Graceful exception handling with tune.run

I am performing a simple hyperparameter search calling tune.run. In my case it is expected that some configurations will not work and the trainable throws an exception. In principle, I think this is not a problem for tune.run as it collects the errors and finishes all the other runs. However, after completion instead of returning the tune.Analysis object, tune.run throws an exception itself.

I am now wondering, what is the intended way of gracefully handling exceptions that occur within the trainable? How do I get access to the tune.Analysis object if some runs failed?

Hmm, you could do tune.run(raise_on_failed_trial=False)?

1 Like

Oh yes! I missed that option, thanks a lot!