This is a PyTorch compatible implementation of the transform coding scheme used in
@inproceedings{valtonen-ornhag-etal-icassp-2023,
author = {Marcus {Valtonen~{\"O}rnhag} and
Stefan Adalbj{\"o}rnsson and
P{\"u}ren G{\"u}ler and
Mojtaba Mahdavi},
title = {A Critical Look at Recent Trends in Compression of Channel State Information},
booktitle = {To be presented at the {IEEE} International Conference on Acoustics, Speech and Signal Processing,
{ICASSP} 2023},
year = {2023},
}
Please cite the paper if you use the code in academic publications.
Please note, that the core message from our paper is not to propose a state-of-the art method, but to have an open discussion on the datasets and metrics used by the community. In particular, we propose to:
- Utilize new datasets, e.g. DeepMIMO.
- Always quantize the data. Do not measure compression ratio in terms of length of the vector!
- Be mindful about using the NMSE metric for intermediate system-level simulations.
- Bring research from academia and industry closer, not further apart.
Our work is using the COST2100 dataset by Wen et al., which can be downloaded from one of the following locations:
Download the dataset and put it in a folder named COST2100 in the root folder.
You can use a virtual environment to run the code:
python3.9 -m venv venv
source venv/bin/activate
pip install -r requirements.txt
We used Python 3.9.14 on UNIX, but the scripts should be compatible for Python versions >= 3.7.
You can now execute the main script
python main.py --data-dir COST2100 --scenario in -b 200 -j 0 --cr 4 --cpu --evaluate
To recreate the results from the paper run with --datadir {in,out} and --cr {4,16,64}.
| Scenario | Compression Ratio | NMSE | rho |
|---|---|---|---|
| indoor | 1/4 | -25.75* | 0.99 |
| indoor | 1/16 | -15.39 | 0.98 |
| indoor | 1/64 | -8.73 | 0.94 |
| outdoor | 1/4 | -17.67 | 0.97 |
| outdoor | 1/16 | -9.81 | 0.93 |
| outdoor | 1/64 | -4.61 | 0.77 |
*In the paper we reported 25.78, which is a typo.
Note: The code is not optimized for performance. By lowering the bit precision and increasing the number of coefficients the performance could be further increased. This, however, is not the main point of this paper.
Use the --bit-depth flag, e.g.
python main.py [...] --cr 16 --bit-depth 4
which would result in a comparison where the corresponding deep-learning based method is sending a vector reduced by 1/16 of its original length together and where each real and imaginary bit is allocated 4 bits.
In the quantization experiments, we used the open-source repositories from
- Lu et al. (2020), CRNet repo
- Cui et al. (2022), TransNet repo
Our code conform with the above-mentioned repositories, in order to utilize the same evaluation script.
To test the quantization impact of these methods one needs to modify the encoder output and decoder
input by passing them through the non-linear quantization methods available in models/quantization.