class pytorch_forecasting.models.base_model.AutoRegressiveBaseModel(log_interval: Union[int, float] = - 1, log_val_interval: Optional[Union[float, int]] = None, learning_rate: Union[float, List[float]] = 0.001, log_gradient_flow: bool = False, loss: pytorch_forecasting.metrics.Metric = SMAPE(), logging_metrics: torch.nn.modules.container.ModuleList = ModuleList(), reduce_on_plateau_patience: int = 1000, reduce_on_plateau_min_lr: float = 1e-05, weight_decay: float = 0.0, optimizer_params: Optional[Dict[str, Any]] = None, monotone_constaints: Dict[str, int] = {}, output_transformer: Optional[Callable] = None, optimizer='ranger')[source]

Bases: pytorch_forecasting.models.base_model.BaseModel

Model with additional methods for autoregressive models.

Adds in particular the decode_autoregressive() method for making auto-regressive predictions.

Assumes the following hyperparameters:

  • target (str) – name of target variable

  • target_lags (Dict[str, Dict[str, int]]) – dictionary of target names mapped each to a dictionary of corresponding lagged variables and their lags. Lags can be useful to indicate seasonality to the models. If you know the seasonalit(ies) of your data, add at least the target variables with the corresponding lags to improve performance. Defaults to no lags, i.e. an empty dictionary.

BaseModel for timeseries forecasting from which to inherit from

  • log_interval (Union[int, float], optional) – Batches after which predictions are logged. If < 1.0, will log multiple entries per batch. Defaults to -1.

  • log_val_interval (Union[int, float], optional) – batches after which predictions for validation are logged. Defaults to None/log_interval.

  • learning_rate (float, optional) – Learning rate. Defaults to 1e-3.

  • log_gradient_flow (bool) – If to log gradient flow, this takes time and should be only done to diagnose training failures. Defaults to False.

  • loss (Metric, optional) – metric to optimize, can also be list of metrics. Defaults to SMAPE().

  • logging_metrics (nn.ModuleList[MultiHorizonMetric]) – list of metrics that are logged during training. Defaults to [].

  • reduce_on_plateau_patience (int) – patience after which learning rate is reduced by a factor of 10. Defaults to 1000

  • reduce_on_plateau_min_lr (float) – minimum learning rate for reduce on plateua learning rate scheduler. Defaults to 1e-5

  • weight_decay (float) – weight decay. Defaults to 0.0.

  • optimizer_params (Dict[str, Any]) – additional parameters for the optimizer. Defaults to {}.

  • monotone_constaints (Dict[str, int]) – dictionary of monotonicity constraints for continuous decoder variables mapping position (e.g. "0" for first position) to constraint (-1 for negative and +1 for positive, larger numbers add more weight to the constraint vs. the loss but are usually not necessary). This constraint significantly slows down training. Defaults to {}.

  • output_transformer (Callable) – transformer that takes network output and transforms it to prediction space. Defaults to None which is equivalent to lambda out: out["prediction"].

  • optimizer (str) – Optimizer, “ranger”, “sgd”, “adam”, “adamw” or class name of optimizer in torch.optim. Alternatively, a class or function can be passed which takes parameters as first argument and a lr argument (optionally also weight_decay) Defaults to “ranger”.


decode_autoregressive(decode_one, …)

Make predictions in auto-regressive manner.

from_dataset(dataset, **kwargs)

Create model from dataset.


Convert network output to rescaled and normalized prediction.

decode_autoregressive(decode_one: Callable, first_target: Union[List[torch.Tensor], torch.Tensor], first_hidden_state: Any, target_scale: Union[List[torch.Tensor], torch.Tensor], n_decoder_steps: int, **kwargs) Union[List[torch.Tensor], torch.Tensor][source]

Make predictions in auto-regressive manner.

Supports only continuous targets.

  • decode_one (Callable) –

    function that takes at least the following arguments:

    • idx (int): index of decoding step (from 0 to n_decoder_steps-1)

    • lagged_targets (List[torch.Tensor]): list of normalized targets. List is idx + 1 elements long with the most recent entry at the end, i.e. previous_target = lagged_targets[-1] and in general lagged_targets[-lag].

