Avoiding numerical issues in show for problematic chains

[torfjelde]

#401 is dope stuff (thanks @sethaxen ) but the tail ess in many cases seem to cause errors upon trying upon calling show, e.g. I just got the following trying a model

DivideError: integer division error

Stacktrace:
  [1] div
    @ ./int.jl:288 [inlined]
  [2] divgcd(x::Int64, y::Int64)
    @ Base ./rational.jl:45
  [3] //
    @ ./rational.jl:73 [inlined]
  [4] /(x::Rational{Int64}, y::Rational{Int64})
    @ Base ./rational.jl:346
  [5] _ess_rhat(::Val{:basic}, chains::Array{Real, 3}; autocov_method::AutocovMethod, split_chains::Int64, maxlag::Int64)
    @ MCMCDiagnosticTools ~/.julia/packages/MCMCDiagnosticTools/oVEPy/src/ess_rhat.jl:520
  [6] #_ess#36
...

This is for a problematic chain of course (I believe the variance is very high?). I also ran into a bug somewhere earlier, again, in the ESS computation.

Maybe we should provide a way to convert these errors into warnings in MCMCChains.describe?

MCMCChains.jl/src/stats.jl

Lines 286 to 331 in ddac60f

	function summarystats(
	chains::Chains;
	sections = _default_sections(chains),
	append_chains::Bool = true,
	autocov_method::MCMCDiagnosticTools.AbstractAutocovMethod = AutocovMethod(),
	maxlag = 250,
	kwargs...
	)
	# Store everything.
	funs = [mean∘cskip, std∘cskip]
	func_names = [:mean, :std]
	# Subset the chain.
	_chains = Chains(chains, _clean_sections(chains, sections))
	# Calculate MCSE and ESS/R-hat separately.
	mcse_df = MCMCDiagnosticTools.mcse(
	_chains; sections = nothing, autocov_method = autocov_method, maxlag = maxlag,
	)
	ess_rhat_rank_df = MCMCDiagnosticTools.ess_rhat(
	_chains; sections = nothing, autocov_method = autocov_method, maxlag = maxlag, kind=:rank
	)
	ess_tail_df = MCMCDiagnosticTools.ess(
	_chains; sections = nothing, autocov_method = autocov_method, maxlag = maxlag, kind=:tail
	)
	nt_additional = (
	mcse=mcse_df.nt.mcse,
	ess_bulk=ess_rhat_rank_df.nt.ess,
	ess_tail=ess_tail_df.nt.ess,
	rhat=ess_rhat_rank_df.nt.rhat,
	ess_per_sec=ess_rhat_rank_df.nt.ess_per_sec,
	)
	additional_df = ChainDataFrame("Additional", nt_additional)
	# Summarize.
	summary_df = summarize(
	_chains, funs...;
	func_names = func_names,
	append_chains = append_chains,
	additional_df = additional_df,
	name = "Summary Statistics",
	sections = nothing
	)
	return summary_df
	end

Maybe just add try-catch statements to all of these and do @warn + NaN in case of failure?

[devmotion]

At first glance (haven't checked anything), it seems we should just fix the implementation in MCMCDiagnosticTools?

[sethaxen]

Agreed this should be fixed upstream. @torfjelde can you by any chance share an MWE?

[torfjelde]

I don't have an MWE atm, but the model is the SDE example from https://turinglang.org/v0.24/tutorials/10-bayesian-differential-equations/ that I ran without adaptation 😕

[sethaxen]

Thanks @torfjelde , that should be enough for me to reproduce the issue when I get to a computer.

[sethaxen]

@torfjelde I just tried and wasn't able to reproduce this. Could you write your chains to a file and share them with me?

[sethaxen]

@torfjelde gentle reminder to send the problematic chains when you get the chance

[sethaxen]

@torfjelde slightly less gentle reminder to send the problematic chains

[torfjelde]

Haha yeah sorry. Basically my computer with the thingy crashed and so I couldn't access it until after I got back from BayesComp, and I've yet to be able to reproduce the example:/

[sethaxen]

I'm not certain what the original issue is, but if it's the same as #447, then MCMCDiagnosticTools v0.3.10 contains a fix for this: TuringLang/MCMCDiagnosticTools.jl#119