Categorical type error with BayesOpt

Ray Libraries (Data, Train, Tune, Serve) Ray Tune
1

I’m optimizing a model with a pretty standard set of hyperparameters, using BayesOpt:

    config = {
        "num_layers": tune.choice([2, 3]),
        "num_trees": tune.choice([512, 768, 1024]),
        "depth": tune.choice([4, 5, 6]),
        "batch_size": tune.choice([128, 512, 1024])
    }

However, when I run the optimization, I get the following:

ValueError: BayesOpt does not support parameters of type `Categorical`

I’m having a hard time interpreting this error message, and therefore I’m not sure how to act on it. These are not categorical parameters.

Running Ray 1.6.0 on Ubuntu, optimizing a Pytorch Tabular model.

2

Hey @fonnesbeck,

Looks like the error message is grouping discrete parameters together with categorical. However, by definition BayesOpt only supports searching across continuous parameters.

Perhaps you can try another one of the Search Algorithms that does support discrete/categorical parameters?

3
  1. Link is invalid.
  2. Is it possible to combine a search algorithm for continuous params with a search algo for discrete?