Manual Warmup

class pytorch_warmup.base.BaseWarmup(optimizer, warmup_params, last_step=-1)

Base class for all warmup schedules.

The learning rate \(\alpha_{t}\) is dampened by multiplying it by the warmup factor \(\omega_{t} \in [0, 1]\) at each iteration \(t\). Thus, the modified learning rate

\[\hat \alpha_{t} = \alpha_{t} \cdot \omega_{t} \]

is used by the optimizer.

Parameters:

• optimizer (Optimizer) – Wrapped optimizer.

• warmup_params (list) – Warmup parameters.

• last_step (int) – The index of last step. Default: -1.

dampen(step=None)

Dampens the learning rate.

It is not recommended to explicitly call this method for PyTorch 1.4.0 or later. Please use the dampening() context manager that calls this method correctly.

Parameters:

step (int) – The index of current step. Default: None.

dampening()

Dampens the learning rate after calling the step() method of the learning rate scheduler.

The step() method calls must be placed in a suite of the with statement having the dampening() context manager.

Examples

>>> # For no LR scheduler
>>> with warmup_scheduler.dampening():
>>>     pass

>>> # For a single LR scheduler
>>> with warmup_scheduler.dampening():
>>>     lr_scheduler.step()

>>> # To chain two LR schedulers
>>> with warmup_scheduler.dampening():
>>>     lr_scheduler1.step()
>>>     lr_scheduler2.step()

>>> # To delay an LR scheduler
>>> iteration = warmup_scheduler.last_step + 1
>>> with warmup_scheduler.dampening():
>>>     if iteration >= warmup_period:
>>>         lr_scheduler.step()

load_state_dict(state_dict)

Loads the warmup scheduler’s state.

Parameters:

state_dict (dict) – Warmup scheduler state. Should be an object returned from a call to state_dict().

state_dict()

Returns the state of the warmup scheduler as a dict.

It contains an entry for every variable in self.__dict__ which is not the optimizer.

warmup_factor(step, **params)

Returns the warmup factor \(\omega_{t}\) at an iteration \(t\).

dampen() uses this method to get the warmup factor for each parameter group. It is unnecessary to explicitly call this method.

Parameters:

• step (int) – The index of current step.

• params (dict) – The warmup parameters. For details, refer to the arguments of each subclass method.

class pytorch_warmup.base.ExponentialWarmup(optimizer, warmup_period, last_step=-1)

Exponential warmup schedule.

The exponential warmup schedule uses the warmup factor

\[\omega_{t}^{\rm expo, \tau} = 1 - \exp \left( - \frac{1}{\tau} \cdot t \right) \]

at each iteration \(t\), where the constant \(\tau\) is analogous to a linear warmup period.

Parameters:

• optimizer (Optimizer) – Wrapped optimizer. RAdam is not suitable because of the warmup redundancy.

• warmup_period (int or list[int]) – The constant \(\tau\) analogous to a linear warmup period.

• last_step (int) – The index of last step. Default: -1.

Example

>>> lr_scheduler = CosineAnnealingLR(optimizer, ...)
>>> warmup_scheduler = ExponentialWarmup(optimizer, warmup_period=1000)
>>> for batch in dataloader:
>>>     optimizer.zero_grad()
>>>     loss = ...
>>>     loss.backward()
>>>     optimizer.step()
>>>     with warmup_scheduler.dampening():
>>>         lr_scheduler.step()

Warning

The warmup schedule must not be initialized before the initialization of the learning rate schedule.

warmup_factor(step, warmup_period)

Returns the warmup factor \(\omega_{t}^{\rm expo, \tau}\) at an iteration \(t\).

Parameters:

• step (int) – The index of current step.

• warmup_period (int) – The constant \(\tau\) analogous to a linear warmup period.

class pytorch_warmup.base.LinearWarmup(optimizer, warmup_period, last_step=-1)

Linear warmup schedule.

The linear warmup schedule uses the warmup factor

\[\omega_{t}^{\rm linear, \tau} = \min \left\{ 1, \frac{1}{\tau} \cdot t \right\} \]

at each iteration \(t\), where \(\tau\) is the warmup period.

Parameters:

• optimizer (Optimizer) – Wrapped optimizer. RAdam is not suitable because of the warmup redundancy.

• warmup_period (int or list[int]) – The warmup period \(\tau\).

• last_step (int) – The index of last step. Default: -1.

Example

>>> lr_scheduler = CosineAnnealingLR(optimizer, ...)
>>> warmup_scheduler = LinearWarmup(optimizer, warmup_period=2000)
>>> for batch in dataloader:
>>>     optimizer.zero_grad()
>>>     loss = ...
>>>     loss.backward()
>>>     optimizer.step()
>>>     with warmup_scheduler.dampening():
>>>         lr_scheduler.step()

Warning

The warmup schedule must not be initialized before the initialization of the learning rate schedule.

warmup_factor(step, warmup_period)

Returns the warmup factor \(\omega_{t}^{\rm linear, \tau}\) at an iteration \(t\).

Parameters:

• step (int) – The index of current step.

• warmup_period (int) – The warmup period \(\tau\).