Current singularity HPO workflow calls train.sh, that downloads the data and sometimes when the data isn't present all processes involved in HPO download the data resulting into undesirable data race condition.
Possible Fix:
HPO spuns two jobs:
- preprocess.sh
- this runs on one processor or as the model data download in preprocess.sh is setup
- with args to get and format data as per requirements
- launch the HPO job
- as per the procs and run configuration specified for running the HPO workflow
Note: Design should consider that: this infrastructure will be useful for the cross-study workflows,
where we run multiple HPO for each differing in the preprocess.sh inputs. cross-study workflow also has to run infer.sh on all of the combinations.
Current singularity HPO workflow calls train.sh, that downloads the data and sometimes when the data isn't present all processes involved in HPO download the data resulting into undesirable data race condition.
Possible Fix:
HPO spuns two jobs:
Note: Design should consider that: this infrastructure will be useful for the
cross-studyworkflows,where we run multiple HPO for each differing in the preprocess.sh inputs.
cross-studyworkflow also has to runinfer.shon all of the combinations.