BerkIGuler/ChaSEDatasets

ChaSE evaluation data

This data set holds the ray-traced channels on which ChaSE (Channel Semantics Evaluation) evaluates wireless channel foundation models. CSIGen generated the channels with Sionna RT 2.2.0. The data set holds the saved paths of every channel, not the channel frequency responses. ChaSE rebuilds the responses from the paths, bit for bit.

Contents

Path	Contents
runs/<scene>/<site>/channel_tx_<i>.npz	The paths and the metadata of every channel of one sector (CSIGen format, without h)
runs/<scene>/<site>/generation.yaml	The CSIGen version, the software versions, a summary of the run and the checksum of each file
runs/<scene>/<site>/metadata.yaml	The settings of the run
partitions/<cell>/partition.npz	The part of every link (train, val, test, excluded, guard) and whether it goes to the serving site
partitions/<cell>/manifest.yaml	The split, the counts per site, the mean power of the pretraining channels and the checksums
quickstart/runs/<site>/channel_tx_<i>.npz	A small sample for the quick start of ChaSE (see below)

Scene	Sites	Evaluation cell
urban	Boston, Chicago, New York, San Francisco	partitions/urban
suburban	Irvine, Palo Alto	partitions/suburban
rural	Agua Dulce, Littlerock	partitions/rural

Each site has base stations with 6 sectors. Each sector has a uniform planar array of 4 rows and 8 columns, that is 8 elements horizontally and 4 vertically. Each user has one isotropic antenna. The channels have 14 OFDM symbols and 32 subcarriers at 30 kHz around 3.5 GHz. One base station per city or town is the test cell. No channel of the test cell is in the train or val parts. The validation links are 50 m tiles with a guard of 10 m.

Build the ChaSE sets

Install ChaSE, download this data set, and build the sets of a cell. The build needs only numpy and runs on the CPU, at about 10 to 20 seconds per file.

chase-build-data runs/urban partitions/urban data/urban

This writes train.npy, val.npy and test.npy (the channels, [n, OFDM symbol, user antenna, base-station antenna, subcarrier] complex64) and <set>_labels.npz (the ground truth from the paths) to data/urban. Set data_dir: data/urban in the ChaSE config. The sizes of the built sets are about 17 GB (urban), 5.4 GB (suburban) and 8.2 GB (rural).

ChaSE takes the settings of the paper by itself. Its default min_path_gain_db is -150, and it reads mean_power from the manifest.yaml of the build. For reference, these are the values of each cell.

Cell	mean_power	min_path_gain_db
urban	3.858271428867973e-11	-150
suburban	1.121667476669e-09	-150
rural	1.4140291090871107e-09	-150

Quick-start sample

The folder quickstart/ holds a small sample for the quick-start notebook of ChaSE. It has the two rural towns, Agua Dulce and Littlerock, with 10% of their users (642 and 651 users, chosen with seed 0). The files keep the CSIGen format, with the rows of the kept users in every array, and the channels rebuilt from them equal the same rows of the full runs, bit for bit. The sample is 25 MB, and its built sets take about 0.9 GB. It has no partition, because the notebook makes one with chase-make-partition.

hf download BerkIGuler/ChaSEDatasets --repo-type dataset --include "quickstart/*" --local-dir chase_data

Sources and attribution

The building and terrain geometry of the scenes comes from OpenStreetMap (© OpenStreetMap contributors, available under the Open Database License) and from lidar data, processed with Geo2SigMap.

Citation

The citation follows when the paper is on arXiv.