Avoid hardcoding 200 updates per chain (#169)

Author: penelopeysm

Project.toml

@@ -3,7 +3,7 @@ uuid = "80f14c24-f653-4e6a-9b94-39d6b0f70001"
 keywords = ["markov chain monte carlo", "probabilistic programming"]
 license = "MIT"
 desc = "A lightweight interface for common MCMC methods."
-version = "5.7.0"
+version = "5.7.1"
 
 [deps]
 BangBang = "198e06fe-97b7-11e9-32a5-e1d131e6ad66"

src/logging.jl

@@ -6,6 +6,8 @@ internally take for single-chain sampling.
 """
 abstract type AbstractProgressKwarg end
 
+DEFAULT_N_UPDATES = 200
+
 """
     CreateNewProgressBar
 
@@ -27,7 +29,7 @@ end
 function finish_progress!(p::CreateNewProgressBar)
     ProgressLogging.@logprogress p.name "done" _id = p.uuid
 end
-get_n_updates(::CreateNewProgressBar) = 200
+get_n_updates(::CreateNewProgressBar) = DEFAULT_N_UPDATES
 
 """
     NoLogging
@@ -38,7 +40,7 @@ struct NoLogging <: AbstractProgressKwarg end
 init_progress!(::NoLogging) = nothing
 update_progress!(::NoLogging, ::Any) = nothing
 finish_progress!(::NoLogging) = nothing
-get_n_updates(::NoLogging) = 200
+get_n_updates(::NoLogging) = DEFAULT_N_UPDATES
 
 """
     ExistingProgressBar
@@ -72,7 +74,7 @@ end
 function finish_progress!(p::ExistingProgressBar)
     ProgressLogging.@logprogress p.name "done" _id = p.uuid
 end
-get_n_updates(::ExistingProgressBar) = 200
+get_n_updates(::ExistingProgressBar) = DEFAULT_N_UPDATES
 
 """
     ChannelProgress

src/sample.jl

@@ -452,13 +452,12 @@ function mcmcsample(
             progress_channel = Channel{Bool}(nchains)
             overall_progress_bar = CreateNewProgressBar(progressname)
             # If we have many chains and many samples, we don't want to force
-            # each chain to report back to the main thread for each sample, as
-            # this would cause serious performance issues due to lock conflicts.
-            # In the overall progress bar we only expect 200 updates (i.e., one
-            # update per 0.5%). To avoid possible throttling issues we ask for
-            # twice the amount needed per chain, which doesn't cause a real
-            # performance hit.
-            updates_per_chain = max(1, 400 ÷ nchains)
+            # each chain to report back to the main thread for each sample, as this would
+            # cause serious performance issues due to lock conflicts. In the overall
+            # progress bar we only expect N updates (by default N = 200, i.e., one update
+            # per 0.5%). To avoid possible throttling issues we ask for twice
+            # the amount needed per chain, which doesn't cause a real performance hit.
+            updates_per_chain = max(1, (2 * get_n_updates(overall_progress_bar)) ÷ nchains)
             init_progress!(overall_progress_bar)
         end
         if progress == :perchain
@@ -483,7 +482,7 @@ function mcmcsample(
                     Ntotal = nchains * updates_per_chain
                     # Determine threshold values for progress logging
                     # (one update per 0.5% of progress)
-                    threshold = Ntotal / 200
+                    threshold = Ntotal / get_n_updates(overall_progress_bar)
                     next_update = threshold
 
                     itotal = 0
@@ -633,7 +632,7 @@ function mcmcsample(
             overall_progress_bar = CreateNewProgressBar(progressname)
             init_progress!(overall_progress_bar)
             # See MCMCThreads method for the rationale behind updates_per_chain.
-            updates_per_chain = max(1, 400 ÷ nchains)
+            updates_per_chain = max(1, (2 * get_n_updates(overall_progress_bar)) ÷ nchains)
             child_progresses = [
                 ChannelProgress(progress_channel, updates_per_chain) for _ in 1:nchains
             ]
@@ -646,9 +645,8 @@ function mcmcsample(
                 # This task updates the progress bar
                 Distributed.@async begin
                     # Determine threshold values for progress logging
-                    # (one update per 0.5% of progress)
                     Ntotal = nchains * updates_per_chain
-                    threshold = Ntotal / 200
+                    threshold = Ntotal / get_n_updates(overall_progress_bar)
                     next_update = threshold
 
                     itotal = 0