Evaluation / evaluate_matrix_game_sc.py
VideoWorldmodel's picture
Upload evaluate_matrix_game_sc.py
6f9d7c9 verified
Raw
History Blame Contribute Delete
14.6 kB
#!/usr/bin/env python3
"""Reproduce Matrix-Game 2.0 self-consistency metrics from released videos."""
from __future__ import annotations
import argparse
import csv
import json
import math
import re
import sys
from collections import defaultdict
from dataclasses import asdict, dataclass
from pathlib import Path
from typing import Sequence
import cv2
import numpy as np
EXPECTED_COUNTS = {"inverse": 448, "loop": 445, "equivalence": 239}
PAPER_RESULTS = {
"inverse": {"lpips": 0.71, "psnr": 10.45},
"loop": {"lpips": 0.72, "psnr": 10.62},
"equivalence": {"lpips": 0.59, "psnr": 12.57},
}
EQUIVALENCE_RE = re.compile(
r"(?P<run>run_\d+_\d+)__pair_(?P<pair>\d+)_(?P<branch>[AB])_traj_\d+\.mp4$",
re.IGNORECASE,
)
@dataclass(frozen=True)
class EvaluationUnit:
relation: str
unit: str
video_a: Path
video_b: Path | None = None
@dataclass(frozen=True)
class MetricRow:
relation: str
unit: str
video_a: str
video_b: str
frame_a: int
frame_b: int
width: int
height: int
psnr: float
lpips: float | None
psnr_exact_match: bool
def is_downloaded_video(path: Path) -> bool:
if not path.is_file() or path.stat().st_size <= 200:
return False
with path.open("rb") as handle:
return not handle.read(64).startswith(b"version https://git-lfs.github.com/spec/v1")
def discover_units(root: Path, strict_counts: bool = True) -> list[EvaluationUnit]:
units: list[EvaluationUnit] = []
for relation in ("inverse", "loop"):
for difficulty in ("easy", "hard"):
folder = root / f"{relation}_{difficulty}"
if not folder.is_dir():
raise FileNotFoundError(f"missing dataset folder: {folder}")
for path in sorted(folder.glob("*.mp4")):
if not is_downloaded_video(path):
raise RuntimeError(f"missing Git LFS video object: {path}")
units.append(EvaluationUnit(relation, path.stem, path))
equivalence_folder = root / "equivalence"
if not equivalence_folder.is_dir():
raise FileNotFoundError(f"missing dataset folder: {equivalence_folder}")
pairs: dict[str, dict[str, Path]] = defaultdict(dict)
for path in sorted(equivalence_folder.glob("*.mp4")):
if not is_downloaded_video(path):
raise RuntimeError(f"missing Git LFS video object: {path}")
match = EQUIVALENCE_RE.fullmatch(path.name)
if not match:
raise ValueError(f"unrecognized Equivalence filename: {path.name}")
unit = f"{match.group('run')}__pair_{match.group('pair')}"
branch = match.group("branch").upper()
if branch in pairs[unit]:
raise ValueError(f"duplicate Equivalence branch {branch}: {unit}")
pairs[unit][branch] = path
for unit, branches in sorted(pairs.items()):
if set(branches) != {"A", "B"}:
raise ValueError(f"incomplete Equivalence pair {unit}: {sorted(branches)}")
units.append(EvaluationUnit("equivalence", unit, branches["A"], branches["B"]))
counts = {relation: sum(unit.relation == relation for unit in units) for relation in EXPECTED_COUNTS}
if strict_counts and counts != EXPECTED_COUNTS:
raise RuntimeError(f"unexpected graph counts: found {counts}, expected {EXPECTED_COUNTS}")
return units
def read_endpoint(path: Path, endpoint: str) -> tuple[np.ndarray, int]:
capture = cv2.VideoCapture(str(path))
if not capture.isOpened():
raise RuntimeError(f"failed to open video: {path}")
frame_count = int(capture.get(cv2.CAP_PROP_FRAME_COUNT))
if frame_count < 2:
capture.release()
raise RuntimeError(f"video has fewer than two frames: {path}")
frame_index = 0 if endpoint == "first" else frame_count - 1
capture.set(cv2.CAP_PROP_POS_FRAMES, frame_index)
ok, frame = capture.read()
capture.release()
if not ok:
raise RuntimeError(f"failed to decode {endpoint} frame: {path}")
return cv2.cvtColor(frame, cv2.COLOR_BGR2RGB), frame_index
def psnr(reference: np.ndarray, prediction: np.ndarray) -> float:
if reference.shape != prediction.shape:
raise ValueError(f"frame shape mismatch: {reference.shape} versus {prediction.shape}")
mse = np.mean((reference.astype(np.float64) - prediction.astype(np.float64)) ** 2)
if mse == 0:
return float("inf")
return 10.0 * math.log10((255.0**2) / mse)
class LPIPSMetric:
