StochasticWeightAveraging

class pytorch_lightning.callbacks.StochasticWeightAveraging(swa_lrs, swa_epoch_start=0.8, annealing_epochs=10, annealing_strategy='cos', avg_fn=None, device=device(type='cpu'))

Bases: pytorch_lightning.callbacks.callback.Callback

Implements the Stochastic Weight Averaging (SWA) Callback to average a model.

Stochastic Weight Averaging was proposed in Averaging Weights Leads to Wider Optima and Better Generalization by Pavel Izmailov, Dmitrii Podoprikhin, Timur Garipov, Dmitry Vetrov and Andrew Gordon Wilson (UAI 2018).

This documentation is highly inspired by PyTorch’s work on SWA. The callback arguments follow the scheme defined in PyTorch’s swa_utils package.

For a SWA explanation, please take a look here.

Warning

StochasticWeightAveraging is in beta and subject to change.

Warning

StochasticWeightAveraging is currently not supported for multiple optimizers/schedulers.

Warning

StochasticWeightAveraging is currently only supported on every epoch.

See also how to enable it directly on the Trainer

Parameters

• swa_lrs (Union[float, List[float]]) –

  The SWA learning rate to use:

  • float. Use this value for all parameter groups of the optimizer.

  • List[float]. A list values for each parameter group of the optimizer.

• swa_epoch_start (Union[int, float]) – If provided as int, the procedure will start from the swa_epoch_start-th epoch. If provided as float between 0 and 1, the procedure will start from int(swa_epoch_start * max_epochs) epoch

• annealing_epochs (int) – number of epochs in the annealing phase (default: 10)

• annealing_strategy (str) –

  Specifies the annealing strategy (default: “cos”):

  • "cos". For cosine annealing.

  • "linear" For linear annealing

• avg_fn (Optional[Callable[[Tensor, Tensor, Tensor], Tensor]]) – the averaging function used to update the parameters; the function must take in the current value of the AveragedModel parameter, the current value of model parameter and the number of models already averaged; if None, equally weighted average is used (default: None)

• device (Union[device, str, None]) – if provided, the averaged model will be stored on the device. When None is provided, it will infer the device from pl_module. (default: "cpu")

static avg_fn(averaged_model_parameter, model_parameter, num_averaged)

Adapted from https://github.com/pytorch/pytorch/blob/v1.7.1/torch/optim/swa_utils.py#L95-L97.

Return type

Tensor

load_state_dict(state_dict)

Called when loading a checkpoint, implement to reload callback state given callback’s state_dict.

Parameters

state_dict (Dict[str, Any]) – the callback state returned by state_dict.

Return type

None

on_fit_start(trainer, pl_module)

Called when fit begins.

Return type

None

on_train_end(trainer, pl_module)

Called when the train ends.

Return type

None

on_train_epoch_end(trainer, *args)

Called when the train epoch ends.

To access all batch outputs at the end of the epoch, either:

1. Implement training_epoch_end in the LightningModule and access outputs via the module OR

2. Cache data across train batch hooks inside the callback implementation to post-process in this hook.

Return type

None

on_train_epoch_start(trainer, pl_module)

Called when the train epoch begins.

Return type

None

reset_batch_norm_and_save_state(pl_module)

Adapted from https://github.com/pytorch/pytorch/blob/v1.7.1/torch/optim/swa_utils.py#L140-L154.

Return type

None

reset_momenta()

Adapted from https://github.com/pytorch/pytorch/blob/v1.7.1/torch/optim/swa_utils.py#L164-L165.

Return type

None

setup(trainer, pl_module, stage)

Called when fit, validate, test, predict, or tune begins.

Return type

None

state_dict()

Called when saving a checkpoint, implement to generate callback’s state_dict.

Return type

Dict[str, Any]

Returns

A dictionary containing callback state.

static update_parameters(average_model, model, n_averaged, avg_fn)

Adapted from https://github.com/pytorch/pytorch/blob/v1.7.1/torch/optim/swa_utils.py#L104-L112.

Return type

None