TimeLLMForecaster

Interface to the Time-LLM.

Quickstart

from sktime.forecasting.time_llm import TimeLLMForecaster

estimator = TimeLLMForecaster(task_name='long_term_forecast', pred_len=24, seq_len=96, llm_model='GPT2', llm_layers=3, llm_dim=768, patch_len=16, stride=8, d_model=128, d_ff=128, n_heads=4, dropout=0.1, device: str | None=None, prompt_domain=False)

Parameters(12)

task_namestr, default=’long_term_forecast’

Task to perform - can be one of [‘long_term_forecast’, ‘short_term_forecast’].

pred_lenint, default=24

Forecast horizon - number of time steps to predict.

seq_lenint, default=96

Length of input sequence.

llm_modelstr, default=’GPT2’

LLM model to use - can be one of [‘GPT2’, ‘LLAMA’, ‘BERT’].

llm_layersint, default=3

Number of transformer layers to use from LLM.

patch_lenint, default=16

Length of patches for patch embedding.

strideint, default=8

Stride between patches.

d_modelint, default=128

Model dimension.

d_ffint, default=128

Feed-forward dimension.

n_headsint, default=4

Number of attention heads.

dropoutfloat, default=0.1

Dropout rate.

devicestr, default=’cuda’ if available else ‘cpu’

Device to run model on.

Examples

>>> from sktime.forecasting.time_llm import TimeLLMForecaster
>>> from sktime.datasets import load_airline
>>> y = load_airline ()
>>> forecaster = TimeLLMForecaster (
... pred_len = 36,
... seq_len = 96,
... llm_model = 'GPT2'
... )
>>> forecaster. fit (y, fh = [1 ]) TimeLLMForecaster(pred_len=36)
>>> y_pred = forecaster. predict (fh = [1 ])

References

1. [1 ] https://github.com/KimMeen/Time-LLM [2 ] Ming Jin, Shiyu Wang, Lintao Ma, Zhixuan Chu, James Y. Zhang, Xiaoming Shi, Pin-Yu Chen, Yuxuan Liang, Yuan-Fang Li, Shirui Pan, Qingsong Wen. Time-LLM: Time Series Forecasting by Reprogramming Large Language Models. https://arxiv.org/abs/2310.01728.