Hi,
In older versions of Ray (1.13 for example), I could set the resume flag for tune.run() to true so that the hyperparameter optimization resumes even from a previously finished experiment. For example, if in a previous experiment, the tuner already explore `{“a”:1, “b”:1.5}, it won’t do it again and only runs for unexplored configurations.
I’v recently started migrating my code to newer versions of Ray and can’t find a similar feature in the most recent release. Has this feature been removed or it’s me who can’t find it?
Experiment restoration is still possible, though it’s only for unfinished experiments that were interrupted/crashed in the middle. Ex: If I had 5 pending trials at the time of the experiment crashing, then restoring will finish up running those 5. If I had a previously finished experiment (all trials terminated), then restoring will not have anything left to run. This is in line with the previous tune.run(resume) flag.
Here’s how you do that: How to Enable Fault Tolerance in Ray Tune — Ray 2.7.0
If you’re looking to start a new run, but start from the saved searcher state from a previous experiment, this is probably what you’re looking for: Tune Search Algorithms (tune.search) — Ray 2.7.0
Let me know if this clarifies things for you!
Thanks @justinvyu for the answer.
I was looking for something more like the Tune search algorithm restoration. I will try it out.
1 Like