Static branching within static branching

[alexpghayes]

I got my previous simulation to work (thank you!) and am now expanding it to include more models. First I statically branch to sample from 2 different models. Now, for each of those two models, I want to fit two different estimators. My first attempt is at https://github.com/alexpghayes/citation-impute-simulations/blob/main/_targets.R but I'm struggling to get the splitting within splitting.

If I use tidyr::expand(models, estimators) and a single tar_map() call, I end up sampling from each model for each estimator, which is not what I want.

[alexpghayes]

So I eventually got the computation graph I wanted in

with some metaprogramming. See https://github.com/alexpghayes/citation-impute-simulations/blob/unholy/_targets.R. But then targets errored on me during the make process.

Error : Target loss_symmetric_svd_model1 tried to branch over svd_model1_symmetric_svd, which is illegal. Patterns must only branch over explicitly declared targets in the pipeline. Stems and patterns are fine, but you cannot branch over branches or global objects. Also, if you branch over a target with format = "file", then that target must also be a pattern.Error: callr subprocess failed: Target loss_symmetric_svd_model1 tried to branch over svd_model1_symmetric_svd, which is illegal. Patterns must only branch over explicitly declared targets in the pipeline. Stems and patterns are fine, but you cannot branch over branches or global objects. Also, if you branch over a target with format = "file", then that target must also be a pattern.

Does this mean I've hit a fundamental limitation of targets? Is the solution here to use tidyr::expand_grid() to generate the full grid of models crossed with estimators from the beginning, and to just deal with the fact the duplication of the distributions?

[wlandau]

Did you mean to type map(distribution) instead of map(population) in the svd* targets? https://github.com/alexpghayes/citation-impute-simulations/pull/4 seems to fix the error on my end.

[alexpghayes]

Thank you!!

[liutiming]

Hi @alexpghayes, the link is broken. I think the others may also benefit from reading your use case if that is possible?

[liutiming]

@wlandau the nested tar_map just impressed me again about how extensible targets is 👍

[alexpghayes]

@liutiming The link is to an ongoing research project that I prefer not be public just yet. I'll make it public again once I've submitted my manuscript!