💡 LITE: Modeling Environmental Ecosystems with Multimodal Large Language Models

📝 Requirements

1. Clone this repository and navigate to LITE folder.

git clone https://github.com/hrlics/LITE.git
cd LITE

2. Install packages.

conda create -n LITE python=3.10 -y
conda activate LITE
pip install -r requirements.txt

📦 Preparation

1. Save the environmental data in the format of Hugging Face Datasets. Suppose it is tabular data, with N features and 1 target variable.
2. Construct temporal trend images.

python draw_temporal_image.py \
 --feature_path path/to/your/raw/features \
 --target_path path/to/your/raw/targets \
 --dataset_path path/to/the/huggingface_datasets \
 --dataset_name name/of/the_parent_dir/of/huggingface_datasets  \
 --len_look_back look_back_window_size (default:30)

🚀 Training

1. The training code is at ./train.py.
2. The relevant script can be found at scripts/train_{dataset_name}.sh.

sh ./scripts/train_{dataset_name}.sh

📚 Citation

If you find our work useful, please consider citing:

@article{li2024lite,
  title={LITE: Modeling Environmental Ecosystems with Multimodal Large Language Models},
  author={Li, Haoran and Liu, Junqi and Wang, Zexian and Luo, Shiyuan and Jia, Xiaowei and Yao, Huaxiu},
  journal={arXiv preprint arXiv:2404.01165},
  year={2024}
}