I’m using “algorithm.restore_from_dir” to continue running the previous finished tune. In the case of using SkOptSearch, the hyperparameter values of all trials have identical values. in another word, all tries run on exactly the same configuration. Any clue on how this happens would be appreciated. Unfortunately I can’t repeat in small scale runs and it only happens in the large scale running. Here are more observations:
- it’s only happens for SkOpt. I also tried hyperopt and it works fine so far.
- max_concurrent = n_initial_points = 100
- when running the tune from scratch, all 100 jobs have different configurations. After the first run finished and I need to continue the tuning using algorithm.restore_from_dir, all 100 jobs have identical configuration.
- when trying to run it a 3rd time, all 100 jobs again running with the same configuration but the value are different from the 2nd try
- I used the “quniform” for some parameters and saw the warning that SkOpt does not allow quantization.
Could this be a ray issue or SkOptSearch issue?