# Setting up multiple chains in Turing.jl
**URL:** https://discourse.julialang.org/t/setting-up-multiple-chains-in-turing-jl/20806
**Category:** General Usage
**Created:** [February 15, 2019, 5:03am UTC](https://discourse.julialang.org/t/setting-up-multiple-chains-in-turing-jl/20806 "2019-02-15T05:03:34Z")
**Posts on this page:** 4
**Page:** 2
### Author:  [@Christopher\_Fisher](https://discourse.julialang.org/u/Christopher_Fisher)
#### Post date: [March 7, 2019, 9:18am UTC](https://discourse.julialang.org/t/setting-up-multiple-chains-in-turing-jl/20806/21 "2019-03-07T09:18:32Z")
> [@cpfiffer](#):
>
> using Distributed addprocs(4) @everywhere using Turing,Distributions @everywhere @model model() = begin a ~ Normal(-10,.1) b ~ Normal(-8,.1) c ~ Normal(-6,.1) d ~ Normal(-4,.1) return nothing end Nsamples = 2000 Nadapt = 1000 δ = .85 specs = NUTS(Nsamples,Nadapt,δ) chns = reduce(chainscat, pmap(x-\>sample(model(),specs),1:4)) subset = chns[Nadapt:Nsamples, :, :]
Thank you. This is indeed much simpler. It might be helpful to other users to add a pmap example to the documentation. Thanks again!
---
### Author:  [@goedman](https://discourse.julialang.org/u/goedman)
#### Post date: [March 7, 2019, 2:45pm UTC](https://discourse.julialang.org/t/setting-up-multiple-chains-in-turing-jl/20806/22 "2019-03-07T14:45:38Z")
Hi Cameron,
It’s not only simpler, but it actually works better! I used something similar to a few functions I am trying to add to TuringModels to redo the sections in a Chains object created by Turing. It seems to work, but at the end doing `chains[:a]` shows it is not correct. A MWE (well ‘minimum’?):
```julia
using Turing
function flatten_name_map(chn::MCMCChains.AbstractChains)
parms = values(chn.name_map)
[String(parms[i][j]) for i in 1:length(parms) for j in 1:length(parms[i])]
end
function construct_a3d(chn::MCMCChains.AbstractChains)
d, p, c = size(chn.value)
a3d = fill(0.0, d, p, c);
for (i, par) in enumerate(flatten_name_map(chn))
a3d[:, i, 1] = reshape(chn[par], d)
end
a3d
end
function move_parameters_to_new_section(chn::MCMCChains.AbstractChains,
new_section::Symbol, parameters_to_move::Vector{String})
a3d = construct_a3d(chn)
parms = values(chn.name_map.parameters)
moved_parameters = String[]
for par in parameters_to_move
(par in parms) ? append!(moved_parameters, [par]) : @warn "$par not in $parms, ignored"
end
return MCMCChains.Chains(a3d,
Symbol.(flatten_name_map(chn)),
Dict(
:parameters => Symbol.(filter(x -> !(x in parameters_to_move), parms)),
new_section => Symbol.(moved_parameters),
:internals => Symbol.(values(chn.name_map.internals))))
end
@model model() = begin
a ~ Normal(-10,.1)
b ~ Normal(-8,.1)
c ~ Normal(-6,.1)
d ~ Normal(-4,.1)
return nothing
end
Nsamples = 2000
Nadapt = 1000
draws = Nadapt+1 : Nsamples
δ = .85
sampler = NUTS(Nsamples, Nadapt, δ)
sampler = Turing.NUTS(2000, 1000, 0.65)
chn = sample(model(), sampler)
chn1 = move_parameters_to_new_section(chn, :pooled, ["b", "d", "zeta"])
describe(chn1)
describe(chn1, section=:pooled)
mean(chn1[:a])
```
You mind taking a look to spot my error? Or if this functionality is better achieved in some other way?
Thanks!
---
### Author:  [@cpfiffer](https://discourse.julialang.org/u/cpfiffer)
#### Post date: [March 7, 2019, 4:36pm UTC](https://discourse.julialang.org/t/setting-up-multiple-chains-in-turing-jl/20806/23 "2019-03-07T16:36:45Z")
I think I can serve this better on the back end by providing the functions `set_name_map(c::Chains, new_name_map::Dict)` and `set_name_map(c::Chains, new_name_map::NamedTuple)`. I can also call them `set_sections` if you’d prefer. People probably shouldn’t have to write all this stuff outside of the library to do basic stuff like this.
---
### Author:  [@goedman](https://discourse.julialang.org/u/goedman)
#### Post date: [March 7, 2019, 5:02pm UTC](https://discourse.julialang.org/t/setting-up-multiple-chains-in-turing-jl/20806/24 "2019-03-07T17:02:55Z")
Fully agreed. My intention was to test these and suggest it via a PR. Something like set\_sections or update\_sections all work fine. As an end user I tend to think in sections, as a developer probably more in terms of name\_map.
It is interesting that the longer I try to build packages the more I think that newer (good!) stuff like NamedTuples is on the developer side (the developer would probably use Parameters.jl and stuff like @unpack …), Dicts and DataFrames are end user. Just a side note.
In line with that, one other thought/suggestion. In StatisticalRethinking I have a function `to_df(chain::MCMCChains.Chains)` which converts a chain to a DataFrame. As an example, you did that great example in TuringExamples using training and test observations (that I hijacked for `04/clip-38s.jl`), it would be nice to continue using DataFrames containing the samples.
Right now I use:
```julia
function to_df(chain::MCMCChains.AbstractChains)
d, p, c = size(chain.value)
df = DataFrame()
a3d = zeros(d*c, p);
indx = 1
for name in Symbol.(keys(chain))
a3d[:, indx] = reshape(convert(Array{Float64}, chain[name]), d*c)
indx += 1
end
cnames = keys(chain)
snames = [Symbol(cnames[i]) for i in 1: length(cnames)]
DataFrame(a3d[:,:], snames)
end
```
[Previous page](https://discourse.julialang.org/t/setting-up-multiple-chains-in-turing-jl/20806.md?page=1)