stable_learning_control.algos.pytorch.common.get_lr_scheduler

Contains functions used for creating Pytorch learning rate schedulers.

Module Contents

Classes

ConstantLRScheduler

A learning rate scheduler that keeps the learning rate constant.

Functions

get_exponential_decay_rate(lr_start, lr_final, steps)	Calculates the exponential decay rate needed to go from a initial learning rate
get_linear_decay_rate(lr_init, lr_final, steps)	Returns a linear decay factor (G) that enables a learning rate to transition
get_lr_scheduler(optimizer, decaying_lr_type, ...)	Creates a learning rate scheduler.
estimate_step_learning_rate(lr_scheduler, lr_start, ...)	Estimates the learning rate at a given step.

class stable_learning_control.algos.pytorch.common.get_lr_scheduler.ConstantLRScheduler(optimizer)

Bases: torch.optim.lr_scheduler.LambdaLR

A learning rate scheduler that keeps the learning rate constant.

Initialize the constant learning rate scheduler.

Parameters:

optimizer (torch.optim.Optimizer) – The wrapped optimizer.

stable_learning_control.algos.pytorch.common.get_lr_scheduler.get_exponential_decay_rate(lr_start, lr_final, steps)

Calculates the exponential decay rate needed to go from a initial learning rate to a final learning rate in N steps.

Parameters:

• lr_start (float) – The starting learning rate.

• lr_final (float) – The final learning rate.

• steps (int) – The number of steps.

Returns:

The exponential decay rate (high precision).

Return type:

decimal.Decimal

stable_learning_control.algos.pytorch.common.get_lr_scheduler.get_linear_decay_rate(lr_init, lr_final, steps)

Returns a linear decay factor (G) that enables a learning rate to transition from an initial value (lr_init) at step 0 to a final value (lr_final) at a specified step (N). This decay factor is compatible with the torch.optim.lr_scheduler.LambdaLR scheduler. The decay factor is calculated using the following formula:

Parameters:

• lr_init (float) – The initial learning rate.

• lr_final (float) – The final learning rate you want to achieve.

• steps (int) – The number of steps/epochs over which the learning rate should decay. This is equal to epochs -1.

Returns:

Linear learning rate decay factor (G).

Return type:

decimal.Decimal

stable_learning_control.algos.pytorch.common.get_lr_scheduler.get_lr_scheduler(optimizer, decaying_lr_type, lr_start, lr_final, steps)

Creates a learning rate scheduler.

Parameters:

• optimizer (torch.optim.Adam) – Wrapped optimizer.

• decaying_lr_type (str) – The learning rate decay type that is used (options are: linear and exponential and constant).

• lr_start (float) – Initial learning rate.

• lr_final (float) – Final learning rate.

• steps (int, optional) – Number of steps/epochs used in the training. This includes the starting step/epoch.

Returns:

A learning rate scheduler object.

Return type:

torch.optim.lr_scheduler

See also

See the pytorch documentation on how to implement other decay options.

stable_learning_control.algos.pytorch.common.get_lr_scheduler.estimate_step_learning_rate(lr_scheduler, lr_start, lr_final, update_after, total_steps, step)

Estimates the learning rate at a given step.

This function estimates the learning rate for a specific training step. It differs from the get_last_lr method of the learning rate scheduler, which returns the learning rate at the last scheduler step, not necessarily the current training step.

Parameters:

• lr_scheduler (torch.optim.lr_scheduler) – The learning rate scheduler.

• lr_start (float) – The initial learning rate.

• update_after (int) – The step number after which the learning rate should start decreasing.

• lr_final (float) – The final learning rate.

• total_steps (int) – The total number of steps/epochs in the training process. Excludes the initial step.

• step (int) – The current step number. Excludes the initial step.

Returns:

The learning rate at the given step.

Return type:

float