etna.models.nn.TFTNativeModel#

class TFTNativeModel(encoder_length: int, decoder_length: int, n_heads: int = 4, num_layers: int = 2, dropout: float = 0.1, hidden_size: int = 160, lr: float = 0.001, static_categoricals: List[str] | None = None, static_reals: List[str] | None = None, time_varying_categoricals_encoder: List[str] | None = None, time_varying_categoricals_decoder: List[str] | None = None, time_varying_reals_encoder: List[str] | None = None, time_varying_reals_decoder: List[str] | None = None, num_embeddings: Dict[str, int] | None = None, loss: Module | None = None, train_batch_size: int = 16, test_batch_size: int = 16, optimizer_params: dict | None = None, trainer_params: dict | None = None, train_dataloader_params: dict | None = None, test_dataloader_params: dict | None = None, val_dataloader_params: dict | None = None, split_params: dict | None = None)[source]#

Bases: DeepBaseModel

TFT model. For more details read the paper.

Model needs label encoded inputs for categorical features, for that purposes use LabelEncoderTransform. Feature values that were not seen during fit should be set to NaN for expected behaviour with strategy=”none”

Passed feature values aren’t validated on being static or being correctly label encoded.

Note

This model requires torch extension to be installed. Read more about this at installation page.

Init TFT model.

Parameters:

  • encoder_length (int) – encoder length

  • decoder_length (int) – decoder length

  • n_heads (int) – number of heads in Multi-Head Attention

  • num_layers (int) – number of layers in LSTM layer

  • dropout (float) – dropout rate

  • hidden_size (int) – size of the hidden state

  • lr (float) – learning rate

  • static_categoricals (List[str] | None) – categorical features that have one unique feature value for the whole series, e.g. segment. The first value in the series is passed to batch for each feature.

  • static_reals (List[str] | None) – continuous features that have one unique feature value for the whole series. The first value in the series is passed to batch for each feature.

  • time_varying_categoricals_encoder (List[str] | None) – time varying categorical features for encoder

  • time_varying_categoricals_decoder (List[str] | None) – time varying categorical features for decoder (known for future)

  • time_varying_reals_encoder (List[str] | None) – time varying continuous features for encoder, default to target

  • time_varying_reals_decoder (List[str] | None) – time varying continuous features for decoder (known for future)

  • num_embeddings (Dict[str, int] | None) – dictionary where keys are feature names and values are number of unique values of that features

  • loss (torch.nn.Module | None) – loss function

  • train_batch_size (int) – batch size for training

  • test_batch_size (int) – batch size for testing

  • optimizer_params (dict | None) – parameters for optimizer for Adam optimizer (api reference torch.optim.Adam)

  • trainer_params (dict | None) – Pytorch lightning trainer parameters (api reference pytorch_lightning.trainer.trainer.Trainer)

  • train_dataloader_params (dict | None) – parameters for train dataloader like sampler for example (api reference torch.utils.data.DataLoader)

  • test_dataloader_params (dict | None) – parameters for test dataloader

  • val_dataloader_params (dict | None) – parameters for validation dataloader

  • split_params (dict | None) –

    dictionary with parameters for torch.utils.data.random_split() for train-test splitting

    • train_size: (float) value from 0 to 1 - fraction of samples to use for training

    • generator: (Optional[torch.Generator]) - generator for reproducible train-test splitting

    • torch_dataset_size: (Optional[int]) - number of samples in dataset, in case of dataset not implementing __len__

Methods

fit(ts)

Fit model.

forecast(ts, prediction_size[, ...])

Make predictions.

get_model()

Get model.

load(path[, ts])

Load an object.

params_to_tune()

Get default grid for tuning hyperparameters.

predict(ts, prediction_size[, return_components])

Make predictions.

raw_fit(torch_dataset)

Fit model on torch like Dataset.

raw_predict(torch_dataset)

Make inference on torch like Dataset.

save(path)

Save the object.

set_params(**params)

Return new object instance with modified parameters.

to_dict()

Collect all information about etna object in dict.

