dymad.utils.scheduler

Functions

make_scheduler([scheduler, scheduler_type])

Factory function to create a scheduler instance.

Classes

SchedulerBase()

Base class for schedulers.

StandardScheduler(scheduler)

Standard scheduler that wraps a PyTorch scheduler.

SweepScheduler([sweep_lengths, sweep_tols, ...])

Scheduler to manage sweep lengths during training.
class dymad.utils.scheduler.SchedulerBase

Bases: object

Base class for schedulers.

As wrapper of standard schedulers, with customized inputs and outputs. Also for other custom schedulers.

diagnostic_info()

Return diagnostic information about the scheduler.

Returns:

Diagnostic information.

Return type:

str

load_state_dict(state_dict)

Load the state dictionary.

Return type:

None

state_dict()

Return the state dictionary for saving.

Return type:

dict[str, Any]

step(**kwargs)

Step through the scheduler logic.

Returns:

True if the scheduler indicates to stop training, False by default.

Return type:

bool

class dymad.utils.scheduler.StandardScheduler(scheduler)

Bases: SchedulerBase

Standard scheduler that wraps a PyTorch scheduler.

Parameters:

scheduler (torch.optim.lr_scheduler._LRScheduler) – PyTorch scheduler instance.

diagnostic_info()

Return diagnostic information about the scheduler.

Returns:

Diagnostic information.

Return type:

str

load_state_dict(state_dict)

Load the state dictionary into the wrapped scheduler.

Return type:

None

state_dict()

Return the state dictionary of the wrapped scheduler.

Return type:

dict[str, Any]

step(**kwargs)

Step through the wrapped scheduler.

Return type:

tuple[bool, bool]

class dymad.utils.scheduler.SweepScheduler(sweep_lengths=None, sweep_tols=None, epoch_step=10, mode='skip')

Bases: SchedulerBase

Scheduler to manage sweep lengths during training. Cycles through predefined sweep lengths and tolerances.

It operates in four modes:

  • No sweep_lengths or sweep_tols: Does nothing.

  • Only sweep_lengths: Change lengths every epoch_step.

  • sweep_lengths AND sweep_tols: Change lengths either every epoch_step or a tolerance is reached, whichever comes first. In particular, when the last tolerance is reached,

    • If mode=’full’: the scheduler resets to the initial length.

    • If mode=’skip’: the scheduler keeps using the current length.

Note

  • If sweep_tols is set, the optimization ends when the last tolerance is reached.

Parameters:

  • sweep_lengths (list) – List of sweep lengths to cycle through.

  • sweep_tols (list, optional) – List of tolerances for each sweep length.

  • epoch_step (int) – Number of epochs after which to switch to the next sweep length.

  • mode (str) – Mode of operation (‘full’ or ‘skip’). Default is ‘skip’.

diagnostic_info()

Return diagnostic information about the scheduler.

Returns:

Diagnostic information.

Return type:

str

get_length()
Return type:

int

load_state_dict(state_dict)

Load the state dictionary.

Return type:

None

state_dict()

Return the state dictionary for saving.

Return type:

dict[str, Any]

step(eploss=None)

Step through the scheduler logic.

Returns:

True if the scheduler indicates to stop training, False by default.

Return type:

bool

dymad.utils.scheduler.make_scheduler(scheduler=None, scheduler_type='', **kwargs)

Factory function to create a scheduler instance.

Parameters:

  • scheduler_type (str) – Type of the scheduler to create.

  • **kwargs – Additional arguments for the scheduler.

Returns:

An instance of the specified scheduler type.

Return type:

SchedulerBase