Schedulers#

KonfAI has two distinct scheduler families, resolved by different loaders. Don’t confuse them. Use this page when filling in a schedulers: block — first check which of the two blocks you are in, then pick a name from the matching table.

A. Criterion-weight schedulers#

These schedule the weight of a loss over training, in the schedulers: subtree of a criterion (konfai/metric/schedulers.py, base class Scheduler):

CrossEntropyLoss:
  is_loss: true
  schedulers:
    Constant: { nb_step: 0, value: 1 }

Name

Purpose

Config args (defaults)

Constant

Fixed weight for all iterations.

value=1 (+ nb_step)

CosineAnnealing

Cosine anneal from start_value to eta_min over t_max.

start_value=1, eta_min=1e-5, t_max=100 (+ nb_step)

Note

Each entry carries an nb_step (window width). Multiple schedulers can be chained into consecutive iteration windows; an nb_step: 0 (or None) window is the terminal, always-on schedule.

B. Learning-rate schedulers#

These schedule the optimizer learning rate, in the model’s schedulers: block. The loader searches both torch.optim.lr_scheduler and konfai.metric.schedulers, so you can use any torch scheduler (StepLR, ReduceLROnPlateau, CosineAnnealingLR, …) plus the two KonfAI additions:

schedulers:
  StepLR: { step_size: 20, gamma: 0.5 }     # any torch LR scheduler works

Name

Purpose

Config args (defaults)

Warmup

Linear LR warmup wrapper (LambdaLR).

warmup_steps=10, last_epoch=-1 (+ nb_step)

PolyLRScheduler

nnU-Net-style polynomial LR decay lr = initial_lr·(1 − step/max_steps)^exponent.

initial_lr (required), max_steps (required), exponent=0.9 (+ nb_step)

Note

The optimizer itself is injected by the framework — you do not write it under schedulers:.

Next steps#