METR-LA Traffic Dataset - TsFile Conversion

This repository contains a TsFile conversion of witgaw/METR-LA, a Hugging Face version of the METR-LA traffic forecasting dataset commonly used with the DCRNN model.

Modalities: Time-series.

Source Dataset

The source dataset contains pre-windowed traffic forecasting records with temporal train, validation, and test splits. Each row represents one traffic sensor at one reference timestamp. METR-LA has 207 sensors, 5-minute temporal resolution, 12 historical input steps, and a 12-step / 1-hour prediction horizon.

Original citation:

@inproceedings{li2018dcrnn_traffic,
  title={Diffusion Convolutional Recurrent Neural Network: Data-Driven Traffic Forecasting},
  author={Li, Yaguang and Yu, Rose and Shahabi, Cyrus and Liu, Yan},
  booktitle={International Conference on Learning Representations},
  year={2018}
}

Converted Data

Total converted rows: 7,089,543. Total TsFile size: 619,695,942 bytes.

• train: 4,962,618 rows, 207 sensors, 48 fields, 436,357,738 bytes
• val: 708,975 rows, 207 sensors, 48 fields, 61,623,430 bytes
• test: 1,417,950 rows, 207 sensors, 48 fields, 121,714,774 bytes

The source sensor graph files are mirrored under sensor_graph/.

TsFile Schema

• Time: parsed from source t0_timestamp as INT64 milliseconds.
• TAG: node_id, the traffic sensor identifier.
• FIELD: all source x_t*_d* historical input features and y_t*_d* future target features, stored as FLOAT.

Conversion Notes

• Source t0_timestamp is not stored as a field because it is encoded as Time.
• node_id is cast to string and declared as the TsFile TAG column.
• Feature columns are preserved, but schema-unsafe offset markers are renamed.

Example renamed columns:

• x_t-11_d0 -> x_t_minus_11_d0
• x_t-11_d1 -> x_t_minus_11_d1
• x_t-10_d0 -> x_t_minus_10_d0
• x_t-10_d1 -> x_t_minus_10_d1
• x_t-9_d0 -> x_t_minus_9_d0
• x_t-9_d1 -> x_t_minus_9_d1
• x_t-8_d0 -> x_t_minus_8_d0
• x_t-8_d1 -> x_t_minus_8_d1
• ... 40 additional renamed columns

Read Example

# Read `metr_la_train.tsfile` with the Apache TsFile Java/Python SDK.

Usage

Install the Apache TsFile Python SDK (pip install tsfile) and read a converted file:

from pathlib import Path
from tsfile import TsFileReader

path = Path("metr_la_test.tsfile")
with TsFileReader(str(path)) as reader:
    schemas = reader.get_all_table_schemas()
    print("tables:", list(schemas))
    table_name = next(iter(schemas))
    table = schemas[table_name]
    columns = [column.get_column_name() for column in table.get_columns()]
    print("columns:", columns)
    field_names = [
        column.get_column_name()
        for column in table.get_columns()
        if column.get_column_name() not in {"Time", "time"}
    ]
    if field_names:
        with reader.query_table(table_name, field_names[:3], batch_size=1024) as result:
            batch = result.read_arrow_batch()
            if batch is not None:
                print(batch.to_pandas().head())