Save trace in memory rather than files?

[osorensen]

BayesMallowsSMC2/src/parameter_tracer.cpp

Lines 8 to 22 in b5cf170

	ParameterTracer::ParameterTracer(bool trace, bool trace_latent, const std::string& trace_directory)
	: trace { trace }, trace_latent { trace_latent }, trace_directory { trace_directory } {
	if(trace) {
	#ifdef _WIN32
	// On Windows, mkdir takes only one argument
	int status = mkdir(trace_directory.c_str());
	#else
	// On POSIX systems, mkdir takes two arguments
	int status = mkdir(trace_directory.c_str(), 0777);
	#endif
	if (status != 0) {
	Rcpp::stop("Error creating trace directory.");
	}
	}
	}

It should potentially speed up the algorithm, and at the same time make it easier to post-process.

Some timing experiments below:

library(BayesMallowsSMC2)
set.seed(2)
system.time({
  mod <- compute_sequentially(
    partial_rankings,
    hyperparameters = set_hyperparameters(n_items = 5),
    smc_options = set_smc_options(n_particles = 100, max_rejuvenation_steps = 5)
  )
})
#>    user  system elapsed 
#>   1.419   0.038   1.459

set.seed(2)
system.time({
  mod <- compute_sequentially(
    partial_rankings,
    hyperparameters = set_hyperparameters(n_items = 5),
    smc_options = set_smc_options(n_particles = 100, max_rejuvenation_steps = 5,
                                  trace = TRUE, trace_latent = FALSE,
                                  trace_directory = paste0(tempdir(), "trace1"))
  )
})
#>    user  system elapsed 
#>   1.494   0.244   1.946


set.seed(2)
system.time({
  mod <- compute_sequentially(
    partial_rankings,
    hyperparameters = set_hyperparameters(n_items = 5),
    smc_options = set_smc_options(n_particles = 100, max_rejuvenation_steps = 5,
                                  trace = TRUE, trace_latent = TRUE,
                                  trace_directory = paste0(tempdir(), "trace2"))
  )
})
#>    user  system elapsed 
#>   2.571   6.438  21.058

^{Created on 2025-04-10 with reprex v2.1.1}

[osorensen]

Another wow!!

library(BayesMallowsSMC2)
set.seed(2)
system.time({
  mod <- compute_sequentially(
    partial_rankings,
    hyperparameters = set_hyperparameters(n_items = 5),
    smc_options = set_smc_options(n_particles = 100, max_rejuvenation_steps = 5)
  )
})
#>    user  system elapsed 
#>   1.369   0.032   1.403
length(mod$alpha_traces)
#> [1] 0
length(mod$latent_rankings_traces)
#> [1] 0

set.seed(2)
system.time({
  mod <- compute_sequentially(
    partial_rankings,
    hyperparameters = set_hyperparameters(n_items = 5),
    smc_options = set_smc_options(n_particles = 100, max_rejuvenation_steps = 5,
                                  trace = TRUE, trace_latent = FALSE)
  )
})
#>    user  system elapsed 
#>   1.362   0.022   1.385
length(mod$alpha_traces)
#> [1] 100
length(mod$latent_rankings_traces)
#> [1] 0

set.seed(2)
system.time({
  mod <- compute_sequentially(
    partial_rankings,
    hyperparameters = set_hyperparameters(n_items = 5),
    smc_options = set_smc_options(n_particles = 100, max_rejuvenation_steps = 5,
                                  trace = TRUE, trace_latent = TRUE)
  )
})
#>    user  system elapsed 
#>   1.383   0.025   1.407
length(mod$alpha_traces)
#> [1] 100
length(mod$latent_rankings_traces)
#> [1] 100

^{Created on 2025-04-11 with reprex v2.1.1}