    • hidden_state (Any): Current hidden state required for prediction. Keys are variable names. Only lags that are greater than idx are included.

    • additional arguments are not dynamic but can be passed via the **kwargs argument

    And returns tuple of (not rescaled) network prediction output and hidden state for next auto-regressive step.

  • first_target (Union[List[torch.Tensor], torch.Tensor]) – first target value to use for decoding

  • first_hidden_state (Any) – first hidden state used for decoding

  • target_scale (Union[List[torch.Tensor], torch.Tensor]) – target scale as in x

  • n_decoder_steps (int) – number of decoding/prediction steps

  • **kwargs – additional arguments that are passed to the decode_one function.


re-scaled prediction

Return type

Union[List[torch.Tensor], torch.Tensor]


LSTM/GRU decoder

def decode(self, x, hidden_state):
    # create input vector
    input_vector = x["decoder_cont"].clone()
    input_vector[..., self.target_positions] = torch.roll(
        input_vector[..., self.target_positions],
    # but this time fill in missing target from encoder_cont at the first time step instead of
    # throwing it away
    last_encoder_target = x["encoder_cont"][
        torch.arange(x["encoder_cont"].size(0), device=x["encoder_cont"].device),
        x["encoder_lengths"] - 1,
    input_vector[:, 0, self.target_positions] = last_encoder_target

    if  # training mode
        decoder_output, _ = self.rnn(

        # from hidden state size to outputs
        if isinstance(, str):  # single target
            output = self.distribution_projector(decoder_output)
            output = [projector(decoder_output) for projector in self.distribution_projector]

        # predictions are not yet rescaled -> so rescale now
        return self.transform_output(output, target_scale=target_scale)

    else:  # prediction mode
        target_pos = self.target_positions

        def decode_one(idx, lagged_targets, hidden_state):
            x = input_vector[:, [idx]]
            x[:, 0, target_pos] = lagged_targets[-1]  # overwrite at target positions

            # overwrite at lagged targets positions
            for lag, lag_positions in lagged_target_positions.items():
                if idx > lag:  # only overwrite if target has been generated
                    x[:, 0, lag_positions] = lagged_targets[-lag]

            decoder_output, hidden_state = self.rnn(x, hidden_state)
            decoder_output = decoder_output[:, 0]  # take first timestep
            # from hidden state size to outputs
            if isinstance(, str):  # single target
                output = self.distribution_projector(decoder_output)
                output = [projector(decoder_output) for projector in self.distribution_projector]
            return output, hidden_state

        # make predictions which are fed into next step
        output = self.decode_autoregressive(
            first_target=input_vector[:, 0, target_pos],

        # predictions are already rescaled
        return output
classmethod from_dataset(dataset:, **kwargs) pytorch_lightning.core.lightning.LightningModule[source]

Create model from dataset.

  • dataset – timeseries dataset

  • **kwargs – additional arguments such as hyperparameters for model (see __init__())



output_to_prediction(normalized_prediction_parameters: torch.Tensor, target_scale: Union[List[torch.Tensor], torch.Tensor], **kwargs) Tuple[Union[List[torch.Tensor], torch.Tensor], torch.Tensor][source]

Convert network output to rescaled and normalized prediction.

Function is typically not called directly but via decode_autoregressive().

  • normalized_prediction_parameters (torch.Tensor) – network prediction output

  • target_scale (Union[List[torch.Tensor], torch.Tensor]) – target scale to rescale network output

  • **kwargs – extra arguments for dictionary passed to transform_output() method.


tuple of rescaled prediction and

normalized prediction (e.g. for input into next auto-regressive step)

Return type

Tuple[Union[List[torch.Tensor], torch.Tensor], torch.Tensor]

property lagged_target_positions: Dict[int, torch.LongTensor]

Positions of lagged target variable(s) in covariates.


dictionary mapping integer lags to tensor of variable positions.

Return type

Dict[int, torch.LongTensor]

property target_positions: torch.LongTensor

Positions of target variable(s) in covariates.


tensor of positions.

Return type