def __init__(self, device: str) -> None:
try:
import lpips
import torch
except ImportError as exc:
raise RuntimeError("install the Python dependencies listed in README.md before using --lpips") from exc
if device == "auto":
device = "cuda" if torch.cuda.is_available() else "cpu"
self.torch = torch
self.device = torch.device(device)
self.model = lpips.LPIPS(net="alex").to(self.device).eval()
def __call__(self, reference: np.ndarray, prediction: np.ndarray) -> float:
if reference.shape != prediction.shape:
raise ValueError(f"frame shape mismatch: {reference.shape} versus {prediction.shape}")
tensors = []
for image in (reference, prediction):
tensor = self.torch.from_numpy(np.ascontiguousarray(image)).permute(2, 0, 1).float()
tensors.append(tensor.div(127.5).sub(1.0).unsqueeze(0).to(self.device))
with self.torch.inference_mode():
value = self.model(tensors[0], tensors[1], normalize=False)
return float(value.item())
def evaluate(units: Sequence[EvaluationUnit], lpips_metric: LPIPSMetric | None) -> list[MetricRow]:
rows: list[MetricRow] = []
for index, unit in enumerate(units, start=1):
if unit.relation in {"inverse", "loop"}:
frame_a, index_a = read_endpoint(unit.video_a, "first")
frame_b, index_b = read_endpoint(unit.video_a, "last")
video_b = unit.video_a
else:
if unit.video_b is None:
raise AssertionError("Equivalence unit is missing branch B")
frame_a, index_a = read_endpoint(unit.video_a, "last")
frame_b, index_b = read_endpoint(unit.video_b, "last")
video_b = unit.video_b
value_psnr = psnr(frame_a, frame_b)
value_lpips = lpips_metric(frame_a, frame_b) if lpips_metric else None
rows.append(
MetricRow(
relation=unit.relation,
unit=unit.unit,
video_a=str(unit.video_a),
video_b=str(video_b),
frame_a=index_a,
frame_b=index_b,
width=int(frame_a.shape[1]),
height=int(frame_a.shape[0]),
psnr=value_psnr,
lpips=value_lpips,
psnr_exact_match=math.isinf(value_psnr),
)
)
if index % 100 == 0 or index == len(units):
print(f"evaluated {index}/{len(units)} graph units", flush=True)
return rows
def bootstrap_ci(values: Sequence[float], seed: int, repetitions: int) -> tuple[float, float]:
array = np.asarray(values, dtype=np.float64)
if len(array) < 2:
return float("nan"), float("nan")
rng = np.random.default_rng(seed)
means = np.empty(repetitions, dtype=np.float64)
for start in range(0, repetitions, 500):
count = min(500, repetitions - start)
indices = rng.integers(0, len(array), size=(count, len(array)))
means[start : start + count] = array[indices].mean(axis=1)
low, high = np.quantile(means, [0.025, 0.975])
return float(low), float(high)
def summarize(rows: Sequence[MetricRow], seed: int, repetitions: int) -> list[dict[str, object]]:
output: list[dict[str, object]] = []
for relation in ("inverse", "loop", "equivalence"):
group = [row for row in rows if row.relation == relation]
summary: dict[str, object] = {
"relation": relation,
"n_graphs": len(group),
"psnr_finite_n": sum(math.isfinite(row.psnr) for row in group),
"psnr_exact_match_n": sum(row.psnr_exact_match for row in group),
}
for metric in ("psnr", "lpips"):
values = [
float(value)
for row in group
if (value := getattr(row, metric)) is not None and math.isfinite(float(value))
]
if not values:
continue
low, high = bootstrap_ci(values, seed, repetitions)
summary.update(
{
metric: float(np.mean(values)),
f"{metric}_std": float(np.std(values, ddof=1)) if len(values) > 1 else 0.0,
f"{metric}_ci95_low": low,
f"{metric}_ci95_high": high,
}
)
output.append(summary)
return output
def write_csv(path: Path, rows: Sequence[dict[str, object]]) -> None:
if not rows:
return
columns: list[str] = []
for row in rows:
for key in row:
if key not in columns:
columns.append(key)
with path.open("w", newline="", encoding="utf-8") as handle:
writer = csv.DictWriter(handle, fieldnames=columns)
writer.writeheader()
writer.writerows(rows)
def paper_check(summaries: Sequence[dict[str, object]]) -> tuple[bool, list[dict[str, object]]]:
checks: list[dict[str, object]] = []
passed = True
for summary in summaries:
relation = str(summary["relation"])
for metric in ("lpips", "psnr"):
value = summary.get(metric)
