COST2100 CSI Feedback

This repository mirrors the eight .mat data files used to train and evaluate CsiCoGen, the implementation of Variable-Length Finite-Rate CSI Feedback With Generative Priors, accepted by IEEE Wireless Communications Letters (WCL 2026).

The files retain the original array values and file contents. Each SHA-256 in manifest.json was checked against the copy used by the authors' server experiments. No additional splitting, normalization, or resampling is applied to this mirror.

Files

File	Scene	Split	Samples	Array key and shape
DATA_Htrainin.mat	Indoor	Train	100,000	HT: 100000 × 2048
DATA_Hvalin.mat	Indoor	Validation	30,000	HT: 30000 × 2048
DATA_Htestin.mat	Indoor	Test	20,000	HT: 20000 × 2048
DATA_HtestFin_all.mat	Indoor	Test frequency-domain reference	20,000	HF_all: 20000 × 4000 (complex)
DATA_Htrainout.mat	Outdoor	Train	100,000	HT: 100000 × 2048
DATA_Hvalout.mat	Outdoor	Validation	30,000	HT: 30000 × 2048
DATA_Htestout.mat	Outdoor	Test	20,000	HT: 20000 × 2048
DATA_HtestFout_all.mat	Outdoor	Test frequency-domain reference	20,000	HF_all: 20000 × 4000 (complex)

HF_all contains paired references for the same test samples, rather than additional samples. The complete download is approximately 6.82 GB (decimal).

Download

The unified code prepares the original dataset automatically. The resources are currently private and use your configured credentials.

python main.py download data
python main.py download data --scene indoor --split test

Array conventions

import scipy.io as sio
import numpy as np

x = sio.loadmat("data/COST2100/DATA_Htestin.mat")["HT"]
x = x.reshape(-1, 2, 32, 32).astype(np.float32)
h = sio.loadmat("data/COST2100/DATA_HtestFin_all.mat")["HF_all"]
h = h.reshape(-1, 32, 125)

HT stores the real and imaginary channels of the truncated angular-delay representation in the standard CSI-feedback dataset's normalized scale. CsiCoGen further standardizes these inputs using the training-set statistics bundled with each checkpoint. For the CRNet-aligned evaluator, subtract 0.5, pad the delay dimension from 32 to 257, apply the FFT along that dimension, and retain the first 125 frequency bins to match HF_all. These released evaluation files contain 125 reference bins; they do not contain all 1024 subcarriers described by the paper's system model.

Attribution

The channel data originate from the COST2100 channel model and the CSI-feedback data distribution used by CsiNet and subsequent work. This repository is a data mirror and does not claim authorship of the underlying channel model or dataset.

• COST2100 channel model
• Python CsiNet implementation and data instructions

Please cite the COST2100 channel-model work and the original CSI-feedback dataset work when using these data. Refer to the original distributions for their applicable terms; this mirror does not assign a new license to third-party data.