VEWOXIC / TGTSF

Official implementation of TGTSF in "Beyond Trend and Periodicity: Guiding Time Series Forecasting with Textual Cues"

TGTSF: Text Guided Time Series Forecasting

The main model is TGTSF_torch. Other versions are deprecated.

We will update the paper on arxiv from time to time. Keep updated here: https://arxiv.org/abs/2405.13522

Dataset Preparation

Toy Dataset

• We have upload the toy dataset together with its generation scripts. You can use it to create your own dataset. The pre-embedding scripts are also included. Please do pre-embedding before training.

Weather-Captioned Dataset

• Weather-captioned dataset is uploaded, including the time series of 10 years, all pre-embeddings files for the captions and hashtables for indexing the embeddings.
• We put the pre-embeddings as tarball and storage them on github with git-lfs. You may need to:
  1. Install git-lfs with sudo apt-get install git-lfs or brew install git-lfs
  2. Run git lfs install in the repository
  3. Run git lfs pull to download the pre-embedding files.
  4. Unzip the tarball with tar -xvf embeddings.tar
  5. We break the embeddings for weather-large into several parts due to the 2GB file size limit of GITHUB. You need to merge them with cat openai_caption_emb_large_part_*.tar > openai_caption_emb_large.tar and then untar it.
• We also upload all the scripts to generate such a dataset, including rawdata, captioning, pre-embedding, and indexing as a seperate repository. You can find it here: Weather Captioned Dataset

⚠ If you have trouble in downloading the pre-embedding files with git-lfs, we also provide google drive links for the pre-embedding files. Click Here

You can use gdown to download the files from google drive.

Run the Model

Run scripts in the ./scripts folder.

Use visualize.ipynb to visualize the results. We may upload the checkpoint we trained later.