XLSTMForecaster

xLSTM Forecaster for time series prediction using Extended LSTM architecture.

Quickstart

from sktime.forecasting.xlstm import XLSTMForecaster

estimator = XLSTMForecaster(input_size=1, hidden_size=64, num_layers=2, block_types=None, num_heads=1, dropout=0.1, learning_rate=0.001, batch_size=32, n_epochs=50, sequence_length=20, device=None)

Parameters(11)

input_sizeint, default=1

Number of input features

hidden_sizeint, default=64

Hidden state size for xLSTM blocks

num_layersint, default=2

Number of xLSTM layers

block_typeslist, default=None

List of block types (‘slstm’ or ‘mlstm’). If None, uses all ‘slstm’

num_headsint, default=1

Number of attention heads for mLSTM blocks

dropoutfloat, default=0.1

Dropout probability

learning_ratefloat, default=0.001

Learning rate for optimization

batch_sizeint, default=32

Batch size for training

n_epochsint, default=50

Number of training epochs

sequence_lengthint, default=20

Length of input sequences

devicestr, default=None

Device to use (‘cuda’ or ‘cpu’). If None, auto-detects

Examples

>>> from sktime.datasets import load_airline
>>> from sktime.forecasting.xlstm import XLSTMForecaster
>>> y = load_airline ()
>>> forecaster = XLSTMForecaster (
... hidden_size = 32,
... num_layers = 2,
... n_epochs = 10
... )
>>> forecaster. fit (y) XLSTMForecaster(
... )
>>> y_pred = forecaster. predict (fh = [1, 2, 3 ])