Attributes

This class stores its __init__ parameters as attributes.

context_size

Context size of the model.
fit(ts: TSDataset) → DeepBaseModel[source]#

Fit model.

Parameters:

ts (TSDataset) – TSDataset with features

Returns:

Model after fit

Return type:

DeepBaseModel

forecast(ts: TSDataset, prediction_size: int, return_components: bool = False) → TSDataset[source]#

Make predictions.

This method will make autoregressive predictions.

Parameters:

  • ts (TSDataset) – Dataset with features and expected decoder length for context

  • prediction_size (int) – Number of last timestamps to leave after making prediction. Previous timestamps will be used as a context.

  • return_components (bool) – If True additionally returns forecast components

Returns:

Dataset with predictions

Return type:

TSDataset

get_model() → DeepBaseNet[source]#

Get model.

Returns:

Torch Module

Return type:

DeepBaseNet

classmethod load(path: Path, ts: TSDataset | None = None) → Self[source]#

Load an object.

Warning

This method uses dill module which is not secure. It is possible to construct malicious data which will execute arbitrary code during loading. Never load data that could have come from an untrusted source, or that could have been tampered with.

Parameters:

  • path (Path) – Path to load object from.

  • ts (TSDataset | None) – TSDataset to set into loaded pipeline.

Returns:

Loaded object.

Return type:

Self

params_to_tune() → Dict[str, BaseDistribution][source]#

Get default grid for tuning hyperparameters.

This grid tunes parameters: num_layers, n_heads, hidden_size, lr, dropout. Other parameters are expected to be set by the user.

Returns:

Grid to tune.

Return type:

Dict[str, BaseDistribution]

predict(ts: TSDataset, prediction_size: int, return_components: bool = False) → TSDataset[source]#

Make predictions.

This method will make predictions using true values instead of predicted on a previous step. It can be useful for making in-sample forecasts.

Parameters:

  • ts (TSDataset) – Dataset with features and expected decoder length for context

  • prediction_size (int) – Number of last timestamps to leave after making prediction. Previous timestamps will be used as a context.

  • return_components (bool) – If True additionally returns prediction components

Returns:

Dataset with predictions

Return type:

TSDataset

raw_fit(torch_dataset: Dataset) → DeepBaseModel[source]#

Fit model on torch like Dataset.

Parameters:

torch_dataset (Dataset) – Torch like dataset for model fit

Returns:

Model after fit

Return type:

DeepBaseModel

raw_predict(torch_dataset: Dataset) → Dict[Tuple[str, str], ndarray][source]#

Make inference on torch like Dataset.

Parameters:

torch_dataset (Dataset) – Torch like dataset for model inference

Returns:

Dictionary with predictions

Return type:

Dict[Tuple[str, str], ndarray]

save(path: Path)[source]#

Save the object.

Parameters:

path (Path) – Path to save object to.

set_params(**params: dict) → Self[source]#

Return new object instance with modified parameters.

Method also allows to change parameters of nested objects within the current object. For example, it is possible to change parameters of a model in a Pipeline.

Nested parameters are expected to be in a <component_1>.<...>.<parameter> form, where components are separated by a dot.

Parameters:

**params (dict) – Estimator parameters

Returns:

New instance with changed parameters

Return type:

Self

Examples

>>> from etna.pipeline import Pipeline
>>> from etna.models import NaiveModel
>>> from etna.transforms import AddConstTransform
>>> model = model=NaiveModel(lag=1)
>>> transforms = [AddConstTransform(in_column="target", value=1)]
>>> pipeline = Pipeline(model, transforms=transforms, horizon=3)
>>> pipeline.set_params(**{"model.lag": 3, "transforms.0.value": 2})
Pipeline(model = NaiveModel(lag = 3, ), transforms = [AddConstTransform(in_column = 'target', value = 2, inplace = True, out_column = None, )], horizon = 3, )
to_dict()[source]#

Collect all information about etna object in dict.

property context_size: int[source]#

Context size of the model.