Montevideo spatial-temporal data¶

The Montevideo bus dataset is the first real source with topology, metric position, scalar edge facts, and an ordered node signal. Tinymesh validates the host values first, then lowers their shared node and edge identity into tinygrad tensors.

Pinned source¶

Tinymesh reads PyG Temporal fe555bc with SHA-256 37d9c6286d474077b5c05173c1570c4da42c387013116daa8862c7a6cab86a75. The JSON payload is 3,131,884 bytes. The reader accepts at most 4 MiB and checks that limit before JSON parsing.

An explicit path supports offline runs and fixtures. The default path uses a 10-second standard-library request, reads at most 4 MiB, and verifies the exact checksum above. Python's standard library owns acquisition, JSON parsing, and source validation. No geo or dataframe package enters the runtime.

Source contract¶

ordered nodes
  bus_stop            stable source identity
  lon, lat            two finite numeric position values
  X.y                 ordered feature observations
  y                   ordered target observations

ordered links
  source, target      directed bus_stop identities
  weight              positive finite road distance

The parser resolves link identities to node rows while preserving both source orders. It rejects duplicate node identities, duplicate directed links, missing endpoints, non-finite geometry, non-positive distance, unequal time axes, and oversized input. It returns plain immutable Python values; it creates no tensor, normalization statistic, or dense topology.

The source calls its projected position fields lon and lat. Tinymesh preserves the numeric values without interpreting or transforming them at the source boundary.

Aligned tensors¶

from tinymesh.datasets import montevideo_bus

data = montevideo_bus(lags=4, device="CPU")
signal = data.signal

print(signal.x.shape, signal.y.shape)
# (740, 675, 4) (740, 675, 1)
print(data.position.shape, data.road_distance.shape)
# (675, 2) (690,)

MontevideoBus composes the existing StaticGraphTemporalSignal with one node-aligned position tensor and one COO-edge-aligned road-distance tensor. The record rejects shape, dtype, or device misalignment. It is dataset-specific; it does not establish a generic spatial container.

Graph                       675 nodes, 690 directed edges
signal.x                    [740, 675, 4] raw lagged inflow
signal.y                    [740, 675, 1] raw next-step inflow
position                    [675, 2] node order
road_distance               [690] original COO link order
coordinate frame            EPSG:32721
length unit                 m

The Uruguay open-data catalog identifies the stop positions as WGS 84 / UTM zone 21S. The Uruguay spatial-data recommendation maps that frame to EPSG:32721, whose position unit is metres. PyG Temporal calls the link weight road distance but does not label its unit. Its magnitude and agreement with UTM edge distance support metres; Tinymesh records m as this dataset interpretation and performs no projection.

Road distance remains data, not an aggregation coefficient: signal.edge_weight is None. The loader also retains raw passenger counts. Train-only normalization belongs to forecasting, after a forward split.

Full-source witness¶

uv run --locked python -m experiments.run montevideo_source

At the pinned revision:

{
  "nodes": 675,
  "edges": 690,
  "steps": 744,
  "duplicate_edges": 0,
  "self_loops": 0,
  "position_dimensions": 2,
  "minimum_road_distance": 23.8,
  "maximum_road_distance": 1991.3
}

This establishes source structure, not model quality or framework parity.

Tensor witness¶

uv run --locked python -m experiments.run montevideo_data DEV=CPU
uv run --locked python -m experiments.run montevideo_data DEV=METAL

Both devices report the aligned shapes above. For the first source edge, the straight-line coordinate distance is 171.59254 m and the observed road distance is 172.2 m. Keeping these as separate tensors makes the later forecast comparison explicit rather than silently treating either one as the graph weight.

This stage proves deterministic lowering and real sparse composition. It makes no forecasting or performance claim.

The cross-dataset network measurement adds the missing structural qualification. All 675 nodes form one weak component, but the source has no reciprocal edges, 675 singleton strong components, and only 17.0% directed pair reachability. Its mean reachable distance is 40.47 hops and its directed diameter is 114; the graph is a sparse directed route relation, not a mutually reachable proximity network.