Make density and mixeddensity unique about mcmcchains.jl HOT 15 CLOSED

mixeddensity allows showing a histogram when chains are discrete.

from mcmcchains.jl.

Thanks for the info. In that case, I wonder if something like :pooleddensity would be useful. If not, I think we can close this issue.

from mcmcchains.jl.

Currently I believe what you're talking about exists. Try plot(chn, colordim = :parameter) to plot pooled densities. Does this do what you've asking about?

from mcmcchains.jl.

Doesn't just plot(chn) provide that?

from mcmcchains.jl.

plot(chn) gives you a plot for each parameter with all chains plotting for the same parameter, while plot(chn, colordim=:parameter) gives you one plot per chain and throws all the parameters on that one plot. The README has this example:

Summarize parameters	Summarize chains
plot(chn; colordim = :chain)	plot(chn; colordim = :parameter)

from mcmcchains.jl.

Sorry for the confusion. Just to clarify my initial post, I would like to do something that is similar to density(chn[:Param1].value[:]) while maintaining the labels and settings that are used in the builtin plot functions. This would pool the samples across chains for a given parameter. As far as I can tell, none of the options discussed so far have this functionality.

from mcmcchains.jl.

Oh, try indexing the chain first if you only want one parameter. plot(chn[:Param1], colordim=:parameter). I think it's possible that plot(chn, :Param1, colordim=:parameter) might also work, but I'm not at a good machine to test that out right now.

from mcmcchains.jl.

Unfortunately, that does not work. It generates a subplot for each chain. Here is a simple example based on the Turing docs:

using Distributed,StatsPlots
addprocs(4)
@everywhere using Turing, Random

@everywhere @model gdemo(x) = begin
    s ~ InverseGamma(2,3)
    m ~ Normal(0, sqrt(s))
    for i in eachindex(x)
        x[i] ~ Normal(m, sqrt(s))
    end
end

num_chains = 4
sampler = NUTS(1000, 0.65)
model = gdemo([1.2, 3.5])

chns = reduce(chainscat, pmap(x->sample(model,sampler),1:num_chains))
plot(chns[:m], colordim=:parameter)

The other issue is that there is no way to generate this type of plot for all parameters. For example, plot(chn,seriestype=:pooleddensity)

from mcmcchains.jl.

I guess I'm maybe not sure exactly what type of graph you're looking for. Here's the plotting output from the above code:

Overlay all the chains on top of eachother, with a plot per parameter:

density(chns[:m])

# As above, but for all parameters.
density(chns)

Overlay all parameters per chain together:

density(chns, colordim=:parameter, legend=true)

Which one of the above provides the functionality you'd like to see in mixeddensity?

from mcmcchains.jl.

Sorry. Its clear to me now that I should have provided a more explicit example. I was thinking that something like plot(chn,seriestype=:pooleddensity) would be equivalent to

chns = reduce(chainscat, pmap(x->sample(model,sampler),1:num_chains))
p1 = density(chns[:m].value[:], title="m",ylabel="Density",xlabel="Sample value",leg=false)
p2 = density(chns[:s].value[:], title="s",ylabel="Density",xlabel="Sample value",leg=false)
p = plot(p1,p2,layout=(2,1))

The basic idea is that you can make smoother density plots by pooling the chain samples for a given parameter.

from mcmcchains.jl.

Ah, I understand now. I think I can put that in for you.

from mcmcchains.jl.

Thanks and sorry again for the confusion. I'll provide better examples in the future.

from mcmcchains.jl.

Another way of doing this is using the DataFrame route:

...
chns = reduce(chainscat, pmap(x->sample(model,sampler),1:num_chains))
df = DataFrame(chns, [:parameters])
p = Vector{Plots.Plot{Plots.GRBackend}}(undef, size(df, 2))
for (i, par) in enumerate(names(df))
  p[i] =density(df[par], lab="$par", xlab="Sample value", ylab="Density")
end
plot(p..., layout=(size(df, 2),1))

chris_test_2.pdf

Above snippet works because by default the DataFrame constructor appends chains. This also applies to the Array constructor, but in this case the list of names (names(df)) is handy.

from mcmcchains.jl.

Thanks and sorry again for the confusion. I'll provide better examples in the future.

No problem! I'm glad we got this. I'll try and get this in this weekend. Does anyone have an issue with me adding a new keyword to the plot functions called append_chains which concatenates the chains together? I only say append_chains as the keyword since that's the convention in #79.

from mcmcchains.jl.

Since pooleddensity has been added, I'll go ahead and close this issue.

from mcmcchains.jl.