Skip to content

METR-LA sensor data

METR-LA is Tinymesh's first real fixed sensor network with explicit timestamps, missing observations, and asymmetric road-network affinity. This stage only establishes the data boundary; it makes no forecast or topology-value claim.

Pinned sources

Traffic readings come from the CC BY 4.0 METR-LA CSV on Zenodo:

METR-LA.csv
72,467,662 bytes
SHA-256 8d67a35472db1719d7d4be851f2bf64cb21d9c52577c8a6b4b873d43205af381

Sensor order and road distances come from DCRNN revision 602afd9d:

graph_sensor_ids.txt
1,448 bytes
SHA-256 3ba026caa2e6263ab0ea54b0fa1b125dbfa7216544cd05313b555e826292b990

distances_la_2012.csv
6,393,348 bytes
SHA-256 a576a2a3e28dbb959be6da22688e24dd1b246b81264595e129147c256cd53de5

The default loader uses tinygrad's cache and verifies every remote checksum and exact byte count. An explicit directory supports offline use and fixtures; it must contain those three canonical names, remains size-bounded, and is validated structurally rather than forced to match the remote checksums.

The CSV identifies values as traffic speed but does not label a unit, so Tinymesh does not invent one. Its timestamps have no zone offset and remain naive source datetimes.

Source identity and missingness

The traffic header matches all 207 DCRNN sensor IDs in exact order. The payload contains 34,272 uninterrupted five-minute rows:

2012-03-01 00:00:00
        ...
2012-06-27 23:55:00

The DCRNN paper describes a broader March 1 through June 30 collection period, while this published CSV ends on June 27. Tinymesh preserves the payload rather than filling the difference.

DCRNN's pinned training code uses zero as the missing-value sentinel. The CSV has:

7,094,304 total values
6,519,002 observed values
  575,302 missing sentinels

The nonzero count exactly matches the paper's reported number of observed METR-LA points. METRLA.observed therefore derives speed != 0 on demand; the loader neither imputes zero nor stores a second source of truth.

Directed affinity

Tinymesh reproduces DCRNN's pinned graph recipe directly from selected road-distance rows:

sigma = population standard deviation of selected distances
w(i,j) = exp(-(distance(i,j) / sigma)^2)
keep w(i,j) >= 0.1

Direction is retained. The result is 1,722 COO-ordered edges, including 207 self-edges and 1,111 edges whose reverse is absent. This is a sparse, distance-derived sensor affinity—not a street graph, learned correlation graph, or claim that each edge is a physical road segment.

The loader never creates an [N, N] tensor. It filters selected distance facts, sorts the surviving source-target pairs, and emits one Graph plus its aligned affinity[E].

An independent audit against the pinned adj_mx.pkl found identical sensor order and sparse support. The largest weight difference was one float32 rounding step, 1.1920928955078125e-07.

Public boundary

from tinymesh.datasets import metr_la

data = metr_la(device="CPU")

print(data.speed.shape, data.observed.shape)
# (34272, 207) (34272, 207)
print(data.graph.nodes, data.graph.edges, data.affinity.shape)
# 207 1722 (1722,)

METRLA owns only facts this source can support:

graph          sparse directed sensor connectivity
sensor_ids     [N] stable source identity
timestamps     [T] ordered naive datetimes
speed          [T, N] raw float32 readings
observed       [T, N] derived zero-sentinel mask
affinity       [E] positive COO-aligned graph weight
sample_minutes 5

It rejects mismatched columns, irregular timestamps, duplicate identities or selected edges, missing zero-distance self-edges, non-finite or negative values, malformed schemas, oversized local files, and tensor-axis misalignment.

No generic temporal container changes in this stage. Forecast windows, time-of-day fields, imputation, normalization, split policy, targets, losses, and model unrolling need a task contract and remain outside the loader.

Full-source witness

Revision 33e81efd26ecf370c21b5ca95880491105a1c994 produced matching CPU and Metal observations:

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

{
  "nodes": 207,
  "edges": 1722,
  "steps": 34272,
  "first_timestamp": "2012-03-01 00:00:00",
  "last_timestamp": "2012-06-27 23:55:00",
  "sample_minutes": 5,
  "values": 7094304,
  "observed_values": 6519002,
  "missing_values": 575302,
  "self_loops": 207,
  "asymmetric_edges": 1111,
  "minimum_affinity": 0.10008395463228226,
  "maximum_affinity": 1.0,
  "affinity_sha256": "d4db2dce0cfd83ec40c115372881cd409059e3ddcb2c4adf8f34502b7a7a00e5"
}

The matching affinity digest proves device-independent lowering at this revision. It does not show that the affinity improves a forecast. The METR-LA forecast now owns the forward split, train-only preprocessing, missing-value-aware loss, temporal controls, and false topologies; its learned graph comparison remains pending.

The cross-dataset network measurement finds a 195-node strong core, 13 strong components, and two weak components. One sensor has only its retained self-loop and is therefore isolated from non-self messages. Directed reachability is 93.8%, with mean reachable distance 5.85 hops and diameter 17. The affinity is broad but not fully connected.