Clement Emmanuel
03/01/2023, 7:55 PM6
points from the search space at random, we then run 6
dagster graphs representing our trials in parallel against each of these points, the results of these 6
trials are recorded, and now we start the next batch of sampling 6 points and running the associated 6 trials. Our sampling of the subsequent 6
points is actually informed by the results that have been recorded thus far, so it's important to stagger the trials into epochs like this.
I can eagerly and statically inform dagster of the subgraph that represents running a single trial, so what I was thinking is I could have an op
that is responsible for sampling the search-space, then running the trial job, recording the results, then generate subsequent samples from the space, etc. Even if possible the thing I don't totally like is that logically it's like the trial subgraphs are child ops of this meta op
that is sampling the search-space, but dagster's data model will not be aware of this, this logical connection only exists by way of the meta-op dynamically invoking the execution of those trial graphs/jobs
If any of this is ill-advised, or not possible, or there is some better way to accomplish this please let me know.sandy
03/01/2023, 8:56 PMop
that is sampling the search-space, but dagster's data model will not be aware of this, this logical connection only exists by way of the meta-op dynamically invoking the execution of those trial graphs/jobs", but it might still be useful for you?sandy
03/01/2023, 8:57 PM