WiFo-CF: Wireless Foundation Model for CSI Feedback

Liu, Xuanyu, et al. "WiFo-CF: Wireless Foundation Model for CSI Feedback." arXiv preprint arXiv:2508.04068 (2025). [paper]

Overview

WiFo-CF is a wireless foundation model designed for CSI feedback under heterogeneous system configurations.
Unlike conventional task-specific models that are usually trained for a fixed antenna setting, feedback rate, or channel distribution, WiFo-CF aims to provide a unified framework that can generalize across diverse CSI dimensions, user configurations, feedback bit-widths, and deployment scenarios.

To achieve this, WiFo-CF combines:

• a multi-user, multi-rate self-supervised pre-training strategy,
• a flexible multi-user scalable autoencoder (MUAE) architecture,
• and a Mixture of Shared and Routed Experts (S-R MoE) design for modeling both shared channel correlations and dataset-specific characteristics.

Pre-training Dataset

We build LH-CDF, a large-scale heterogeneous CSI feedback dataset for foundation-model pre-training.

LH-CDF includes:

• Statistical-modeling-based data generated by QuaDriGa,
• Ray-tracing-based data from DeepMIMO and SynthSoM,
• Real-world measured data from Argos and Dichasus.

The dataset covers diverse:

• carrier frequencies,
• antenna configurations,
• user numbers,
• propagation scenarios,
• and channel distributions.

This heterogeneous design enables WiFo-CF to learn more generalizable channel representations for CSI compression and reconstruction.

Dependencies and Installation

• Python 3.8 (Recommend to use Anaconda)
• Pytorch 2.0.0
• NVIDIA GPU + CUDA
• Python packages: pip install -r requirements.txt

Evaluation

After installing the dependencies, download the LH-CDF test dataset and pretrained checkpoint:

• Test dataset: LH-CDF
• Pretrained checkpoint: wifo_cf_base.pth

Place the LH-CDF dataset under ./dataset/, or use your own channel data with the same data structure. WiFo-CF supports channel compression and reconstruction with arbitrary numbers of antennas and subcarriers.

Put wifo_cf_base.pth in the project root directory, then run:

bash test.sh
python get_result.py

test.sh runs the evaluation, and get_result.py summarizes the final results.

Current Status

At this stage, we have released:

• Model architecture code
• Testing / inference code
• Pretrained model weights
• Pretraining dataset

TODO

The following components will be released progressively:

• Fine-tuning code for downstream tasks
• Training scripts
• More comprehensive documentation

Notes

This repository is actively maintained.
We will progressively release additional resources to facilitate reproducibility and follow-up research, including pretrained checkpoints, pretraining datasets, and downstream fine-tuning pipelines.