Summarize

Extract summary values

mean(chn) and summarystats(chn) return a ChainDataFrame with parameter names in the parameters column:

using MCMCChains
chn = Chains(reshape(1:120, 20, 3, 2), [:a, :b, :c])
summary = mean(chn)

Mean

  parameters      mean 
      Symbol   Float64 

           a   40.5000
           b   60.5000
           c   80.5000

Index by parameter and statistic to get a value or vector:

summary[:a, :mean]

40.5

summary[[:a, :c], :mean]

2-element Vector{Float64}:
 40.5
 80.5

Use : to keep all columns in a smaller table:

summary[[:a, :c], :]

Mean

  parameters      mean 
      Symbol   Float64 

           a   40.5000
           c   80.5000

Use Tables.jl to extract named columns, or convert to a DataFrame:

using Tables
Tables.columntable(summary).mean

3-element Vector{Float64}:
 40.5
 60.5
 80.5

using DataFrames
DataFrame(summary)

Array(summary) returns a vector for one statistic or a matrix for several statistics. Matrix rows follow summary[:, :parameters], and columns follow names(summary)[2:end].

convert(Array, summary) keeps the two-dimensional table layout even for one statistic.

The vector follows the parameter order in the chain. Sort the grouped chain first if you need natural index order:

grouped_chn = Chains(reshape(1:120, 20, 3, 2), ["x[2]", "x[1]", "y"])
grouped_mean = mean(sort(group(grouped_chn, :x)))
Array(grouped_mean)

2-element Vector{Float64}:
 60.5
 40.5

The conversion interface preserves the statistic dimension:

size(convert(Array, grouped_mean))

(2, 1)

API

The methods listed below are defined in src/summarize.jl.

ChainDataFrame

summarize(chains, funs...[; sections, func_names = [], name = "", append_chains = true])