CosineAnnealingLR

class torch.optim.lr_scheduler.CosineAnnealingLR(optimizer, T_max, eta_min=0.0, last_epoch=-1)

Set the learning rate of each parameter group using a cosine annealing schedule.

The learning rate is updated recursively using:

$$\eta_{t+1} = \eta_{\min} + (\eta_t - \eta_{\min}) \cdot \frac{1 + \cos\left(\frac{(T_{cur}+1) \pi}{T_{max}}\right)} {1 + \cos\left(\frac{T_{cur} \pi}{T_{max}}\right)}$$

This implements a recursive approximation of the closed-form schedule proposed in SGDR: Stochastic Gradient Descent with Warm Restarts:

$$\eta_t = \eta_{\min} + \frac{1}{2}(\eta_{\max} - \eta_{\min}) \left( 1 + \cos\left(\frac{T_{cur} \pi}{T_{max}}\right) \right)$$

where:

• $\eta_t$ is the learning rate at step $t$

• $T_{cur}$ is the number of epochs since the last restart

• $T_{max}$ is the maximum number of epochs in a cycle

Note

Although SGDR includes periodic restarts, this implementation performs cosine annealing without restarts, so $T_{cur} = t$ and increases monotonically with each call to step().

Parameters

• optimizer (Optimizer) – Wrapped optimizer.

• T_max (int) – Maximum number of iterations.

• eta_min (float) – Minimum learning rate. Default: 0.

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

Example

>>> num_epochs = 100
>>> scheduler = CosineAnnealingLR(optimizer, T_max=num_epochs)
>>> for epoch in range(num_epochs):
>>>     train(...)
>>>     validate(...)
>>>     scheduler.step()

get_last_lr()

Return last computed learning rate by current scheduler.

Return type

list[float]

get_lr()

Retrieve the learning rate of each parameter group.

Return type

list[float]

load_state_dict(state_dict)

Load the scheduler’s state.

Parameters

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

state_dict()

Return the state of the scheduler as a dict.

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

Return type

dict[str, Any]

step(epoch=None)

Perform a step.