > > A different approach would be for DelayedArray to respect the user's configuration and use bpparam(), or perhaps look at the class of bpparam() and tell the user they should, e.g., BiocParallel::register(SerialParam()) if that's appropriate, or use registered("MulticoreParam") or registered("SerialParam") if available (they are by default) rather than creating an ad-hoc instance. > > > > Martin > > > > On 4/11/19, 10:17 AM, "Bioc-devel on behalf of Steffi Grote" <[email protected] on behalf of [email protected]> wrote: > > > > Hi all, > > I found out that example code for my package GOfuncR yields a different result the first time it's executed, despite setting a seed. All the following executions are identical. > > It turned out that loading the database package 'Homo.sapiens' changed the random numbers: > > > > set.seed(123) > > runif(1) > > # [1] 0.2875775 > > > > set.seed(123) > > suppressWarnings(suppressMessages(require(Homo.sapiens))) > > runif(1) > > # [1] 0.7883051 > > > > set.seed(123) > > runif(1) > > # [1] 0.2875775 > > > > Is that known or expected behaviour? > > Should I not load a package inside a function that later uses random numbers? > > > > Thanks in advance, > > Steffi > > > > _______________________________________________ > > [email protected] mailing list > > https://urldefense.proofpoint.com/v2/url?u=https-3A__stat.ethz.ch_mailman_listinfo_bioc-2Ddevel&d=DwIGaQ&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=8XXamcpEeef966i7IGk_3aE9GMJodKAzXwWW4fL_hrI&s=KoHGLM0HbP4whRZLG4ol66_q1qkg9E0LjFHObDqgNuo&e= <https://urldefense.proofpoint.com/v2/url?u=https-3A__stat.ethz.ch_mailman_listinfo_bioc-2Ddevel&d=DwIGaQ&c=eRAMFD45gAfqt84VtBcfhQ&r=BK7q3XeAvimeWdGbWY_wJYbW0WYiZvSXAJJKaaPhzWA&m=8XXamcpEeef966i7IGk_3aE9GMJodKAzXwWW4fL_hrI&s=KoHGLM0HbP4whRZLG4ol66_q1qkg9E0LjFHObDqgNuo&e=> > > > > -- > Hervé Pagès > > Program in Computational Biology > Division of Public Health Sciences > Fred Hutchinson Cancer Research Center > 1100 Fairview Ave. N, M1-B514 > P.O. Box 19024 > Seattle, WA 98109-1024 > > E-mail: [email protected] > Phone: (206) 667-5791 > Fax: (206) 667-1319 > > _______________________________________________ [email protected] mailing list https://stat.ethz.ch/mailman/listinfo/bioc-devel -- Best, Kasper _______________________________________________ [email protected] mailing list https://stat.ethz.ch/mailman/listinfo/bioc-devel
