Unofficial implementation of the paper:

B. Park, H. Do and N. Lee, "Multi-Rate Variable-Length CSI Compression for FDD Massive MIMO," ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP), Seoul, Korea, Republic of, 2024, pp. 7715-7719.

This implementation is built on top of the CompressAI library.

1. Install the requirements.

2. Download the dataset:

   curl -L -o COST2100_dataset.zip "https://www.dropbox.com/scl/fo/tqhriijik2p76j7kfp9jl/h?rlkey=4r1zvjpv4lh5h4fpt7lbpus8c&e=2&st=pmf7duk6&dl=1"
   

   unzip COST2100_dataset.zip -d COST2100_dataset
   

   rm -f COST2100_dataset.zip
   

3. Edit the dataset path in cost_loader.py line 39:

   general_path = '/MY_DATASETS/COST2100_dataset/'
   

python3 main.py -train --name test1

python3 test_bit_budgets.py --run lambda-5e-4_div11.8

The results closely match those presented in the paper.

If you find this repo useful please cite:

@software{Rizzello_csi-feedback-vbr_2026,
    author = {Rizzello, Valentina},
    doi = {10.5281/zenodo.19597095},
    month = apr,
    title = {{csi-feedback-vbr}},
    url = {https://github.com/vrizz/csi-feedback-vbr},
    version = {1.0.0},
    year = {2026}
}

@INPROCEEDINGS{park2024-multi-rate,
    author={Park, Bumsu and Do, Heedong and Lee, Namyoon},
    booktitle={ICASSP 2024 - 2024 IEEE International Conference on Acoustics, Speech and Signal Processing (ICASSP)}, 
    title={Multi-Rate Variable-Length CSI Compression for FDD Massive MIMO}, 
    year={2024},
    volume={},
    number={},
    pages={7715-7719},
}