examples/baseline_constant_velocity.py

#!/usr/bin/env python3
"""Constant-velocity baseline for EnvShip-Bench v2 (Track A).

Reads the paper-default filtered subset of one jurisdiction and reports
average and final displacement error in metres. This is the simplest
sanity-check predictor: assume each vessel keeps the velocity it had
at the last history point.

The same loader works for Track B — change the path and the trajectory
will be 90 + 180 points instead of 30 + 30.

Usage
-----
    python examples/baseline_constant_velocity.py \\
        --csv data/envship_v2/track_a_short-term_Cross-domain_Datasets/dma_track_v1/test/part-000.csv.gz

Expected output (DMA Track A test split, filtered):
    samples: 14878
    ADE: 87.4 m   FDE: 184.2 m
"""
from __future__ import annotations

import argparse
import gzip
import json
import csv

import numpy as np


def load_split(csv_path: str) -> tuple[np.ndarray, np.ndarray, np.ndarray]:
    """Return (hist_xy, fut_xy, keep_mask).

    Each row of the gzipped CSV holds:
      - hist_x_json, hist_y_json (30 history points for Track A,
                                  90 for Track B)
      - fut_x_json,  fut_y_json  (30 / 180 future points)
    Positions are metres relative to the trajectory anchor.
    """
    hist_x, hist_y, fut_x, fut_y, keep = [], [], [], [], []
    with gzip.open(csv_path, "rt", encoding="utf-8") as fh:
        reader = csv.DictReader(fh)
        has_flag = "osm_temporal_consistent" in (reader.fieldnames or [])
        for row in reader:
            hist_x.append(json.loads(row["hist_x_json"]))
            hist_y.append(json.loads(row["hist_y_json"]))
            fut_x.append(json.loads(row["fut_x_json"]))
            fut_y.append(json.loads(row["fut_y_json"]))
            keep.append(row.get("osm_temporal_consistent") == "true" if has_flag else True)
    hist = np.stack([np.asarray(hist_x), np.asarray(hist_y)], axis=-1)  # (N, T_hist, 2)
    fut  = np.stack([np.asarray(fut_x),  np.asarray(fut_y)],  axis=-1)  # (N, T_fut,  2)
    return hist, fut, np.asarray(keep, dtype=bool)


def constant_velocity_predict(hist: np.ndarray, n_fut: int) -> np.ndarray:
    """Extrapolate from the last two history points (last-step velocity)."""
    v = hist[:, -1] - hist[:, -2]                       # (N, 2)
    steps = np.arange(1, n_fut + 1)                     # (T_fut,)
    return hist[:, -1:] + steps[None, :, None] * v[:, None, :]


def displacement_errors(pred: np.ndarray, target: np.ndarray) -> tuple[float, float]:
    """Per-step euclidean error, then ADE (mean across time) and FDE (last step)."""
    d = np.linalg.norm(pred - target, axis=-1)          # (N, T_fut)
    return float(d.mean()), float(d[:, -1].mean())


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--csv", required=True,
                    help="path to <subset>/{split}/part-000.csv.gz")
    ap.add_argument("--apply-default-filter", action="store_true", default=True,
                    help="drop rows where osm_temporal_consistent != 'true' (paper default)")
    args = ap.parse_args()

    hist, fut, keep = load_split(args.csv)
    if args.apply_default_filter:
        hist, fut = hist[keep], fut[keep]
    n_fut = fut.shape[1]
    pred = constant_velocity_predict(hist, n_fut)
    ade, fde = displacement_errors(pred, fut)

    print(f"samples: {len(fut)}")
    print(f"ADE: {ade:.1f} m   FDE: {fde:.1f} m")


if __name__ == "__main__":
    main()