tubeulator-models 0.1.5

Models of the TfL transit network

Tubeulator Models

Graph neural network models of the London Underground trained on TfL timetable data.

Two models share a GATv2 encoder over the station topology graph and differ in their decoder head:

• nexthop policy — given a current station and destination, predicts the optimal next station to visit. Greedy rollout achieves 1.09× Dijkstra ratio with 100% success across all 239,610 origin–destination pairs.
• nexthop value — predicts shortest travel time in minutes between any pair of stations. MAE of ~0.5 min vs. Floyd–Warshall ground truth.

Installation

pip install tubeulator-models[inference]

For the CLI:

pip install tubeulator-models[cli]

Usage

Python API

from tubeulator_models import TubeRouter

router = TubeRouter.from_pretrained("permutans/tube-nexthop-policy")
route = router.route("West Ham", "Shoreditch")
print(route)
# West Ham
#   → [district] Bromley-by-Bow (2.0m)
#   ...
# ✓ 8 hops · 2 lines · 1 transfer · 16.0 min

Route through waypoints:

route = router.route("Camden Town", "Canary Wharf", via=["King's Cross"])

The Route object has .steps, .total_minutes, .lines_used, .n_transfers, .success, and a .to_dict() method.

CLI

tm-infer policy --model permutans/tube-nexthop-policy -o "West Ham" -d Shoreditch
tm-infer policy --model permutans/tube-nexthop-policy -o "Camden Town" -d "Canary Wharf" -v "King's Cross"
tm-infer value  --model permutans/tube-nexthop-value  -o "West Ham" -d Shoreditch

Training from source

Clone the repo and use the justfile recipes:

just all        # fetch data + build graph + train

Or step by step:

just fetch      # pull timetables from TfL API → GTFS zip
just graph      # GTFS → GeoParquet → PyG graph objects
just train nexthop full
just train nexthop full --value-primary --batch-size 1024 --epochs 200

All hyperparameters live in defaults.toml. CLI flags override the TOML but never replace it:

tm-train --model nexthop --profile full --lr 5e-5

Profiles control the training regime:

Profile	Epochs	Batch size	LR	Notes
dev	20	512	5e-4	Fast iteration
full	200	256	1e-4	Production, d_model=512, 16-layer encoder

Data pipeline

Step	CLI	Recipe	Output
Fetch timetables	tm-build-gtfs	just fetch	data/tfl_station_data_gtfs.zip
Build graph	tm-gtfs2pyg	just graph	data/graph/, data/pyg/
Plot network	tm-plot	just plot	data/graph/network.png
Train	tm-train	just train	checkpoints/
Export	tm-export	just upload	models/ → HuggingFace

Staged graph conversion is also available:

uv run --group prep    tm-gtfs2graph   # GTFS → GeoParquet
uv run --group pyg     tm-graph2pyg    # GeoParquet → PyG .pt

Configuration

defaults.toml is the single source of truth for all tuneable values. The merge order is:

[base] → [model.<type>] → [profiles.<name>] → [profiles.<name>.model.<type>] → CLI overrides

Development

just lint       # ruff check
just clean      # remove data/ and checkpoints/

Links

• Policy model on HuggingFace
• Value model on HuggingFace
• tubeulator-models on PyPI