scvi.train.JaxTrainingPlan

class scvi.train.JaxTrainingPlan(module, *, optimizer='Adam', optimizer_creator=None, lr=0.001, weight_decay=1e-06, eps=0.01, max_norm=None, n_steps_kl_warmup=None, n_epochs_kl_warmup=400, **loss_kwargs)

Bases: TrainingPlan

Lightning module task to train Pyro scvi-tools modules.

Parameters:

• module (JaxBaseModuleClass) – An instance of JaxModuleWraper.

• optimizer (Literal['Adam', 'AdamW', 'Custom'] (default: 'Adam')) – One of “Adam”, “AdamW”, or “Custom”, which requires a custom optimizer creator callable to be passed via optimizer_creator.

• optimizer_creator (Callable[[], GradientTransformation] | None (default: None)) – A callable returning a GradientTransformation. This allows using any optax optimizer with custom hyperparameters.

• lr (float (default: 0.001)) – Learning rate used for optimization, when optimizer_creator is None.

• weight_decay (float (default: 1e-06)) – Weight decay used in optimization, when optimizer_creator is None.

• eps (float (default: 0.01)) – eps used for optimization, when optimizer_creator is None.

• max_norm (float | None (default: None)) – Max global norm of gradients for gradient clipping.

• n_steps_kl_warmup (int | None (default: None)) – Number of training steps (minibatches) to scale weight on KL divergences from min_kl_weight to max_kl_weight. Only activated when n_epochs_kl_warmup is set to None.

• n_epochs_kl_warmup (int | None (default: 400)) – Number of epochs to scale weight on KL divergences from min_kl_weight to max_kl_weight. Overrides n_steps_kl_warmup when both are not None.

Attributes table

training

Methods table

backward(*args, **kwargs)	Called to perform backward on the loss returned in training_step().
configure_optimizers()	Shim optimizer for PyTorch Lightning.
forward(*args, **kwargs)	Passthrough to the module's forward method.
get_optimizer_creator()	Get optimizer creator for the model.
jit_training_step(state, batch, rngs, **kwargs)	Jit training step.
jit_validation_step(state, batch, rngs, **kwargs)	Jit validation step.
optimizer_step(*args, **kwargs)	Override this method to adjust the default way the Trainer calls the optimizer.
set_train_state(params[, state])	Set the state of the module.
training_step(batch, batch_idx)	Training step for Jax.
transfer_batch_to_device(batch, device, ...)	Bypass Pytorch Lightning device management.
validation_step(batch, batch_idx)	Validation step for Jax.

Attributes

JaxTrainingPlan.training: bool

Methods

JaxTrainingPlan.backward(*args, **kwargs)

Called to perform backward on the loss returned in training_step(). Override this hook with your own implementation if you need to.

Parameters:

loss – The loss tensor returned by training_step(). If gradient accumulation is used, the loss here holds the normalized value (scaled by 1 / accumulation steps).

Example:

def backward(self, loss):
    loss.backward()

JaxTrainingPlan.configure_optimizers()

Shim optimizer for PyTorch Lightning.

PyTorch Lightning wants to take steps on an optimizer returned by this function in order to increment the global step count. See PyTorch Lighinting optimizer manual loop.

Here we provide a shim optimizer that we can take steps on at minimal computational cost in order to keep Lightning happy :).

JaxTrainingPlan.forward(*args, **kwargs)

Passthrough to the module’s forward method.

JaxTrainingPlan.get_optimizer_creator()

Get optimizer creator for the model.

Return type:

Callable[[], GradientTransformation]

static JaxTrainingPlan.jit_training_step(state, batch, rngs, **kwargs)

Jit training step.

JaxTrainingPlan.jit_validation_step(state, batch, rngs, **kwargs)

Jit validation step.

JaxTrainingPlan.optimizer_step(*args, **kwargs)

Override this method to adjust the default way the Trainer calls the optimizer.

By default, Lightning calls step() and zero_grad() as shown in the example. This method (and zero_grad()) won’t be called during the accumulation phase when Trainer(accumulate_grad_batches != 1). Overriding this hook has no benefit with manual optimization.

Parameters:

• epoch – Current epoch

• batch_idx – Index of current batch

• optimizer – A PyTorch optimizer

• optimizer_closure – The optimizer closure. This closure must be executed as it includes the calls to training_step(), optimizer.zero_grad(), and backward().

Examples:

def optimizer_step(self, epoch, batch_idx, optimizer, optimizer_closure):
    # Add your custom logic to run directly before `optimizer.step()`

    optimizer.step(closure=optimizer_closure)

    # Add your custom logic to run directly after `optimizer.step()`

JaxTrainingPlan.set_train_state(params, state=None)

Set the state of the module.

JaxTrainingPlan.training_step(batch, batch_idx)

Training step for Jax.

static JaxTrainingPlan.transfer_batch_to_device(batch, device, dataloader_idx)

Bypass Pytorch Lightning device management.

JaxTrainingPlan.validation_step(batch, batch_idx)

Validation step for Jax.