I'm aware that rand()
is not guaranteed to give the same output even when using the same seed in general, and that people who need this have to use StableRNGs.jl
. I'm wondering though whether there are cases where I should expect the same outputs? E.g. if I use the same Julia binary on the same OS would that be expected to yield the same numbers?
If someone wants to try, I'm on 1.10-rc2 on Windows 10:
julia> using Random
julia> Random.seed!(123)
TaskLocalRNG()
julia> rand()
0.521213795535383
it definitely should give the same output for the same seed (on the same version)
I think this even goes as far as being the same on the same architecture
julia> using Random
julia> Random.seed!(123)
TaskLocalRNG()
julia> rand()
0.521213795535383
julia> versioninfo()
Julia Version 1.11.0-DEV.822
Commit 528fefb66a7* (2023-11-12 12:24 UTC)
Platform Info:
OS: Linux (x86_64-pc-linux-gnu)
CPU: 24 × AMD Ryzen 9 7900X 12-Core Processor
WORD_SIZE: 64
LLVM: libLLVM-15.0.7 (ORCJIT, znver3)
Threads: 34 on 24 virtual cores
Environment:
JULIA_PKG_USE_CLI_GIT = true
the issue with relying on this to never change is that it may change between versions & architectures; generally though, for the same julia version it should give the same output.
Does it really depend on the arch though? How? Does it use FMA?
I'm pretty sure it shouldn't, but I don't think Julia actually guarantees it
Hard to tell: The reproducibility section only mentions Julia version changes
that section doesn't give a whole lot guarantees at all anyway
from the way the RNGs are currently implemented, it should be consistent across architectures
Julia just doesn't guarantee that to be the case
e.g. I'm not sure you're guaranteed to get the same result on a GPU, which has a different default RNG
Cool thanks - it sounds to me like I'm okay to share a notebook that includes versioninfo()
and Random.seed!(123)
in the expectation that someone would be able to get the same outputs.
Currently, in 1.9 and 1.10, specifying the exact Julia version is sufficient. However, IMO by doing this you are depending on internals (i.e. there is no guarantee that this will still work when you and the folks reproducing your code are on 1.12)
Sure, but that's expected. I'm just after having a historical record that says "this sample was drawn with Julia 1.10rc-2 on Windows 10, using the random seed 123" so that someone else can download the same Julia version and get the same results with the same seed.
Seems like that's what StableRNGs.jl is for
Yes as I said originally that's the correct solution, I just wondered under which circumstances I can go without it.
Or https://github.com/wildart/Xkcd221 for even stronger reproducibility guarantees
Last updated: Nov 22 2024 at 04:41 UTC