Skip to content

Latest commit

 

History

History
68 lines (50 loc) · 2.6 KB

output-reports.md

File metadata and controls

68 lines (50 loc) · 2.6 KB
Raw
CurrentModule = Pigeons

[Interpreting pigeons' standard output](@id output-reports)

During the execution of parallel tempering, interim diagnostics can be computed and printed to standard out at the end of every iteration (this can be disabled using show_report = false):

using Pigeons

pigeons(target = toy_mvn_target(100))
nothing # hide

The functions called to emit each of these can be found at all_reports(). Some key quantities:

  • Λ: the global communication barrier, as described in Syed et al., 2021 and estimated using the sum of rejection estimator analyzed in the same reference. Syed et al., 2021 also developed a rule of thumb to configure the number of chains: PT should be set to roughly 2Λ.
  • time and allc: the time (in second) and allocation (in bytes) used in each round.
  • log(Z₁/Z₀): the stepping_stone() estimator for the log of the normalization constant, see [the documentation page on approximation of the normalization constant](@ref output-normalization).
  • min(α) and mean(α): minimum and average swap acceptance rates over the PT chains.

Additional statistics can be shown when more [recorders](@ref collecting-statistics) are added. For example, to accumulate other constant-memory summary statistics:

pigeons(target = toy_mvn_target(100), record = record_online(), explorer = AutoMALA())
nothing # hide
  • max|ρ| and mean|ρ|: maximum and average (across chains) correlation of the random variables L^t_i = V(X_i) and L^{t+1}_i = V(X_i) where V = \log \pi_N / \pi_1, X_i \sim \pi_{\beta_i}, and t, t+1 are indices just before and after a call to step!().
  • min(αₑ) and mean(αₑ): minimum and average (across chains) of the explorer's acceptance rates.

Programmatic access

The tables described above can also be accessed as a DataFrame via:

using Pigeons
pt = pigeons(target = toy_mvn_target(100))
pt.shared.reports.summary

Detailed statistics can be accessed via these DataFrames. For example to obtain mean swap acceptance probabilities for each round and pair of communicating chains, use:

first(pt.shared.reports.swap_prs, 20)

Creating a plot from this:

using StatsPlots
plotlyjs()
my_plot =  @df pt.shared.reports.swap_prs StatsPlots.plot(:round, :mean, group = :first)
StatsPlots.savefig(my_plot, "swap_prs.html"); 
nothing # hide

<iframe src="../swap_prs.html" style="height:500px;width:100%;"></iframe>