Retrieve RNG seed in julia

Question

In julia, the seed for the global RNG can be set with

srand(SEED)

How can I retrieve the seed or current state of the global RNG, e.g. again at a later point?

The goal is to get the state of the RNG at any given point in time, and recreate it in a different session, without knowing the initial seed or all calls to RNG that happened in the meanwhile.

For example, R allows access to the current seed through

.Random.seed

I was hoping that an equivalent way would exist in julia.

Answer 1

Base.Random.RANDOM_SEED is your friend for getting the seed:

julia> srand(10)

julia> Base.Random.RANDOM_SEED
1-element Array{Uint32,1}:
 0x0000000a

julia> srand(1)

julia> Base.Random.RANDOM_SEED
1-element Array{Uint32,1}:
 0x00000001

julia> srand(0xF)

julia> Base.Random.RANDOM_SEED
1-element Array{Uint32,1}:
 0x0000000f

This isn't documented, but the source is easy enough to read. I'm not sure how to get the current state of the RNG, but it might be in the dSFMT module

Answer 2

You should get the seed like this

reinterpret(Int32, Base.Random.GLOBAL_RNG.seed)

Test:

julia> srand(123456789);

julia> reinterpret(Int32, Base.Random.GLOBAL_RNG.seed)
1-element Array{Int32,1}:
 123456789

For saving an restoring the full rng state you can do the simple thing and just store the whole Base.Random.GLOBAL_RNG object. An easy way would be using JLD package.

In my private package I do manually save/read the rng state to HDF5, see here.

EDIT: This of course a more explicit version of @IainDunning's answer

Answer 3

Using a specialized MersenneTwister with an explicit variable (instead of the hidden global one provided by the default random values functions), the functionality you require can be provided:

newmt = Base.Random.MersenneTwister(123)
randvec = [rand(newmt) for i=1:100]
# save state now
savestate = deepcopy(newmt.state)
randvec2 = [rand(newmt) for i=1:20]
# rewind state to old state
newmt.state = savestate
randvec3 = [rand(newmt) for i=1:20]
if randvec2==randvec3
    println("works!")
end

The deepcopy threw me off for a second there. Also, it would have been easier to access the global random generator state, but it might require ccalling libdSFMT library (see random.jl and dSFMT.jl in Base.