v0.3.11

v0.3.11 (2019-07-28)

Diff since v0.3.10

Closed issues:

• Release 0.3.10 (#105)

Merged pull requests:

• Added a Chains constructor for a vector-of-vectors (#114) (cpfiffer)
• Added mean function (#109) (cpfiffer)
• Support digits keyword for stats functions (#108) (cpfiffer)

v0.3.10

v0.3.10 (2019-05-21)

Full Changelog

Closed issues:

• lf_eps included as a parameter on Turing master (#101)
• Release 0.3.9 (#100)

Merged pull requests:

• Increment version number to 0.3.10 (#104) (cpfiffer)
• Remove Turing from dataframe/summarize tests (#103) (cpfiffer)
• Update chains.jl (#102) (dalejordan)

v0.3.9

• Corrected ESS calculation (#98, #86, #90)
• Added Rhat to the summarystats output (#98)

v0.3.8

• Fixed the size function (#92)
• The show(chain) function now prints quantiles in addition to summary stats.
• Fixed a bug where ChainDataFrame objects would not appear in Jupyter output cells.

DataFrames, hpd export.

• Fixes a quantile issue
• Exports the hpd function
• Uses DataFrames for all statistics functions

Quantiles Hotfix

• Fixes an improperly specified quantile function (#76, #75)

Sampling, Arrays, and DataFrames

• You can now use the sampling API from StatsBase to pull random samples from your chain (#68)
• Supports Array(chn) to create a simple Array object from a Chains object (#73)
• Export your chain to a DataFrame (#73)

The README file contains more information on how to use the above features, but here's the relevant section:

Exporting Chains

A few utility export functions have been provided to convert Chains objects to either an Array or a DataFrame:

# Several examples of creating an Array object:
Array(chns)
Array(chns[:s])
Array(chns, [:parameters])
Array(chns, [:parameters, :internals])

# By default chains are appended. This can be disabled
# using the append_chains keyword argument:
Array(chns, append_chains=false)

# This will return an `Array{Array, 1}` object containing
# an Array for each chain.

# A final option is:
Array(chns, remove_missing_union=false)

# This will not convert the Array columns from a 
`Union{Missing, Real}` to a `Vector{Real}`.

Similarly, for DataFrames:

DataFrame(chns)
DataFrame(chns[:s])
DataFrame(chns, [:parameters])
DataFrame(chns, [:parameters, :internals])
DataFrame(chns, append_chains=false)
DataFrame(chns, remove_missing_union=false)

See also ?MCMCChains.DataFrame and ?MCMCChains.Array for more help.

Sampling Chains

MCMCChains overloads several sample() methods as defined in StatsBase:

# Sampling `n` samples from the chain `a`. Optionally
# weighting the samples using `wv`.
sample([rng], a, [wv::AbstractWeights], n::Integer)

# E.g. creating 10000 weighted samples:
c = kde(Array(chn[:s]))
chn_weighted_sample = sample(c.x, Weights(c.density), 100000)

# As above, but supports replacing and ordering.
sample([rng], a, [wv::AbstractWeights], n::Integer; replace=true, 
  ordered=false)

See also ?MCMCChains.sample for additional help.

Hotfix

• Removed restriction on Chains.value which required a Union{Missing, A} matrix.
• Added lastindex functions
• Fixed leftover indexing by string functions from #57.

Index changes and bug fixes

• Diagnostic functions now support section = :section_name as keyword arguments (#54)
• Corner plots have been fixed to work with multiple chains (#55)
• chn[:param] and chn[:, :param, :] now both return the same thing, a Chains object containing only :param.
• Strings are now enforced as parameter names. Symbols no longer work as parameter names. (#57)
• Added get_params, which packages all the parameters in a chain into a NamedTuple. Used like get_params(chn). (#57)
• Added an overload on get accepting a keyword section, which can be a vector or a single symbol. Works like get(chn; section = :pooled) or get(chn; section = [:parameters, :pooled]). (#57)
• All variants of get and get_params accept a keyword argument flatten, which can be true or false. If flatten=true, the tuple returned will have a flat structure. This is handy if you want to go straight to a dataframe, because flatten=false (the default) will clump variables with the same names together (as with "P[1]", "P[2]", "P[3]"). (#57)
• Fixed bug where chains would not show if their length was less than 200 (c2d0827)
• Fixed bug where chains would not show if they did not contain a :parameters section

Plot labl improvements

• Fixed quantile deprecation warning
• Supported label overriding for plot
• Set default chain name to an integer