preprocess.py

#!/usr/bin/env python
"""
Build train.csv and test.csv from raw GPS and constellation data.

Run this once before any training, evaluation, or ablation:
    python preprocess.py

Output
------
  data/train.csv  — aggregated, feature-engineered, labeled training rows
  data/test.csv   — same for the cross-test split
"""

import pandas as pd

from const import (
    DEFAULT_TRAIN_CONFIG,
    DEFAULT_TEST_CONFIG,
    EXCLUDED_TESTS,
    TRAIN_CSV,
    TEST_CSV,
)
from utils import read_raw_csvs, create_agg_fe_labeled_from_df


def build_split(config, gps_df, const_df):
    frames = []
    for c in config:
        df = create_agg_fe_labeled_from_df(
            gps_df=gps_df,
            constellation_df=const_df,
            day=c["day"],
            hour_1=c["hour_1"],
            hour_2=c["hour_2"],
            agg_size="1S",
            loc_id=c["loc_id"],
        )
        frames.append(df[~df["test_id"].isin(EXCLUDED_TESTS)])
    return pd.concat(frames, ignore_index=True)


def main():
    print("Reading raw CSV data …")
    gps_df, const_df = read_raw_csvs()

    print("Building train split …")
    train_df = build_split(DEFAULT_TRAIN_CONFIG, gps_df, const_df)
    train_df.to_csv(TRAIN_CSV, index=False)
    print(f"  {len(train_df):,} rows → {TRAIN_CSV}")

    print("Building test split …")
    test_df = build_split(DEFAULT_TEST_CONFIG, gps_df, const_df)
    test_df.to_csv(TEST_CSV, index=False)
    print(f"  {len(test_df):,} rows → {TEST_CSV}")


if __name__ == "__main__":
    main()