1.5 KiB
Running Experiments with Lambda:
This is not supposed to be a instruction on how to do it properly, but it is a writeup on how i did it. If you want to do it properly, extend the command line Arguments for haga-lambda and allow runtime tweaking of Hyperparams and Datasets. While at it, generalizing LamdaCalculusV1 would be smart, too. You can use LamdaCalculusV2 as a template on how to do it more properly. (I wrote that later, and was IMO quite a bit smarter about it. I sadly didn't have time to fix up V1...)
You just want to do the same hack i did or know about it?
create a branch for each Dataset-experiment pair. e.g. iris_1 ... iris_9
here git is your friend, especially if you inevitably screw up.
e.g. echo git\ checkout\ iris_{1..9};\ git\ cherry-pick\ 7ced1e1; will create a command for applying the commit 7ced1e1 to every iris branch.
Adapt the build.sbatch and run.sbatch and commit them! clone the branch you committed to on the cluster. create the required folders! If you forget the output one, slurm will fail silently!
Make sure to sbatch an adapted build.sbatch before run.sbatch! build.sbatch will need to be adapted for and run on every node you will use! Otherwise stuff WILL break!
sbatch run.sbatch
You can use squeue to monitor progress.
A huge slew of raw data will be dumped into the output Folder. The error files contain results, the output files stats during training.
On how to process these results, see: https://merl.dnshome.de/git/Hans/haga-graphics