biapy.engine.schedulers.warmup_cosine_decay

Learning rate scheduler with warmup and cosine decay for BiaPy.

This module provides the WarmUpCosineDecayScheduler class, which implements a learning rate schedule with a linear warmup phase followed by cosine decay, as commonly used in modern deep learning training pipelines.

class biapy.engine.schedulers.warmup_cosine_decay.WarmUpCosineDecayScheduler(lr: float, min_lr: float, warmup_epochs: int, epochs: int)

Bases: object

Learning rate scheduler with linear warmup and cosine decay.

This scheduler increases the learning rate linearly for a specified number of warmup epochs, then decays it following a half-cycle cosine schedule down to a minimum learning rate.

adjust_learning_rate(optimizer: Optimizer, epoch: float | int) → float

Decay the learning rate with half-cycle cosine after warmup.

Parameters:

• optimizer (Optimizer) – PyTorch optimizer whose learning rate will be adjusted.

• epoch (float or int) – Current epoch (can be fractional for finer granularity).

Returns:

lr – The adjusted learning rate.

Return type:

float