86Cao/china-city-graphml

VisNavMap City GraphML Snapshots

This dataset contains the city-level navigation graph cache derived from the city-level OpenStreetMap PBF files released in the Hugging Face dataset 86Cao/china-city-osm-pbf. It is intended to make VisNavMap training, rollout, and evaluation reproducible without rebuilding a graph from a PBF at runtime.

Data provenance

The data-generation chain is:

Hugging Face: 86Cao/china-city-osm-pbf
        │
        │ download city-level <adcode>_<official_name>.osm.pbf
        ▼
VisNavMap graph reader: read_pbf()
        │
        ▼
city-level raw directed walking graph: <adcode>_<official_name>.graphml

Each GraphML file is generated from the PBF with the same read_pbf() and graph_hash() implementation used by the project. The city filename is kept unchanged across the source PBF, GraphML, and manifest so that the matching source file can be recovered unambiguously. The original PBF files are not duplicated in this dataset.

Contents

data/
  graphml/       # <adcode>_<official_name>.graphml
  manifests/     # one manifest per city plus index.json

The GraphML files are the raw directed walking MultiDiGraph cache. VisNavMap still applies its runtime graph policy and constructs the task-specific decision graph from this cache; GraphML is not a rendered map or a serialized episode.

The city namespace is shared with the source PBF dataset:

110000_北京市.graphml
110000_北京市.json

Each manifest records the source PBF name, graph snapshot hash, and node/edge counts. data/manifests/index.json is the dataset-level catalog and should be used to enumerate available cities.

Loading

import networkx as nx

graph = nx.read_graphml(
    "data/graphml/110000_北京市.graphml",
    node_type=str,
)

For VisNavMap runtime loading, pass the GraphML root through the project’s graph cache configuration; the loader falls back to the source PBF only when the matching GraphML file is absent.

Reproduction

The conversion script is included in this dataset under scripts/. Run it from the VisNavMap-sandbox repository, where the project source code and its graph-reading policy are available:

python data/china-city-graphml/scripts/build_graphml_from_pbf.py \
  --pbf-root data/china-city-osm-pbf/data/pbf \
  --output-root data/china-city-graphml

The input directory is searched recursively for *.osm.pbf. For each city, the script writes one GraphML file to data/graphml/, one manifest to data/manifests/, and rebuilds data/manifests/index.json. Existing GraphML files are skipped by default; add --overwrite to regenerate them.

The script depends on the VisNavMap runtime source under src/ and on NetworkX. It uses the same read_pbf() and graph_hash() implementation as the training and inference code, so this is a reproducibility script rather than an independent OSM conversion policy.

Data and licensing

These files are derived from OpenStreetMap data. Preserve OpenStreetMap attribution and comply with the Open Database License (ODbL). The source PBF files are not duplicated in this artifact dataset; obtain them from 86Cao/china-city-osm-pbf.