This repository contains the PyTorch implementation of the Multimodal Adaptive Gaussian Network (MAGNeT), designed for moving target selection.

To get started, ensure you have the following installed:

• Python: Version 3.10 or newer
• PyTorch: Version 2.4.0 or newer
• Scikit-learn: Version 1.5.2 or newer
• Pandas: Version 2.2.2 or newer

MAGNeT is computationally efficient and can run on CPU/GPU with at least 2GB of memory. Our experiments were conducted on an RTX 3090 (24GB GPU), utilizing approximately 1.8GB of memory for a batch size of 32. This setup achieved an impressive processing speed of $2.84 \times 10^{-4}$ seconds per sample during training and $1.43 \times 10^{-4}$ seconds during testing.

To use the provided datasets, follow these steps:

1. Download the dataset package:

   • You can download the compressed dataset package MTS_dataset.zip.

2. Extract and place the folders:

   • After downloading, extract the MTS_dataset.zip file.
   • Move the extracted dataset_2d and dataset_3d folders to the same directory as your fusion_file_2d.py or fusion_file_3d file.

3. Using your own datasets (optional):

   • If you prefer to prepare your own datasets, you can refer to the structure and format of the provided dataset_2d and dataset_3d folders as a reference.

Once the datasets are properly placed, you should be able to run the code without additional configuration.

We use 5 randomly generated seeds for dataset splitting and weight initialization. The random numbers will be generated in random_number.npy file automatically.

To train and test MAGNeT, execute the following command:

For 2D moving target selection:

python fusion_file_2d.py --use_env --epochs 50 --num_shot [NUM_FEW_SHOT] --result_path [PATH_TO_SAVE]

For 3D moving target selection:

python fusion_file_3d.py --use_env --epochs 50 --num_shot [NUM_FEW_SHOT] --result_path [PATH_TO_SAVE]

Detailed parameter settings can be found in parse.py.

@article{li2025magnet,
  title={MAGNeT: Multimodal Adaptive Gaussian Networks for Intent Inference in Moving Target Selection across Complex Scenarios},
  author={Li, Xiangxian and Zheng, Yawen and Zhang, Baiqiao and Ma, Yijia and XianhuiCao, XianhuiCao and Liu, Juan and Bian, Yulong and Huang, Jin and Yang, Chenglei},
  journal={arXiv preprint arXiv:2508.12992},
  year={2025}
}