phenotypic.tune.ScreeningConfig#

class phenotypic.tune.ScreeningConfig(free_param_floor: int = 6, freeze_epsilon: float = 0.1, warmup_floor: int = 20, warmup_c: int = 3, top_k: int = 10)[source]#

Bases: object

Conservative, config-exposed freeze thresholds (screening-importance §14).

Parameters:

  • free_param_floor (int) – Screening is off until the space has more than this many free (non-Fixed) params (default 6).

  • freeze_epsilon (float) – The cumulative-tail budget ε: freeze the least-important params whose total importance collectively stays < ε (default 0.10 → keep the params covering ~90% of explained variance).

  • warmup_floor (int) – The absolute warm-up trial floor (default 20).

  • warmup_c (int) – The per-param warm-up multiplier; the freeze-grade floor is max(warmup_floor, warmup_c · n_params) (default 3).

  • top_k (int) – How many best explore trials feed the central-tendency freeze value and the focused round’s warm-start (default 10).

Methods

__init__

Attributes

free_param_floor

freeze_epsilon

top_k

warmup_c

warmup_floor

free_param_floor: int = 6#
freeze_epsilon: float = 0.1#
warmup_floor: int = 20#
warmup_c: int = 3#
top_k: int = 10#