PrecisionPlugin

class pytorch_lightning.plugins.precision.PrecisionPlugin

Bases: pytorch_lightning.plugins.base_plugin.Plugin

Base class for all plugins handling the precision-specific parts of the training. The static classattributes EPSILON and precision must be overwritten in child-classes and their default values reflect fp32 training.

backward(model, closure_loss, optimizer, opt_idx, should_accumulate, *args, **kwargs)

performs the actual backpropagation

Parameters

• model (LightningModule) – the model to be optimized

• closure_loss (Tensor) – the loss value obtained from the closure

• optimizer (Optimizer) – the optimizer to perform the step lateron

• opt_idx (int) – the optimizer’s index

• should_accumulate (bool) – whether to accumulate gradients or not

Return type

Tensor

clip_grad_by_norm(optimizer, clip_val, norm_type=2.0, eps=1e-06)

Clip gradients by norm

Return type

None

clip_grad_by_value(optimizer, clip_val)

Clip gradients by value

Return type

None

clip_gradients(optimizer, clip_val, gradient_clip_algorithm=<GradClipAlgorithmType.NORM: 'norm'>, model=None)

Clips the gradients

Return type

None

connect(model, optimizers, lr_schedulers)

Connects this plugin to the accelerator and the training process

Return type

Tuple[Module, List[Optimizer], List[Any]]

master_params(optimizer)

The master params of the model. Returns the plain model params here. Maybe different in other precision plugins.

Return type

Iterator[Parameter]

post_optimizer_step(optimizer, optimizer_idx)

Hook to do something after each optimizer step.

Return type

None

pre_optimizer_step(pl_module, optimizer, optimizer_idx, lambda_closure, **kwargs)

Hook to do something before each optimizer step.

Return type

bool