Colleagues, hello,
Why do you use loguniform distribution in the case for RandomizedSearchCV? Do you have any intuitives/math/empirical reasons to use it here or it is just an example?
If any, can you describe it please.
The empirics in the general case of hyperparameter tuning is that a loguniform allows coarse tuning whereas a np.linspace is used for fine tuning once the global minimum is more or less identified. In some cases the former option is enough, mostly when the hyperparameter has a relatively small impact on the data, like l2_regularization in this case, that can vary along several orders of magnitude.
3 Likes