Script issueHello,
Customers at Solvay are developing a script and faced some issue. I tested it on JMP-16 and JMP-17: on v16 sometimes it works, on v17 it crashes each time. I copy below his message:
"... I have been working on some code to train a neural network with a hyperparameter optimization routine. The script is called "fit_NN_loop" in the attached file. If you run the script, a window pops up to select the x values, y values, and validation columns:
Basically, we do a hyperparameter tuning of a NN for each validation column. The hyperparameter space is sampled based on a DOE (Space filling design), which you will see in the script from lines 34 to 56. When we try to run the code, JMP crashes randomly. Sometimes we get the following message:
We suspect that perhaps launching the DOE platform from the same script could be the problem, but we are not certain..."
Example of data with script (this is the last one) is attached, help will be appreciated (and serve as learning opportunity for myself).
Emmanuel
Re: Script issueHi Emmanuel,
I am far from being an expert but this might not be a script issue.
I observed the following phenomenon.
If you try to create the DoE they want using the Space Filling Platform, JMP also crashes.
- With 1 continuous variable and 3 categorical variables with 4 levels, this is OK
- With 1 continuous variable and 3 categorical variables with 8 levels, this is still OK
- With 1 continuous variable and 3 categorical variables with 9 levels, the PC is not responding anymore
- With 1 continuous variable and 3 categorical variables with 10 levels, I have a direct crash.
Re: Script issueHello Stephane,
good point this is probably the main part of the problem.
That being said by reducing the #levels to 8 it works on v16 but still not in v17. On v16 if you run it a second time it still crash, I have to close & re-open JMP-16 for having it working.
Emmanuel
Re: Script issueBonjour Emmanuel,
Please ask Solvay to report this to Technical Support (support@jmp.com). They can copy you on the report. Technical Support will investigate, report any problem to development, and help the customer find a workaround if possible.
Merci!
Re: Script issueHi Jeff,
they did it indeed. I was thinking (may be I'm wrong) it is a script issue rather than a bug. Let see what JMP-Support will say. Emmanuel
Re: Script issueThis is a bug - I've opened up a defect where TS can update with the customer track (https://jmpdiscovery.atlassian.net/browse/JMP-16127).
@Jeff_Perkinson, if you want to point TS here, I thought I'd chime in about a general workaround, since it's outside of the bug.
For only 10 runs, all that the code ends up doing is a random sample of one level from each of those categorical, and then just space-filling one dimension of the continuous. So, I wouldn't really call it much of a space-filling design. If they prefer, in JMP 17, we can handle discrete numeric factors. Based on the number of runs, it will likely pick a couple of runs at 1 and a couple at 10 for each of those factors (and not all 10 values), but at least it tries to spread out those choses in a numerical nature.
Re: Script issueI'm glad to hear that they've contacted Tech Support.
Remember, a crash is always a bug and should be reported to Tech Support. You should not be able to make a scripting mistake and cause JMP to crash.
Re: Script issueFollowing up here - the crash for this example was reported to Development, and fixed in JMP 18 and JMP 17.2.
Re: Script issueAny crash should be reported directly to Tech Support (support@jmp.com). They are best positioned to respond and suggest any workarounds that may exist.
Please do not try to do this support yourself. We need to track the reports in order to ensure they are corrected in subsequent releases.
