SANet

Repository Layout

SANet/
├── configs/     # Training and inference defaults
├── data/        # Dataset format docs and local experiment data
├── models/      # Model definitions
├── scripts/     # Train/inference entrypoints
└── util/        # Data loading, normalization, and optimization helpers

Requirements

• Python 3.9+
• CUDA is optional but recommended for training

Install PyTorch first using the command recommended for your platform on the official PyTorch site, then install the remaining dependencies:

pip install -r requirements.txt

Quick Start

Train from the repository root:

python scripts/train.py

Example:

python scripts/train.py \
  --epochs 200 \
  --log_dir logs/exp_log/exp_my_run \
  --bandwidth_B 64 \
  --bandwidth_loss_weight 0.5 \
  --use_bandwidth_adaptive_compression 1

Run inference with:

python scripts/inference.py

Main defaults are defined in configs/config.py.

Data Format

Each dataset directory should contain:

• CSI.npy: shape (T, 2) for I/Q channels
• traffic.npy: shape (T,) or (T, 1)
• user_intent.npy: shape (T,) or (T, 1), optional
• scaler.npz: optional, auto-generated if missing

Example layout:

data/
  train_band_n1/
    CSI.npy
    traffic.npy
    user_intent.npy
    scaler.npz

More details are in data/README.md.

License

This project is released under the MIT License. See LICENSE.

Citation

If you use SANet in your research, please cite the following paper:

@article{xiao2026sanet,
  title   = {{SANet}: A Semantic-aware Agentic {AI} Networking Framework for Cross-layer Optimization in {6G}},
  author  = {Xiao, Yong and Li, Xubo and Zhou, Haoran and Li, Yingyu and Gao, Yayu and Shi, Guangming and Zhang, Ping and Krunz, Marwan},
  journal = {IEEE Transactions on Mobile Computing},
  year    = {2026},
  note    = {Early Access}
}