Hello everybody!
I am currently working on a demand forecasting project in which I have to train a Prophet model using custom hyperparameters. I am pretty new with your library (before I was using Optuna’s framework for the hp job). However, I wanted to see how much of a improvement (in terms of speed) does actually RayTune gives to my script.
Besides that as I was walking through your documentation, I’ve understood that you could early stop a trial using scheduler (I found an example using ASHAScheduler and implemented it myself). But I really struggle in understaning of how it works.
What I want to accomplish is the next thing: I’ve seen in previous hyperparameter searches that I’ve done that some trials just froze (my intuition is that it didn’t like some combination of parameters). After the trial froze I had to manually end the python proccess and restart it once again excluding some parameters or minimising the search_space (in order to avoid that combination of parameters happen again). Is there any workaround doing this programmatically? Like given a scheduler the maximum time of a trial to compute and if it is exceeded pass over that trial and continue training?