search

keras-team / keras-tuner

A Hyperparameter Tuning Library for Keras
https://keras.io/keras_tuner/
Apache License 2.0
2.85k stars 395 forks source link

control n_max in Hyperband #323

Open PaulAmosKreiner opened 4 years ago

PaulAmosKreiner commented 4 years ago

[Feature Request] As stated on page 11 of the hyperband paper (section 3.6), in scenarios with high overhead costs per model == per hyperparameter combination one might want to set n_max (meaning a maximum number of different hparam combinations to use).

sebastianlutter commented 4 years ago

This would be very helpful in my opinion.

I work with hp.Float("learning_rate", 0.000001, 0.001) values. As far as I understood I can calulate the number of combinations it tries by calculating the number of iterations like this:

(1 + log(max_epochs)/log(factor) ** 2)

Given this, the only way to modify the number of hyperparameter combinations is changing max_epochs or the factor. Would be great if we could set a limit below this.