CityDeploy-Data

Urban scenes. Coordinated deployments. Physics-grounded supervision.

Code  ·  Explore  ·  Load data  ·  Download layers  ·  Schema  ·  License

CityDeploy-Bench manuscript figure. Geometry: © OpenStreetMap contributors.

Where should a set of transmitters be placed in a city? CityDeploy-Data pairs urban geometry with feasible transmitter sets and radio-simulation labels for scene-conditioned utility learning, spatial-set planning and geographic evaluation.

Complete archived corpus available. All deployment records, cached scene tensors, 67 portable scene packages and 36,599 archived Radio Maps are uploaded and verified. See release verification. Download only the layers you need. No pretrained weights are included.

At a glance

Cities	Primary scenes	Unique deployments	Spatial scales	Archived Radio Maps
6	60	54,193	128 / 256 / 512 m	36,599

London · New York · Paris · Shanghai · Sydney · Tokyo

Deployment region	Scenes	TXs per deployment	Records
128 × 128 m	12	1, 2	5,811
256 × 256 m	18	2, 3, 4	17,782
512 × 512 m	30	5, 6, 7, 8, 9	30,600

One row is one deployment, not one TX. The primary corpus has 54,900 separate family memberships. The adaptation extension adds 3,590 unique deployments with 3,600 memberships and no archived maps. All 67 local scene packages are included as portable assets: 60 primary scenes, two additional adaptation scenes, and five supplementary scene packages. Do not count supplementary geometry as additional primary benchmark training data.

Start small

Read the complete scalar tables without downloading meshes or Radio Maps:

from datasets import load_dataset

train = load_dataset("Chenyang-Yuan/CityDeploy-Data", "deployments", split="train")
sample = train[0]
print(sample["scene_id"], sample["num_tx"])
print(sample["tx_xyz_m"])                 # scene-centered meters
print(sample["tx_xy_norm01"])             # XY in [0, 1]
print(sample["joint_4metric_coverage"])   # fraction, not percent

Use streaming=True for row iteration, or select scale_128, scale_256, scale_512, scene_catalog, family_memberships, adaptation, or preview. The 18-row preview remains a tutorial, not a representative statistical subset.

Train the four utility models

Download the loader and install its lightweight dependencies:

from huggingface_hub import snapshot_download
snapshot_download("Chenyang-Yuan/CityDeploy-Data", repo_type="dataset",
                  allow_patterns=["examples/*"], local_dir="citydeploy_data")

python -m pip install -r citydeploy_data/examples/requirements.txt
python citydeploy_data/examples/download_dataset.py --destination citydeploy_data

After installing CityDeploy-Bench:

citydeploy train --model hpem --dataset-root citydeploy_data/core/urban_multicity --scene-root citydeploy_data/scenes --device cuda

Select regressor, in, or nri for the other models. Cached 11-channel scene tensors are already included in the core; training from these labels does not require the large spatial archives or a ray tracer. Select training budgets and record the downloaded revision. See download layers.

What one sample contains

Group	Content
Identity	Sample, scene, geographic group, contract, family and split
TX set	TX count, candidate indices, XYZ meters and normalized XY
Coverage	Path Loss, SS-RSRP, SINR, throughput and joint coverage; counts and denominator
Statistics	Metric means and percentiles
Provenance	Sampling strategy, simulation seed and numerical RF contract
Optional spatial outputs	Map path; shard/member/checksum in the map index

36,599 / 54,193 primary deployments have archived Radio Maps. The remaining 17,594 still have scalar labels. Missing maps do not mean zero coverage. The release includes every archived map; it does not fabricate maps that were never saved. Map shards contain the original numerical payloads, including per-TX path gain where available, without reducing precision or rerunning propagation.

Evaluation protocol

Primary split	Scenes	Deployments
Train	42	38,162
Validation	9	8,019
Test	9	8,012

Preserve geographic-group assignments and family relationships. Random row-level resplitting can leak correlated deployments. The adaptation extension is separate: paris_11 (128 m, 2 TX), new_york_12 (256 m, 4 TX), tokyo_01 (512 m, 8 TX). Its historical split fields are not within-target few-shot partitions. Training on these targets changes the evaluation regime and must be reported explicitly.

Per-cell metric	Covered when
Path Loss	≤ 106 dB
SS-RSRP proxy	≥ −110 dBm
SINR	≥ 5 dB
Effective Throughput	≥ 10 Mbps

Joint coverage is an intersection: the fraction of road evaluation cells meeting all four criteria. Other active TXs contribute interference relative to the strongest-received-power serving TX. A learned utility is a surrogate, not physical verification of a new deployment.

The reference configuration uses 3.65 GHz, 20 MHz, isotropic vertically polarized antennas, 24 dBm per TX, TX/RX heights of 10/1.5 m, 1 m map cells and 5 million tracing samples per TX. LOS, specular/diffuse reflection, refraction and diffraction are enabled with maximum depth 5. See RF parameters.

Limits and responsible use

• Simulation labels depend on geometry, material priors and propagation settings; they are not measured network performance or installation permission.
• Road coverage is not whole-rectangle or indoor coverage.
• Coverage fields are fractions in [0, 1]; plots may multiply by 100.
• The cached evaluation_mask is a context channel, not necessarily the label denominator. Use each map's road_evaluation_mask to recompute coverage.
• Subsets can share a master simulation. Do not treat all family memberships as independent deployments or silently mix adaptation rows into primary training.
• Public IDs and linked contracts were migrated together. Original coordinates, numeric labels, model-input tensors and split assignments remain unchanged.

License and attribution

Original simulation outputs: noncommercial scientific research only. Geographic layers derived from OpenStreetMap retain ODbL-1.0 and attribution to © OpenStreetMap contributors; their existing rights are not restricted by the simulation-output terms. Loader code is MIT. See the complete, component-specific license. No satellite tiles, model checkpoints or experiment logs are bundled. Citation details will be added with the archival paper record.

Data specification · Download layout · Corpus statistics · Checksums