vis4d.engine.optim.scheduler

LR schedulers.

Classes

ConstantLR(optimizer, max_steps[, factor, ...])	Constant learning rate scheduler.
LRSchedulerDict	LR scheduler.
LRSchedulerWrapper(lr_schedulers_cfg, optimizer)	LR scheduler wrapper.
PolyLR(optimizer, max_steps[, power, ...])	Polynomial learning rate decay.
QuadraticLRWarmup(optimizer, max_steps[, ...])	Quadratic learning rate warmup.

class ConstantLR(optimizer, max_steps, factor=0.3333333333333333, last_epoch=-1)

Constant learning rate scheduler.

Parameters:

• optimizer (Optimizer) – Wrapped optimizer.

• max_steps (int) – Maximum number of steps.

• factor (float) – Scale factor. Default: 1.0 / 3.0.

• last_epoch (int) – The index of last epoch. Default: -1.

Initialize ConstantLR.

get_lr()

Compute current learning rate.

Return type:

list[float]

class LRSchedulerDict

LR scheduler.

class LRSchedulerWrapper(lr_schedulers_cfg, optimizer, steps_per_epoch=-1)

LR scheduler wrapper.

Initialize LRSchedulerWrapper.

get_lr()

Get current learning rate.

Return type:

list[float]

load_state_dict(state_dict)

Load state dict.

Return type:

None

state_dict()

Get state dict.

Return type:

dict[int, Dict[str, Any]]

step(epoch=None)

Step on training epoch end.

Return type:

None

step_on_batch(step)

Step on training batch end.

Return type:

None

class PolyLR(optimizer, max_steps, power=1.0, min_lr=0.0, last_epoch=-1)

Polynomial learning rate decay.

Example

Assuming lr = 0.001, max_steps = 4, min_lr = 0.0, and power = 1.0, the learning rate will be: lr = 0.001 if step == 0 lr = 0.00075 if step == 1 lr = 0.00050 if step == 2 lr = 0.00025 if step == 3 lr = 0.0 if step >= 4

Parameters:

• optimizer (Optimizer) – Wrapped optimizer.

• max_steps (int) – Maximum number of steps.

• power (float, optional) – Power factor. Default: 1.0.

• min_lr (float) – Minimum learning rate. Default: 0.0.

• last_epoch (int) – The index of last epoch. Default: -1.

Initialize PolyLRScheduler.

get_lr()

Compute current learning rate.

Return type:

list[float]

class QuadraticLRWarmup(optimizer, max_steps, last_epoch=-1)

Quadratic learning rate warmup.

Parameters:

• optimizer (Optimizer) – Wrapped optimizer.

• max_steps (int) – Maximum number of steps.

• last_epoch (int) – The index of last epoch. Default: -1.

Initialize QuadraticLRWarmup.

get_lr()

Compute current learning rate.

Return type:

list[float]