Zurück zu den Modellen
Forecaster

AuroraForecaster

API-ReferenzQuellcode
Categorical in XPred intMultivariate

Zero-shot forecaster wrapping Aurora via the aurora-model package.

Aurora is a multimodal time series foundation model supporting generative probabilistic forecasting. Besides multivariate support, it can optionally condition on free-text domain context and vision inputs.

Inference follows the official aurora-model API [2] and Hugging Face model card [1] examples.

Schnellstart

python
from sktime.forecasting.aurora import AuroraForecaster

estimator = AuroraForecaster(repo_id: str='DecisionIntelligence/Aurora', weights_filename: str='model.safetensors', cache_dir: str | None=None, force_download: bool=False, device: str | None=None, context_length: int | None=None, inference_token_len: int=48, num_samples: int=100, max_text_length: int=200, text: str | None=None, vision=None)

Parameter(11)

repo_idstr, default=”DecisionIntelligence/Aurora”
Hugging Face repository id for model weights.
weights_filenamestr, default=”model.safetensors”
Weights file name in the Hugging Face repository.
cache_dirstr, optional, default=None
Local cache directory for downloaded weights.
force_downloadbool, default=False
Whether to force re-download of weights from the Hub.
devicestr, optional, default=None

Device for inference. If None, uses CUDA when available, else CPU.

context_lengthint, optional, default=None

Number of trailing history steps passed to the model. If None, uses the full series seen at predict time.

inference_token_lenint, default=48
Patch length for inference. Using the series period length when known is recommended.
num_samplesint, default=100

Number of stochastic forecast trajectories from flow matching. Point forecasts use the sample mean. Quantile forecasts require num_samples > 1.

max_text_lengthint, default=200

Maximum token length when text is provided.

textstr, optional, default=None
Optional text context for multimodal forecasting (e.g. domain metadata or event descriptions). The same text is applied to every target variable (channel-independent inference).
visionPIL.Image, array-like, or torch.Tensor, optional, default=None

Optional external RGB image for multimodal conditioning. When None (default), Aurora renders a pseudo-image from the historical series internally (period-based 2D layout).

Set vision only when a real image should provide additional context alongside the series (e.g. a chart or domain photograph). Accepted inputs are passed to Aurora’s ViT preprocessor: a single PIL.Image, a list of images (batch size must match the model batch: 1 for univariate, n_vars for multivariate), a numpy array, or a torch.Tensor of shape (batch, 3, H, W) in RGB channel order. Images are resized to 224x224 and normalized inside the model.

Beispiele

>>> from sktime.datasets import load_airline
>>> from sktime.forecasting.aurora import AuroraForecaster
>>> from sktime.split import temporal_train_test_split
>>> y = load_airline ()
>>> y_train, _ = temporal_train_test_split (y)
>>> f = AuroraForecaster (num_samples = 10)
>>> f. fit (y_train)
>>> y_pred = f. predict (fh = [1, 2, 3 ]) Multimodal forecasting with text context:
>>> f = AuroraForecaster (
... text = "Monthly airline passenger totals.",
... num_samples = 10,
... )
>>> f. fit (y_train)
>>> y_pred = f. predict (fh = [1, 2, 3 ])

Referenzen