Adding a covariate as time_varying_unknown_real to deepAR model

[janavl]

• PyTorch-Forecasting version: 0.10.3
• Torch version: 1.13.1
• Python version: 3.9.11
• Operating System:
• PyTorch-Lightning version: 1.9.5

Expected behavior

I want to add a numerical feature to the deepAR model for which future values are unknown (the feature is the price of a product, that changes over time based on discounts, promotion periods etc. and is unknown for the future). Reading the DeepAR documentation, I expected to be able to add price as a time_varying_unkown_real in the TimeSeriesDataSet. I expected the model to run without errors.

Actual behavior

However, I receive the following error when fitting the model: input.size(-1) must be equal to input_size. Expected 244, got 245.

Code to reproduce the problem

training = TimeSeriesDataSet(
        data,
        time_idx="time_idx",
        allow_missing_timesteps=True,
        target="target_name",
        group_ids=["article"],
        min_encoder_length=args.min_encoder_length,
        max_encoder_length=args.max_encoder_length,
        min_prediction_length=args.min_prediction_length,
        max_prediction_length=args.max_prediction_length,
        static_categoricals=static_cat_cols,
        static_reals=[],
        time_varying_known_categoricals=args.known_categoricals + ["campaigns"],
        variable_groups={"campaigns": campaigns},
        time_varying_known_reals=["time_idx"],
        time_varying_unknown_categoricals=[],
        time_varying_unknown_reals=["target_name","price"],
        add_relative_time_idx=True,
        add_target_scales=True,
        add_encoder_length=True,
        target_normalizer=GroupNormalizer(
            groups=["article"],
            transformation=None,
            center=True,
            eps=1e-4,
        ), 
    )

train_dataloader = training.to_dataloader(
        train=True, batch_size=args.batch_size, num_workers=args.num_workers, shuffle=True
    )

model = DeepAR.from_dataset(
        training,
        cell_type="LSTM",
        rnn_layers=args.rnn_layers,
        learning_rate=args.learning_rate,
        hidden_size=args.hidden_dimension,
        dropout=args.dropout_rate,
        loss=loss_type,
        log_interval=1,
        optimizer="ranger",
        reduce_on_plateau_patience=4,
        time_varying_reals_decoder=['time_idx','price'],
        time_varying_reals_encoder=['time_idx','price','target_name'],
    )

trainer = pl.Trainer(
        max_epochs=args.max_epochs,
        gpus=n_gpus,
        weights_summary="top",
        gradient_clip_val=gradient_clip_value,
        gradient_clip_algorithm="value",
        callbacks=[early_stop_callback],
    )

trainer_object = trainer.fit(model, train_dataloaders=train_dataloader)

When I fit the model (trainer.fit()), it fails and gives error: input.size(-1) must be equal to input_size. Expected 244, got 245.

Please help me understand if it is possible to add unkown reals other than the target.