MODULE REFERENCE

torch

anomx.models.torch

A generic PyTorch network behind the Darts forecasting interface.

PyTorchModel

PyTorchModel(input_chunk_length: int = 12, output_chunk_length: int = 1, module: nn.Module | None = None, n_epochs: int = 20, batch_size: int = 32, learning_rate: float = 0.001, random_state: int = 42, device: str = 'cpu', loss_fn: nn.Module | None = None)

Bases: DartsModel

Train any torch.nn.Module on deterministic time-series windows.

The module receives (batch, input_chunk_length, channels) and must return (batch, output_chunk_length, channels). Omit it for a tiny linear baseline. Training accepts one or many Darts TimeSeries; prediction rolls the network forward for an arbitrary horizon. No normalization is implicit. Use Darts transformers fitted on the training split when scaling is needed.

The exported ONNX graph uses the identical tensor contract. The adjacent JSON manifest describes channel names, window sizes and training KPIs.

Signature parameters
Parameter	Type	Default
`input_chunk_length`	`int`	`12`
`output_chunk_length`	`int`	`1`
`module`	`nn.Module \| None`	`None`
`n_epochs`	`int`	`20`
`batch_size`	`int`	`32`
`learning_rate`	`float`	`0.001`
`random_state`	`int`	`42`
`device`	`str`	`'cpu'`
`loss_fn`	`nn.Module \| None`	`None`

Members

input_chunk_lengthproperty

# PyTorchModel.input_chunk_length Source ↗

input_chunk_length: int

Returns int

output_chunk_lengthproperty

# PyTorchModel.output_chunk_length Source ↗

output_chunk_length: int

Returns int

supports_multivariateproperty

# PyTorchModel.supports_multivariate Source ↗

supports_multivariate: bool

Returns bool

supports_sample_weightproperty

# PyTorchModel.supports_sample_weight Source ↗

supports_sample_weight: bool

Returns bool

supports_optimized_historical_forecastsproperty

# PyTorchModel.supports_optimized_historical_forecasts Source ↗

supports_optimized_historical_forecasts: bool

Returns bool

min_train_series_lengthproperty

# PyTorchModel.min_train_series_length Source ↗

min_train_series_length: int

Returns int

min_train_samplesproperty

# PyTorchModel.min_train_samples Source ↗

min_train_samples: int

Returns int

extreme_lagsproperty

# PyTorchModel.extreme_lags Source ↗

extreme_lags: tuple[int, int, None, None, None, None, int]

Returns tuple[int, int, None, None, None, None, int]

fitmethod

# PyTorchModel.fit Source ↗

fit(series: TimeSeries | Dataset | Sequence[TimeSeries | Dataset], past_covariates: TimeSeries | Sequence[TimeSeries] | None = None, future_covariates: TimeSeries | Sequence[TimeSeries] | None = None, verbose: bool | None = None, *, on_epoch: Callable[[dict[str, float]], Any] | None = None) -> PyTorchModel

Fit on chronological sliding windows and record mean loss per epoch.

Signature parameters
Parameter	Type	Default
`series`	`TimeSeries \| Dataset \| Sequence[TimeSeries \| Dataset]`	`required`
`past_covariates`	`TimeSeries \| Sequence[TimeSeries] \| None`	`None`
`future_covariates`	`TimeSeries \| Sequence[TimeSeries] \| None`	`None`
`verbose`	`bool \| None`	`None`
`on_epoch`keyword only	`Callable[[dict[str, float]], Any] \| None`	`None`

Returns PyTorchModel

predictmethod

# PyTorchModel.predict Source ↗

predict(n: int, series: TimeSeries | Dataset | Sequence[TimeSeries | Dataset] | None = None, past_covariates: TimeSeries | Sequence[TimeSeries] | None = None, future_covariates: TimeSeries | Sequence[TimeSeries] | None = None, num_samples: int = 1, verbose: bool | None = None, **kwargs: object) -> TimeSeries | list[TimeSeries]

Autoregressively forecast n steps, preserving components and index.

Signature parameters
Parameter	Type	Default
`n`	`int`	`required`
`series`	`TimeSeries \| Dataset \| Sequence[TimeSeries \| Dataset] \| None`	`None`
`past_covariates`	`TimeSeries \| Sequence[TimeSeries] \| None`	`None`
`future_covariates`	`TimeSeries \| Sequence[TimeSeries] \| None`	`None`
`num_samples`	`int`	`1`
`verbose`	`bool \| None`	`None`
`**kwargs`	`object`	`—`

Returns TimeSeries | list[TimeSeries]

export_onnxmethod

# PyTorchModel.export_onnx Source ↗

export_onnx(path: str | Path, *, opset_version: int = 17) -> Path

Export a checked, self-contained ONNX graph and JSON input manifest.

Signature parameters
Parameter	Type	Default
`path`	`str \| Path`	`required`
`opset_version`keyword only	`int`	`17`

Returns Path

Inherited members