Note

We release a open codebase OpenLTM to explore the design philosophy of large time-series models, which contains a simple pipeline to train large time-series models :)

Timer (Large Time-Series Model)

This repo provides official code, datasets and checkpoints for Timer: Generative Pre-trained Transformers Are Large Time Series Models. [Poster], [Slides].

Updates

🚩 News (2024.12) Timer is enhanced with our further work and pre-trained on 260B time points. Checkpoint is now available: [HuggingFace] [Benchmark]. An example of zero-shot forecasting is provided here.

🚩 News (2024.10) We release numpy format of UTSD. An easier and more efficient dataloader can be found here.

🚩 News (2024.6) Pre-training dataset (UTSD) is available in HuggingFace. Dataloader is also contained.

🚩 News (2024.5) Accepted by ICML 2024, a camera-ready version of 31 pages.

🚩 News (2024.2) Releasing model checkpoints and code for fine-tuning.

Introduction

Time Series Transformer (Timer) is a Generative Pre-trained Transformer for general time series analysis. Please visit our Homepage for a detailed introduction.

Zero-Shot Forecasting

We provide the checkpoint to make predictions without training samples. See our HuggingFace Repo for the detialed information and usage.

A inference example (minimal dependencies required):

import torch
from transformers import AutoModelForCausalLM

# load pretrain model
model = AutoModelForCausalLM.from_pretrained('thuml/timer-base-84m', trust_remote_code=True)

# prepare input
batch_size, lookback_length = 1, 2880
seqs = torch.randn(batch_size, lookback_length)

# normalize the input to mitigate different scale
mean, std = seqs.mean(dim=-1, keepdim=True), seqs.std(dim=-1, keepdim=True) 
normed_seqs = (seqs - mean) / std

# generate forecast
prediction_length = 96
normed_output = model.generate(normed_seqs, max_new_tokens=prediction_length)
output = std * normed_output + mean # rescale the output to the original scale

print(output.shape)

Datasets

We collect Unified Time Series Datasets (UTSD), which encompass well-curated time series to facilitate the research on large time-series models. Our dataset is released in HuggingFace.

Usage

You can access and load UTSD in the style of TSLib based on the following steps:

# huggingface-cli login
# export HF_ENDPOINT=https://hf-mirror.com 

python ./scripts/UTSD/download_dataset.py

# dataloader
python ./scripts/UTSD/utsdataset.py

For Developers

For developers interest in large model adaptation, we provide fine-tuning code based on non-HuggingFace checkpoints, which is a smaller version of Timer developed in the TSLib style.

Note

We recommend using checkpoints on HuggingFace for model evaluation (e.g., zero-shot forecasting). However, it is not compatiable with the following fine-tuning code (but we are working on it :)

Supported Tasks

Forecasting: We provide all scripts for few-shot forecasting in this repo.

Imputation: We propose segment-level imputation, which is more challenging than point-level imputation.

Anomaly Detection: We provide new benchmarks of predictive anomaly detection on UCR Anomaly Archive.

We provide the README files illustrating each task under the folder ./scripts/.

Code for Fine-tuning

1. Use Python 3.10 and install necessary dependencies.

pip install -r requirements.txt

2. Put downstream datasets from Google Drive and Baidu Drive under the folder ./dataset/.

3. Put the checkpoint from Google Drive and Baidu Drive under the folder ./checkpoints/.

4. Train and evaluate the model. We provide the above tasks under the folder ./scripts/.

# forecasting
bash ./scripts/forecast/ECL.sh

# segement-level imputation
bash ./scripts/imputation/ECL.sh

# anomaly detection
bash ./scripts/anomaly_detection/UCR.sh

Train on Custom Dataset

To fine-tune on your time series dataset, you can try out the following steps:

1. The key is to reload the customized dataloader and load the pre-trained checkpoint (See ./scripts/ folder).
2. CIDatasetBenchmark/CIAutoRegressionDatasetBenchmark in the data_provider folder can train and evaluate models in direct / iterative multi-step mode.

Approach

Pre-training and Adaptation

To pre-train on heterogeneous time series, we propose single-series sequence (S3), reserving series variations into the unified 1D context. Further, we convert forecasting, imputation, and anomaly detection into a unified generative task.

Model Architecture

We evaluate various candidate backbones and eventually adopt the decoder-only Transformer, which provides notable generalization performance and length-flexibility that accommodate various time series.

Performance

Timer achieves state-of-the-art performance in zero-shot forecasting, general time series analysis, and present the pre-training benefit on few-shot scenarios.

Scalability

By scaling, Timer achieves notable performance improvement. Currently, we provide the base version containing 84M paramaters that is pre-trained on 260B time points, which supports a maximum context length of 2880.

Futher Improvement

We enhanced Timer by this paper with longer context and TimeAttention.

Citation

If you find this repo helpful, please cite our paper.

@inproceedings{liutimer,
  title={Timer: Generative Pre-trained Transformers Are Large Time Series Models},
  author={Liu, Yong and Zhang, Haoran and Li, Chenyu and Huang, Xiangdong and Wang, Jianmin and Long, Mingsheng},
  booktitle={Forty-first International Conference on Machine Learning}
}

@article{liu2024timer,
  title={Timer-XL: Long-Context Transformers for Unified Time Series Forecasting},
  author={Liu, Yong and Qin, Guo and Huang, Xiangdong and Wang, Jianmin and Long, Mingsheng},
  journal={arXiv preprint arXiv:2410.04803},
  year={2024}
}

Contributors

If you have any questions or want to use the code, feel free to contact:

• Yong Liu (liuyong21@mails.tsinghua.edu.cn)
• Guo Qin (qinguo24@mails.tsinghua.edu.cn)
• Haoran Zhang (zhang-hr24@mails.tsinghua.edu.cn)
• Chenyu Li (lichenyu20@mails.tsinghua.edu.cn)