TransformerLightningModuleBase

class rectools.models.nn.transformers.lightning.TransformerLightningModuleBase(torch_model: TransformerBackboneBase, model_config: Dict[str, Any], dataset_schema: Dict[str, Any], item_external_ids: Union[Sequence[Hashable], ndarray], item_extra_tokens: Sequence[Hashable], data_preparator: TransformerDataPreparatorBase, lr: float, gbce_t: float, loss: str, verbose: int = 0, train_loss_name: str = 'train_loss', val_loss_name: str = 'val_loss', adam_betas: Tuple[float, float] = (0.9, 0.98), logits_t: float = 1, **kwargs: Any)

Bases: LightningModule

Base class for transfofmers lightning module. To change train procedure inherit from this class and pass your custom LightningModule to your model parameters.

Parameters

• torch_model (TransformerBackboneBase) – Torch model to make recommendations.

• model_config (Dict[str, Any]) – Model config.

• dataset_schema (DatasetSchemaDict) – Dataset schema.

• item_external_ids (ExternalIds) – External item ids from train dataset.

• item_extra_tokens (Sequence(Hashable)) – Elements used for sequence padding.

• lr (float) – Learning rate.

• gbce_t (float) – Calibration parameter for gBCE loss.

• loss (str, default "softmax") – Loss function.

• adam_betas (Tuple[float, float], default (0.9, 0.98)) – Coefficients for running averages of gradient and its square.

• data_preparator (TransformerDataPreparatorBase) – Data preparator.

• verbose (int, default 0) – Verbosity level.

• train_loss_name (str, default "train_loss") – Name of the training loss.

• val_loss_name (str, default "val_loss") – Name of the training loss.

• logits_t (float, default 1) – Scale factor for logits.

• kwargs (Any) –

Methods

configure_optimizers()	Choose what optimizers and learning-rate schedulers to use in optimization
get_loss_calculator()	Return loss calculator.
requires_negatives(loss)	Return flag for determining the need for negatives.
training_step(batch, batch_idx)	Training step.
validation_step(batch, batch_idx)	Validate step.

Attributes

epsilon_cosine_dist
u2i_dist_available

configure_optimizers() → Adam

Choose what optimizers and learning-rate schedulers to use in optimization

Return type

Adam

get_loss_calculator() → Optional[Callable[[Tensor, Tensor, Tensor], Tensor]]

Return loss calculator.

Return type

Optional[Callable[[Tensor, Tensor, Tensor], Tensor]]

static requires_negatives(loss: str) → Optional[bool]

Return flag for determining the need for negatives.

Parameters

loss (str) –

Return type

Optional[bool]

training_step(batch: Dict[str, Tensor], batch_idx: int) → Tensor

Training step.

Parameters

• batch (Dict[str, Tensor]) –

• batch_idx (int) –

Return type

Tensor

validation_step(batch: Dict[str, Tensor], batch_idx: int) → Dict[str, Tensor]

Validate step.

Parameters

• batch (Dict[str, Tensor]) –

• batch_idx (int) –

Return type

Dict[str, Tensor]