sage.factory.schedulers

Filename : schedulers.py Description : Short description of the file

Created on 2026-03-06 16:29:32

__author__ = Narenraju Nagarajan __copyright__ = Copyright 2026, ProjectName __license__ = MIT Licence __version__ = 0.0.1 __maintainer__ = Narenraju Nagarajan __affiliation__ = N/A __email__ = N/A __status__ = [‘inProgress’, ‘Archived’, ‘inUsage’, ‘Debugging’]

GitHub Repository: NULL

Documentation: NULL

Classes

ManageScheduler

Unified adapter for PyTorch learning-rate schedulers with multiple

Module Contents

class ManageScheduler(scheduler, mode='batch')

Unified adapter for PyTorch learning-rate schedulers with multiple step-trigger modes.

Different scheduler types require step() to be called at different times (per batch, per epoch, or after a metric update). This wrapper standardises the interface by routing batch_step() and epoch_step() calls to the underlying scheduler according to the configured mode.

Parameters:

• scheduler (torch.optim.lr_scheduler._LRScheduler) – The underlying scheduler instance.

• mode (str) –

  When to advance the scheduler:

  • "batch" — call scheduler.step() on every batch.

  • "fractional" — call scheduler.step(epoch + batch/total) for schedulers that accept a fractional epoch argument.

  • "epoch" — call scheduler.step() once per epoch (in epoch_step()).

  • "metric" — call scheduler.step(metric) once per epoch (e.g. torch.optim.lr_scheduler.ReduceLROnPlateau).

scheduler

mode = 'batch'

batch_step(nepoch=None, nbatch=None, num_batches=None)

Advance the scheduler at the end of a batch.

Only takes effect when mode is "batch" or "fractional".

Parameters:

• nepoch (int) – Current epoch index (required for "fractional" mode).

• nbatch (int) – Current batch index within the epoch (required for "fractional").

• num_batches (int) – Total number of batches per epoch (required for "fractional").

epoch_step(metric=None)

Advance the scheduler at the end of an epoch.

Only takes effect when mode is "epoch" or "metric".

Parameters:

metric (float or None) – Validation metric value (required for "metric" mode, e.g. validation loss passed to ReduceLROnPlateau).