Blame for scripts/verify.py

"""verify.py —— 换机器后先跑这个。检查下载是否完整、切分是否自洽。

检查项：
  V1 计数：每个 (generator, split) 的样本数与 meta/counts.csv 一致
  V2 切分互斥：train / val / test 三段的 md5 两两无交集
  V3 真图唯一：generator == "nature" 的 md5 不重复，且每折共享同一批
  V4 字节完整：抽样重算 md5，与行内记录一致（说明下载没被截断/损坏）
  V5 生成器齐全：8 个生成器的 fake 都存在

用法：
    python scripts/verify.py --root .
    python scripts/verify.py --root . --samples 500
"""
from __future__ import annotations

import argparse
import hashlib
import random
import sys
from pathlib import Path

import pandas as pd
import pyarrow.parquet as pq

OK, BAD = "PASS", "FAIL"


def image_bytes(cell):
    if isinstance(cell, dict):
        return cell.get("bytes")
    if isinstance(cell, (bytes, bytearray, memoryview)):
        return bytes(cell)
    return None


def load_meta(root: Path) -> pd.DataFrame:
    frames = []
    for f in sorted((root / "data").glob("*.parquet")):
        t = pq.read_table(f, columns=["md5", "label", "generator", "family",
                                      "class_id", "split", "bytes"])
        frames.append(t.to_pandas())
    if not frames:
        sys.exit("[!] data/ 下没有 parquet")
    return pd.concat(frames, ignore_index=True)


def main() -> None:
    ap = argparse.ArgumentParser()
    ap.add_argument("--root", required=True)
    ap.add_argument("--samples", type=int, default=200)
    ap.add_argument("--seed", type=int, default=20261029)
    args = ap.parse_args()
    root = Path(args.root)
    results = []

    print(f"读索引 {root / 'data'} …")
    df = load_meta(root)
    print(f"  总样本 {len(df):,}  /  {df['bytes'].sum() / 1e9:.1f} GB")

    # V1
    counts_csv = root / "meta" / "counts.csv"
    if counts_csv.exists():
        ref = pd.read_csv(counts_csv)
        ref_set = {(r.generator, r.label, r.split): int(r.n) for r in ref.itertuples()}
        have = df.groupby(["generator", "label", "split"]).size().to_dict()
        mism = {k: (ref_set[k], have.get(k, 0)) for k in ref_set if ref_set[k] != have.get(k, 0)}
        results.append((f"V1 计数与 meta/counts.csv 一致（{len(ref_set)} 组）", OK if not mism else BAD, mism))
    else:
        results.append(("V1 counts.csv 缺失，跳过", "SKIP", {}))

    # V2
    sets = {s: set(df.loc[df.split == s, "md5"]) for s in ("train", "val", "test")}
    inter = {
        "train∩val": len(sets["train"] & sets["val"]),
        "train∩test": len(sets["train"] & sets["test"]),
        "val∩test": len(sets["val"] & sets["test"]),
    }
    results.append(("V2 train/val/test 互斥", OK if not any(inter.values()) else BAD, inter))

    # V3
    nat = df[df.generator == "nature"]
    dup = int(len(nat) - nat["md5"].nunique())
    results.append(("V3 真图池 md5 唯一", OK if dup == 0 else BAD, {"duplicates": dup, "n": len(nat)}))

    # V4 抽样校验 md5
    rng = random.Random(args.seed)
    files = sorted((root / "data").glob("*.parquet"))
    want = rng.sample(range(len(df)), min(args.samples, len(df)))
    want_set = set(df.iloc[want]["md5"])
    checked = bad = 0
    for f in files:
        if not want_set:
            break
        pf = pq.ParquetFile(f)
        for batch in pf.iter_batches(batch_size=64):
            d = batch.to_pydict()
            for i in range(batch.num_rows):
                m = d["md5"][i]
                if m not in want_set:
                    continue
                want_set.discard(m)
                b = image_bytes(d["image"][i])
                checked += 1
                if b is None or hashlib.md5(b).hexdigest() != m:
                    bad += 1
            if not want_set:
                break
    results.append((f"V4 抽样 md5 校验（{checked} 张）", OK if bad == 0 and checked > 0 else BAD,
                    {"checked": checked, "corrupt": bad}))

    # V5
    gens = sorted(df.loc[df.label == 1, "generator"].unique())
    results.append((f"V5 生成器齐全（{len(gens)}）", OK if len(gens) >= 8 else "WARN", gens))

    print()
    fail = 0
    for name, status, detail in results:
        mark = {"PASS": "✅", "FAIL": "❌", "WARN": "⚠️", "SKIP": "⏭"}[status]
        print(f"{mark} {name}  {status}")
        if detail and status != "PASS":
            print(f"     {detail}")
        if status == "FAIL":
            fail += 1

    print()
    print("逐生成器 × split 样本数：")
    print(df.groupby(["generator", "split"]).size().unstack(fill_value=0).to_string())
    print()
    if fail:
        sys.exit(f"[!] {fail} 项未通过，先别开始训练")
    print("全部通过，可以开始训练。")


if __name__ == "__main__":
    main()