Back to models
Forecaster

NeuralForecastTCN

API referenceSource
Categorical in XPred int insampleExogenous

NeuralForecast TCN model.

Quickstart

python
from sktime.forecasting.neuralforecast import NeuralForecastTCN

estimator = NeuralForecastTCN(freq: str | int='auto', local_scaler_type: Literal ['standard', 'robust', 'robust-iqr', 'minmax', 'boxcox' ] | None=None, futr_exog_list: list [str ] | None=None, verbose_fit: bool=False, verbose_predict: bool=False, input_size: int=-1, inference_input_size: int=-1, kernel_size: int=2, dilations: list [int ] | None=None, encoder_hidden_size: int=200, encoder_activation: str='ReLU', context_size: int=10, decoder_hidden_size: int=200, decoder_layers: int=2, loss=None, valid_loss=None, max_steps: int=1000, learning_rate: float=0.001, num_lr_decays: int=-1, early_stop_patience_steps: int=-1, val_check_steps: int=100, batch_size: int=32, valid_batch_size: int | None=None, scaler_type: str='robust', random_seed: int=1, num_workers_loader=0, drop_last_loader=False, optimizer=None, optimizer_kwargs: dict | None=None, lr_scheduler=None, lr_scheduler_kwargs: dict | None=None, trainer_kwargs: dict | None=None, broadcasting: bool=False)

Parameters(33)

freqUnion[str, int] (default=”auto”)

frequency of the data, see available frequencies [4] from pandas use int freq when using RangeIndex in y

default (“auto”) interprets freq from ForecastingHorizon in fit

local_scaler_typestr (default=None)

scaler to apply per-series to all features before fitting, which is inverted after predicting

can be one of the following:

  • ‘standard’

  • ‘robust’

  • ‘robust-iqr’

  • ‘minmax’

  • ‘boxcox’

futr_exog_liststr list, (default=None)
future exogenous variables
verbose_fitbool (default=False)
print processing steps during fit
verbose_predictbool (default=False)
print processing steps during predict
input_sizeint (default=-1)

maximum sequence length for truncated train backpropagation

default (-1) uses all history

inference_input_sizeint (default=-1)

maximum sequence length for truncated inference

default (-1) uses all history

kernel_sizeint (default=2)
size of the convolving kernel
dilationsint list (default=None)

controls the temporal spacing between the kernel points also known as the à trous algorithm by default set to [1, 2, 4, 8, 16]

encoder_hidden_sizeint (default=200)
units for the TCN’s hidden state size
encoder_activationstr (default=”ReLU”)
type of TCN activation from tanh or relu
context_sizeint (default=10)
size of context vector for each timestamp on the forecasting window
decoder_hidden_sizeint (default=200)
size of hidden layer for the MLP decoder
decoder_layersint (default=2)
number of layers for the MLP decoder
losspytorch module (default=None)
instantiated train loss class from losses collection [Rcdb94044b921-5]
valid_losspytorch module (default=None)
instantiated validation loss class from losses collection [Rcdb94044b921-5]
max_stepsint (default=1000)
maximum number of training steps
learning_ratefloat (default=1e-3)
learning rate between (0, 1)
num_lr_decaysint (default=-1)
number of learning rate decays, evenly distributed across max_steps
early_stop_patience_stepsint (default=-1)
number of validation iterations before early stopping
val_check_stepsint (default=100)
number of training steps between every validation loss check
batch_sizeint (default=32)
number of different series in each batch
valid_batch_sizeOptional[int] (default=None)
number of different series in each validation and test batch
scaler_typestr (default=”robust”)
type of scaler for temporal inputs normalization
random_seedint (default=1)
random_seed for pytorch initializer and numpy generators
num_workers_loaderint (default=0)

workers to be used by TimeSeriesDataLoader

drop_last_loaderbool (default=False)

whether TimeSeriesDataLoader drops last non-full batch

optimizerpytorch optimizer (default=None) [Rcdb94044b921-7]

optimizer to use for training, if passed with None defaults to Adam

optimizer_kwargsdict (default=None) [Rcdb94044b921-8]
dict of parameters to pass to the user defined optimizer
lr_schedulerpytorch learning rate scheduler (default=None) [Rcdb94044b921-9]

user specified lr_scheduler instead of the default choice StepLR [Rcdb94044b921-10]

lr_scheduler_kwargsdict (default=None)

list of parameters used by the user specified lr_scheduler

trainer_kwargsdict (default=None)
keyword trainer arguments inherited from PyTorch Lighning’s trainer [Rcdb94044b921-6]
broadcastingbool (default=False)
if True, a model will be fit per time series. Panels, e.g., multiindex data input, will be broadcasted to single series, and for each single series, one copy of this forecaster will be applied.

Examples

>>> 
>>> # importing necessary libraries
>>> from sktime.datasets import load_longley
>>> from sktime.forecasting.neuralforecast import NeuralForecastTCN
>>> from sktime.split import temporal_train_test_split
>>> 
>>> # loading the Longley dataset and splitting it into train and test subsets
>>> y, X = load_longley ()
>>> y_train, y_test, X_train, X_test = temporal_train_test_split (y, X, test_size = 4)
>>> 
>>> # creating model instance configuring the hyperparameters
>>> model = NeuralForecastTCN (
... "A-DEC", futr_exog_list = ["ARMED", "POP" ], max_steps = 5
... )
>>> 
>>> # fitting the model
>>> model. fit (y_train, X = X_train, fh = [1, 2, 3, 4 ]) Seed set to 1 Epoch 4: 100%|████████████████████████████| 1/1 [00:00<00:00, 48.19it/s, v_num=5, train_loss_step=0.833, train_loss_epoch=0.833] NeuralForecastTCN(freq='A-DEC', futr_exog_list=['ARMED', 'POP'], max_steps=5)
>>> 
>>> # getting point predictions
>>> model. predict (X = X_test) Predicting DataLoader 0: 100%|███████████████████████████████████████████████████████████████████| 1/1 [00:00<00:00, 183.39it/s] 1959 63611.593750 1960 63528.542969 1961 63529.167969 1962 63718.667969 Freq: A-DEC, Name: TOTEMP, dtype: float64
>>>

References

  1. [1 ] https://nixtlaverse.nixtla.io/neuralforecast/models.tcn.html#tcn [2 ] https://nixtlaverse.nixtla.io/neuralforecast/core.html#neuralforecast [3 ] https://github.com/Nixtla/neuralforecast/ [4 ] https://pandas.pydata.org/pandas-docs/stable/user_guide/timeseries.html#offset-aliases.. [Rcdb94044b921-5] https://nixtlaverse.nixtla.io/neuralforecast/losses.pytorch.html.. [Rcdb94044b921-6] https://lightning.ai/docs/pytorch/stable/api/pytorch_lightning.trainer.trainer.Trainer.html#lightning.pytorch.trainer.trainer.Trainer.. [Rcdb94044b921-7] https://pytorch.org/docs/stable/optim.html.. [Rcdb94044b921-8] https://pytorch.org/docs/stable/optim.html#algorithms.. [Rcdb94044b921-9] https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.LRScheduler.html.. [Rcdb94044b921-10] https://pytorch.org/docs/stable/generated/torch.optim.lr_scheduler.StepLR.html