A data scientist uses `hp.loguniform("learning_rate", np.log(1e-5), np.log(1e-1))` in a Hyperopt search space. What is the practical implication of using `loguniform` versus `uniform` for the learning rate?

Question

Accepted Answer

B. Sampling is uniform in log space, so the algorithm explores small and large learning rates with equal probability per order of magnitude. Using `hp.loguniform(label, low, high)` samples values such that `log(value)` is uniformly distributed between `low` and `high`. For learning rates, equal probability is given to each order of magnitude. This is the appropriate distribution for parameters spanning multiple orders of magnitude.

A data scientist uses `hp.loguniform("learning_rate", np.log(1e-5), np.log(1e-1))` in a Hyperopt search space. What is the practical implication of using `loguniform` versus `uniform` for the learning rate?

Related Questions

Discussion

A data scientist uses hp.loguniform("learning_rate", np.log(1e-5), np.log(1e-1)) in a Hyperopt search space. What is the practical implication of using loguniform versus uniform for the learning rate?

Related Questions

Discussion

A data scientist uses `hp.loguniform("learning_rate", np.log(1e-5), np.log(1e-1))` in a Hyperopt search space. What is the practical implication of using `loguniform` versus `uniform` for the learning rate?