Population Based Training

See schedy.PopulationBasedTraining for a description of the experiment type.

schedy.pbt.MINIMIZE = 'min'

Minimize the objective

schedy.pbt.MAXIMIZE = 'max'

Maximize the objective

class schedy.pbt.Truncate(proportion=0.2)[source]

Truncate exploit strategy: if the selected candidate job is in the worst n%, use a candidate job in the top n% instead.

Parameters:proportion (float) – Proportion of jobs that are considered to be “best” jobs, and “worst” jobs. For example, if proportion = 0.2, if the selected candidate job is in the bottom 20%, it will be replaced by a job in the top 20%. Must satisfy 0 < proportion <= 0.5.
class schedy.pbt.Perturb(min_factor=0.8, max_factor=1.2)[source]

Perturb explore strategy: multiply the designated hyperparameter by a random factor, sampled from a uniform distribution.

Parameters:
  • min_factor (float) – Minimum value for the factor (inclusive).
  • max_factor (float) – Maximum value for the factor (exclusive).