scripts/demo_train_eval.py

#!/usr/bin/env python3
"""End-to-end demonstration: eval an untrained model, train briefly, eval again,
and show the metrics improve. Also writes the §4 protocol-template report and an
extrapolation comparison image. Synthetic data only (sanity, not research scale)."""

import argparse
import time

import torch

from mapgs.config import load_config
from mapgs.data import SyntheticDataset, collate_samples
from mapgs.eval import Evaluator, write_report
from mapgs.train import Trainer


def fmt(d):
    return {k: (round(v, 3) if isinstance(v, float) else v) for k, v in d.items()}


def main():
    ap = argparse.ArgumentParser()
    ap.add_argument("--iters", type=int, default=400)
    ap.add_argument("--n-scenes", type=int, default=24)
    ap.add_argument("--out", type=str, default="/mnt/william/eval_report.md")
    args = ap.parse_args()

    cfg = load_config("/mnt/william/configs/synthetic_smoke.yaml", [
        "data.root=/mnt/william/data/synthetic", "train.batch_size=4",
        f"train.iters={args.iters}", "train.log_every=50", "train.ckpt_every=0",
        "train.extrap_ramp_iter=100",
    ])
    print("generating data ...")
    train_ds = SyntheticDataset(cfg, "train", n_scenes=args.n_scenes, n_sup_views=6, device="cuda")
    val_ds = SyntheticDataset(cfg, "val", n_scenes=8, n_sup_views=6, device="cuda")

    trainer = Trainer(cfg)
    ev = Evaluator(trainer.model, cfg, device="cuda")

    print("\n=== BEFORE training (random init) ===")
    before = ev.interpolation(val_ds, max_scenes=8)
    before_ext = ev.extrapolation_sweep(val_ds, shifts=[2.0], max_scenes=6, frame=cfg.data.num_frames // 2)
    print("interp:", fmt(before), "| extrap@2m:", fmt(before_ext[2.0]))

    print(f"\n=== training {args.iters} iters ===")
    t = time.time()
    trainer.fit(train_ds, max_iters=args.iters)
    print(f"trained in {time.time()-t:.0f}s")

    print("\n=== AFTER training ===")
    trainer.model.eval()
    after = ev.interpolation(val_ds, max_scenes=8)
    after_ext = ev.extrapolation_sweep(val_ds, shifts=list(cfg.eval.lateral_shifts), max_scenes=6,
                                       frame=cfg.data.num_frames // 2)
    lane = ev.lane_consistency(val_ds, max_scenes=6, frame=cfg.data.num_frames // 2)
    print("interp:", fmt(after))
    for sh, m in after_ext.items():
        print(f"  extrap@{sh}m:", fmt(m))
    print("lane:", fmt(lane))

    print("\n=== IMPROVEMENT (interpolation) ===")
    print(f"  PSNR : {before['PSNR']:.2f} -> {after['PSNR']:.2f}")
    print(f"  SSIM : {before['SSIM']:.3f} -> {after['SSIM']:.3f}")
    print(f"  LPIPS: {before['LPIPS']:.3f} -> {after['LPIPS']:.3f}")
    print(f"  D-RMSE(m): {before['D-RMSE']:.3f} -> {after['D-RMSE']:.3f}")

    write_report(args.out, interpolation={"Synthetic(val)": after}, extrapolation=after_ext, lane=lane,
                 header=f"_Demo run: {args.iters} iters, {args.n_scenes} synthetic scenes, "
                        f"config synthetic_smoke. Numbers are sanity-scale, not research results._")
    trainer.save("/mnt/william/runs/mapgs_synth/ckpt_demo.pt")
    print(f"\nreport -> {args.out}")


if __name__ == "__main__":
    main()