LSCodec

A Time series long-short term codec for compression and representation

This repository supplements our paper "LSCodec: A Time series long-short term codec for compression and representation".

TODO

• Implement of LSCodec model
• Training scripts and configs
• Downstream task scripts
• Pre-trained checkpoints
• how to cite our paper

Framework

Environment Setup

Create and activate a virtual environment to work in, e.g. using env:

python -m venv your_env_lscodec 
your_env_lscodec/Scripts/activate  

Install CUDA and PyTorch 1.13. For CUDA 11.6, this would look like:

pip install torch==1.13.1+cu116 torchvision==0.14.1+cu116 torchaudio==0.13.1 --extra-index-url https://download.pytorch.org/whl/cu116

Install the remaining requirements with pip:

pip install -r requirements.txt

The proposed framework is high relied on vector_quantize, it is an excellent VQ library.

Datasets

LSCodec supports Multi-Channel regular Time Series Sequences. For data preparing, we follow DTAAN scripts for data normalization and train/valid splitting.

python ./preprocess.py <datasets>  

where is space separated list of datasets like 'synthetic', 'SMD', 'SWaT', 'SMAP', 'MSL', 'WADI', 'MSDS', 'UCR', 'MBA', 'NAB' ... The data should be placed in './data' folder. And the preprocessed data will be generated in './processed' folder. Considering that datasets is not allowed for distribution, we recommand to access dataset by its official website.

Training

For CPU training

python ./train.v1.py --task lscodec --config ./config/config_lscodec_xxx.toml --force_cpu  

For single GPU training

python ./train.v1.py --task lscodec --config ./config/config_lscodec_xxx.toml  

For Multi-GPU training, we build our training code upon Accelerate for multi-gpu training, you need to prepare your gpu-env configurations.

CUDA_VISIBLE_DEVICES=xxxx accelerate launch --config_file ./your_accelerate_config  ./train.v1.py --task lscodec --config ./config/config_lscodec_xxx.toml    

License

BSD-3-Clause.
Copyright (c) 2024, Haiwei Zuo.
All rights reserved. See License file for more details.