EngineBuilder.add_adaptation()

EngineBuilder.add_adaptation()#

EngineBuilder.add_adaptation(duration, init=0.1, term=0.2, base=25, thinning=1)[source]#

Adds adaptation epochs, for tuning kernel hyperparameters.

This follows the Stan Development Team (Stan Reference Manual 2021, Chapter 15.2 [1]).

Note that the order in which you add epochs matters. Adaptation epochs should usually come before all other epochs.

Parameters:

  • duration (int) – The number of adaptation samples.

  • init (float | int, default: 0.1) – The share of samples in the initial fast adaptation epoch, relative to the total number of adaptation samples (if given as a float), or the total number of init samples (if given as an integer).

  • term (float | int, default: 0.2) – The share of samples in the final fast adaptation epoch, relative to the total number of adaptation samples (if given as a float), or the total number of term samples (if given as an integer).

  • base (int, default: 25) – The number of samples in the first slow adaptation epoch.

  • thinning (int, default: 1) – Thinning applied in each adaptation.

Warning

Kernels which rely on history tuning might not be able to deal with thinning during adaptation.

See also

add_posterior

For adding posterior epochs.

add_burnin

For adding burnin epochs.

add_adaptation

For adding adaptation epochs.

set_duration

An alternative method to specify adaptation and posterior epochs, see notes for details.

Notes

There are some differences in the details between the methods add_posterior(), add_adaptation(), add_burnin() and set_duration().

  • set_duration() overwrites all existing epochs with the epochs defined in this method. In contrast, the add_ methods only add additional epochs that are appended to the potentially existing ones.

  • set_duration() sets both adaptation and posterior epochs, while each add_ method only adds epochs of one type.

  • set_duration() has an argument term_duration that defines the number of samples drawn in the final fast adaptation epoch, and it defaults to a fixed value of 50 samples. In contrast, add_adaptation() defines the number of samples drawn in the final fast adaptation epoch as a fraction of the total number of adaptation samples, and allows users to change this fraction via the argument term. As a result, if add_adaptation() is used, the length of the final fast adaptation phase increases automatically, if the total number of warmup samples increases, while it remains constant if set_duration() is used.

References