Batch label for tar_stan_mcmc_rep_summary

[mike-lawrence]

When I use a tar_stan_mcmc_rep_summary target for SBC, I see there's a .rep column but I'm not seeing a .batch column; is there a way to get a column indicating which batch a given chunk of the summary comes from?

[wlandau]

There is no .batch column or similar. Would that be useful? The batch just indicates how the computation was parallelized and has no intrinsic meaning as far as the results are concerned.

[mike-lawrence]

Actually, on further consideration, since each batch is associated with a single simulated dataset, and reps re-use this dataset, it could be useful to have a batch identifier after all. The use case for using reps at all (assuming using chains>=1) would be to explore the influence of different warmup strategies, in which case the use case for a batch identifier would be to account for possible weirdness of a specific dataset that will affect all associated reps.

[wlandau]

since each batch is associated with a single simulated dataset...

Each rep within batch gets its own dataset, and batches have multiple datasets.

https://github.com/wlandau/stantargets/blob/df816f34ce45dcf61a46295ad37db522520b6eb2/R/tar_stan_mcmc_rep.R#L122-L130

The purpose of batching is to define good targets that balance the overhead of scale vs vulnerability to invalidation. Otherwise, batches have no meaning.

library(targets)
library(stantargets)
tar_stan_example_file("a.stan")
  
tar_script({
library(stantargets)
tar_stan_mcmc_rep_summary(
  your_model,
  stan_files = "a.stan", 
  data = tar_stan_example_data(),
  batches = 1,
  reps = 2,
  stdout = R.utils::nullfile(),
  stderr = R.utils::nullfile()
)
})

tar_make(reporter = "silent")

str(tar_read(your_model_data, branches = 1))
#> List of 1
#>  $ :List of 2
#>   ..$ :List of 5
#>   .. ..$ n         : int 10
#>   .. ..$ x         : num [1:10] -1 -0.778 -0.556 -0.333 -0.111 ...
#>   .. ..$ y         : num [1:10] -0.145 1.156 0.675 0.46 0.639 ...
#>   .. ..$ true_beta : num -0.0748
#>   .. ..$ .join_data:List of 2
#>   .. .. ..$ beta : num -0.0748
#>   .. .. ..$ y_rep: num [1:10] -0.145 1.156 0.675 0.46 0.639 ...
#>   ..$ :List of 5
#>   .. ..$ n         : int 10
#>   .. ..$ x         : num [1:10] -1 -0.778 -0.556 -0.333 -0.111 ...
#>   .. ..$ y         : num [1:10] -1.32 -1.55 -2.06 -1.36 1.18 ...
#>   .. ..$ true_beta : num 0.713
#>   .. ..$ .join_data:List of 2
#>   .. .. ..$ beta : num 0.713
#>   .. .. ..$ y_rep: num [1:10] -1.32 -1.55 -2.06 -1.36 1.18 ...

^{Created on 2021-03-23 by the reprex package (v1.0.0)}

[mike-lawrence]

Ah! I clearly still had the wrong understanding, thanks for clarifying.

[mike-lawrence]

I think the help docs were what led me astray; under the entry for the batches argument it says "Number of batches. Each batch is a branch target that generates a dataset and runs the model reps times."

[wlandau]

Oh, right, that's definitely grammatically ambiguous. Now fixed.