Multivariate

chronax.gluonts_predictor.Multivariate

A dataclass used for configuring multivariate batching parameters.

__init__(self, batch_size=16)

Parameter	Type	Default	Description
batch_size	int	16	(undocumented)

TotoSampleForecastGenerator

chronax.gluonts_predictor.TotoSampleForecastGenerator · inherits SampleForecastGenerator

Wrapper class for generating GluonTS forecasts from Toto models.

This class differs from the behavior of standard GluonTS generators in one major way:

Most GluonTS models are encoder-style models with a fixed prediction length, whereas Toto is a decoder-only model that can generate forecasts of arbitrary length using autoregressive sampling. This wrapper handles the iterative sampling process to generate forecasts of the required length, which can be set to any value when the generator is called.

In addition, this class can also provide both point predictions (when num_samples is None) or samples from the predictive distribution (when num_samples is not None). Note that when num_samples is not None, the memory usage is significantly higher, so the batch_size and num_variates should be set accordingly.

__call__(self, inference_data_loader, prediction_net, prediction_length, num_samples, input_names, output_transform, mode=Multivariate(batch_size=4), samples_per_batch=10, use_kv_cache=False) -> Iterator[SampleForecast]

Parameter	Type	Default	Description
inference_data_loader	DataLoader	-	(undocumented)
prediction_net	Toto	-	(undocumented)
prediction_length	int	-	(undocumented)
num_samples	Optional[int]	-	(undocumented)
input_names	List[str]	-	(undocumented)
output_transform	Optional[OutputTransform]	-	(undocumented)
mode	Multivariate	Multivariate(batch_size=4)	(undocumented)
samples_per_batch	int	10	(undocumented)
use_kv_cache	bool	False	(undocumented)

TotoSampleForecast

chronax.gluonts_predictor.TotoSampleForecast · inherits SampleForecast

Wrapper around GluonTS's SampleForecast class that adds a deterministic mean forecast to the samples. By default, SampleForecast calculates the mean forecast by taking the mean of the samples, but since Toto predicts parametrically, we can calculate the mean forecast more efficiently by directly computing the mean from the model's output. This is useful for evaluation metrics that require a point prediction.

__init__(self, samples, mean, start_date, item_id=None, info=None)

Parameter	Type	Default	Description
samples	np.ndarray	-	(undocumented)
mean	np.ndarray	-	(undocumented)
start_date	pd.Period	-	(undocumented)
item_id	Optional[str]	None	(undocumented)
info	Optional[dict]	None	(undocumented)

TotoPredictor

chronax.gluonts_predictor.TotoPredictor · inherits PyTorchPredictor

Predictor class for Toto models in GluonTS. This class is a thin wrapper that adapts Toto to the GluonTS interface for evaluation and forecasting. Most of the actual work is done by the TotoSampleForecastGenerator class.

__init__(self, input_names, prediction_net, prediction_length, input_transform, forecast_generator, output_transform=None, lead_time=0, device='auto', mode=Multivariate(batch_size=16), samples_per_batch=10)

Parameter	Type	Default	Description
input_names	List[str]	-	(undocumented)
prediction_net	torch.nn.Module	-	(undocumented)
prediction_length	int	-	(undocumented)
input_transform	Transformation	-	(undocumented)
forecast_generator	TotoSampleForecastGenerator	-	(undocumented)
output_transform	Optional[OutputTransform]	None	(undocumented)
lead_time	int	0	(undocumented)
device	str \| torch.device	"auto"	(undocumented)
mode	Multivariate	Multivariate(batch_size=16)	(undocumented)
samples_per_batch	int	10	(undocumented)

create_for_eval(cls, model, prediction_length, context_length, mode=Multivariate(batch_size=1), samples_per_batch=10) -> "TotoPredictor"

Parameter	Type	Default	Description
model	Toto	-	(undocumented)
prediction_length	int	-	(undocumented)
context_length	int	-	(undocumented)
mode	Multivariate	Multivariate(batch_size=1)	(undocumented)
samples_per_batch	int	10	(undocumented)

custom_stack_fn(self, data, device=None) -> Dict[str, Union[torch.Tensor, List[Any]]]

Custom stack function for the InferenceDataLoader. Attempts to use GluonTS's batchify for stacking. Falls back to manual logic with consistent padding.

Parameter	Type	Default	Description
data	List[Dict[str, Any]]	-	(undocumented)
device	torch.types.Device	None	Ensure we have a device to use

predict(self, dataset, num_samples=None, use_kv_cache=False, eval=True) -> Iterator[Forecast]

Parameter	Type	Default	Description
dataset	Dataset	-	(undocumented)
num_samples	Optional[int]	None	(undocumented)
use_kv_cache	bool	False	(undocumented)
eval	bool	True	(undocumented)