Channel Charting for Streaming CSI

This is the code for the results in the paper "Channel Charting for Streaming CSI Data", S. Taner, M. Guillaud, O. Tirkkonen, and C. Studer (c) 2025 Sueda Taner

email: taners@ethz.ch

Important Information

If you are using this code (or parts of it) for a publication, then you must cite the following paper:

S. Taner, M. Guillaud, O. Tirkkonen and C. Studer, "Channel Charting for Streaming CSI Data," Asilomar Conference on Signals, Systems, and Computers, 2023, pp. 1648-1653.

How to use this code...

Step 1: Download and save the data

• From dichasus-cf0x Dataset: Distributed Antenna Setup in Industrial Environment, Day 1, download the dichasus-cf02 and dichasus-cf03 files (which should have the .tfrecord extension) and their offset estimates (which should be reftx-offsets-dichasus-cf02.json and reftx-offsets-dichasus-cf03.json) into a folder called ```data_raw``.
• Run preprocess_dichasus.py. This will store .np versions of the CSI, timestamps, and ground-truth positions extracted from the .tfrecord files in a folder called data.

Step 2: Core memory curation for streaming CSI and channel charting

• Set your training and core memory curation parameters as explained on top of main.py and run for the results in the paper. This code does the following:
  • We use the cf02 dataset to simulate the streaming CSI.
  • We store a subset of this dataset in the core memory.
  • We train a neural network for channel charting using the CSI features in the core memory.
  • We test the channel charting neural network on the cf03 dataset.

Version history

Version 0.1: taners@ethz.ch - initial version for GitHub release.

Acknowledgments

This project makes use of the following external data and code:

• dichasus-cf0x Dataset: Distributed Antenna Setup in Industrial Environment, Day 1, accessed on 1/3/2025: Our code uses the cf02 and cf03 datasets.
• Dissimilarity Metric-Based Channel Charting by F. Euchner, accessed on 1/3/2025: We use this code for (i) pre-processing the data from tfrecords and json files, and (ii) computing the angle delay profile (ADP)-based distance metric and then the geodesic distances for training the channel charting network.