Skip to content

⏳ Schedules API

The diff_integrator.schedules module provides weight schedule utilities for integrative refinement.

Weight schedules allow loss term weights to evolve dynamically during optimization. The primary motivation is the annealed geometry restraint pattern: start with a strong positional anchor (preventing fold distortion in early epochs) and gradually relax it so that experimental gradients dominate in later epochs.

ExponentialDecaySchedule

diff_integrator.schedules.ExponentialDecaySchedule

Exponential decay weight schedule. Returns a weight that decreases smoothly from initial_weight toward final_weight over decay_epochs epochs.

Constructor

ExponentialDecaySchedule(
    initial_weight: float,
    final_weight: float,
    decay_epochs: int,
)

LinearSchedule

diff_integrator.schedules.LinearSchedule

Linear interpolation weight schedule. Returns a weight that changes at a constant rate from initial_weight to final_weight over decay_epochs epochs, then holds at final_weight.

Constructor

LinearSchedule(
    initial_weight: float,
    final_weight: float,
    decay_epochs: int,
)

CosineAnnealingSchedule

diff_integrator.schedules.CosineAnnealingSchedule

Cosine annealing weight schedule. Returns a weight that follows a half-cosine curve from initial_weight to final_weight over decay_epochs epochs, then holds at final_weight.

Constructor

CosineAnnealingSchedule(
    initial_weight: float,
    final_weight: float,
    decay_epochs: int,
)