Error in { : task 14 failed - "unable to start png() device"

[gedkitchener-CT]

Project Robyn

Describe issue

Ran model with 5 trials - fine.

Ran again with 10 trials - I get the following error message:

>>> Exporting all charts into directory: C:/Users/.../2022-07-20 16.06 init...
>>> Plotting summary charts...
No spend-exposure modelling needed. all media variables used for mmm are spend variables 
>>> Plotting 171 Pareto optimum models on 6 cores...
  |                                                                                              |   0%Error in { : task 14 failed - "unable to start png() device"
In addition: Warning messages:
1: package ‘doRNG’ was built under R version 4.1.2 
2: package ‘foreach’ was built under R version 4.1.2 
 Error in { : task 14 failed - "unable to start png() device"

6. stop(simpleError(msg, call = expr)) 
5. e$fun(obj, substitute(ex), parent.frame(), e$data) 
4. structure(list(args = uniqueSol(.doRNG.stream = list(c(10407L, 
-76253865L, -1148296788L, -844762435L, 234093146L, -1972368365L, 
-1737907656L), c(10407L, -1793080333L, -348917575L, 1074043502L, 
494141499L, 1248036383L, 367895649L), c(10407L, -982643044L,  ... 
3. do.call("%dopar%", list(obj, ex), envir = parent.frame()) 
2. foreach(sid = uniqueSol) %dorng% {
    plotMediaShareLoop <- plotMediaShare[solID == sid]
    rsq_train_plot <- plotMediaShareLoop[, round(unique(rsq_train), 
        4)] ... at model.R#501
1. robyn_run(InputCollect = InputCollect, plot_folder = robyn_object, 
    pareto_fronts = 3, plot_pareto = TRUE)

Extremely frustrating as this was training for 5+ hours and seemed to be fine - but as a result of a plotting error, the whole model gets discarded! To use an earlier analogy, this feels like baking a beautiful cake, but because my camera can't fit the whole cake in shot, the cake just vanishes!

What does this error mean? I can't make sense of the output to know what's happening (i.e., what part of the loop in step 2 is causing the error?)

[laresbernardo]

Hi @gedkitchener-CT
I'm really sorry to hear that! We did split robyn_run(..., export = FALSE) and robyn_outputs() to avoid these potential issues, so you train the models and then process/export the results in 2 separate steps.
We've seen this is potentially caused for use cases with a lot of media variables. How many variables are you modeling with?
On the other hand, I think you have an outdated version. Can you please update to the latest (dev) version and retry to see if you still get this error? If not, please share a reproducible example so I can take a further look into this potential bug.

[gedkitchener-CT]

Thanks for such a speedy response.

I've got about 30 features and 1,000-ish rows (daily) - in your experience, did that amount of features (30+) sometimes cause this issue in the past?

To be honest, I do have an out-dated version - I'll try switching to the latest (dev) version and run again.

Thanks again for the advice.

[laresbernardo]

We encourage a ~1:10 relation between columns and rows minimum as a rule of thumb. Let me know how it goes with an updated version.

[gedkitchener-CT]

I'm now using R 4.2.1 and Robyn 3.6.3

Using 100 iterations x 5 trials worked fine, both in terms of training the models and producing the output CSVs and plots.

I increased to 2000 iterations x 5 trials, and while robyn_run(...) worked fine, it's the OutputCollect <- robyn_outputs(...) part that's now returning a new error:

>> Exporting pareto results as CSVs into directory...
>> Exporting general plots into directory...
>>> Calculating clusters for model selection using Pareto fronts...
>> Auto selected k = 12 (clusters) based on minimum WSS variance of 5%
>>> Exporting pareto one-pagers into directory...
>> Generating only cluster results one-pagers (12)...
>> Plotting 12 selected models on 6 cores...
  |=========================================                                         |  50%
Failed exporting results, but returned model results anyways:
 Error in serialize(data, node$con): error writing to connection

I should mention that the output CSVs look to be present, but 6 of the one-pagers are missing.

I had a look through the source code, but can't make sense of this new error - any ideas?

[kyletgoldberg]

try using one less core than you have available. serialize issues are due to the parallel processing killing one of the workers.

[gedkitchener-CT]

@kyletgoldberg That's very handy information - I didn't know that.

Dropping the number of cores appears to have solved this serialise issue.

Thanks again.