expected = PAPER_RESULTS[relation][metric]
metric_passed = value is not None and round(float(value), 2) == expected
checks.append(
{
"relation": relation,
"metric": metric,
"computed": value,
"paper_rounded": expected,
"pass_at_2_decimals": metric_passed,
}
)
passed = passed and metric_passed
return passed, checks
def report(summaries: Sequence[dict[str, object]], check_passed: bool | None) -> str:
lines = [
"# Matrix-Game 2.0 SC Reproduction",
"",
"| Relation | Graph N | LPIPS (95% CI) | PSNR dB (95% CI) | Exact PSNR pairs |",
"| --- | ---: | ---: | ---: | ---: |",
]
for row in summaries:
lpips_text = "not computed"
if "lpips" in row:
lpips_text = (
f"{row['lpips']:.4f} [{row['lpips_ci95_low']:.4f}, "
f"{row['lpips_ci95_high']:.4f}]"
)
psnr_text = f"{row['psnr']:.4f} [{row['psnr_ci95_low']:.4f}, {row['psnr_ci95_high']:.4f}]"
lines.append(
f"| {str(row['relation']).title()} | {row['n_graphs']} | {lpips_text} | "
f"{psnr_text} | {row['psnr_exact_match_n']} |"
)
if check_passed is not None:
lines.extend(["", f"Paper rounded-value check: **{'PASS' if check_passed else 'FAIL'}**."])
lines.extend(
[
"",
"Inverse/Loop compare the generated first and final frames. Equivalence compares",
"the generated final frames of paired A/B rollouts. Confidence intervals use",
"10,000 graph-level bootstrap resamples by default.",
"",
]
)
return "\n".join(lines)
def parse_args(argv: Sequence[str] | None = None) -> argparse.Namespace:
parser = argparse.ArgumentParser(description=__doc__)
parser.add_argument("--data", type=Path, default=Path("data/Nips_WM_Eval_qzf"))
parser.add_argument("--output", type=Path, default=Path("results/matrix_game_sc"))
parser.add_argument("--lpips", action="store_true", help="compute LPIPS 0.1.4 with AlexNet")
parser.add_argument("--device", default="auto", help="auto, cpu, cuda, or cuda:<index>")
parser.add_argument("--seed", type=int, default=2026)
parser.add_argument("--bootstrap-repetitions", type=int, default=10_000)
parser.add_argument("--allow-partial", action="store_true", help="disable published-count checks")
parser.add_argument("--check-paper", action="store_true", help="check values at paper precision")
return parser.parse_args(argv)
def main(argv: Sequence[str] | None = None) -> int:
args = parse_args(argv)
if not args.data.is_dir():
raise SystemExit(f"dataset directory does not exist: {args.data}")
if args.check_paper and not args.lpips:
raise SystemExit("--check-paper requires --lpips")
args.output.mkdir(parents=True, exist_ok=True)
units = discover_units(args.data, strict_counts=not args.allow_partial)
metric = LPIPSMetric(args.device) if args.lpips else None
rows = evaluate(units, metric)
summaries = summarize(rows, args.seed, args.bootstrap_repetitions)
check_passed: bool | None = None
checks: list[dict[str, object]] = []
if args.check_paper:
check_passed, checks = paper_check(summaries)
write_csv(args.output / "per_graph.csv", [asdict(row) for row in rows])
write_csv(args.output / "summary.csv", summaries)
if checks:
write_csv(args.output / "paper_check.csv", checks)
audit = {
"data": str(args.data.resolve()),
"definitions": {
"inverse": "generated first frame versus generated final frame",
"loop": "generated first frame versus generated final frame",
"equivalence": "generated branch-A final frame versus generated branch-B final frame",
},
"expected_counts": EXPECTED_COUNTS,
"observed_counts": {row["relation"]: row["n_graphs"] for row in summaries},
"lpips": "lpips==0.1.4, AlexNet, RGB in [-1,1]" if args.lpips else "not computed",
"psnr": "RGB uint8, MAX=255; exact matches excluded from finite PSNR mean and counted separately",
"bootstrap_seed": args.seed,
"bootstrap_repetitions": args.bootstrap_repetitions,
"paper_check_passed": check_passed,
}
(args.output / "audit.json").write_text(
json.dumps(audit, indent=2, ensure_ascii=True) + "\n", encoding="utf-8"
)
(args.output / "report.md").write_text(report(summaries, check_passed), encoding="utf-8")
print(report(summaries, check_passed))
print(f"Outputs: {args.output.resolve()}")
return 0 if check_passed is not False else 1
if __name__ == "__main__":
raise SystemExit(main())