Continue jobs

Continue an old job with the same options

Continuing an old job with the same options as previously, but with awareness of the hyperparameter-constellations that have already been tested, is as simple as this, assuming your job is in runs/my_experiment/0 :

./omniopt --continue runs/my_experiment/0

This will start a new run with the same settings as the old one, but load all already tried out data points, and continue the search from there.

Continue an old job with the changed options

In continued runs, some options can be changed. For example,

./omniopt --continue runs/my_experiment/0 --time=360

Will run the continuation for 6 hours ( = 360 minutes), independent of how long the old job ran.
It is also possible to change parameter borders, though narrowing is not currently supported, i.e. the parameters need to be equal or wider than in the previous run. You cannot remove parameters here or add names of parameters that have previously not existed, though.
Also, parameter names and types must stay the same. This code, for example, runs a continued run but changes the epochs parameter that was previously defined and could have been, for example, between 0 and 10. In this new run, all old values will be considered, but new values of epochs can be between 0 and 1000.

./omniopt --continue runs/my_experiment/0 --parameter epochs range 0 1000 int

All parameters that are not specified here are taken out of the old run, and thus stay in the same borders.

In which folder will a run continue in?

It will create a new folder. Imagine there are already the subfolders 0 , 1 and 2 for your experiment. If you continue the job 0 , it's job data will be in the subfolder 3 then, since it is the first non-existing folder for that project..

Caveat

It is currently not possible to decrease the search space on a continued run. Attempts to do that will be ignored and the original limits will automatically be restored.