Datasets:

moiaraya
/

I-AsSET

ArXiv:

I-AsSET / Scripts /iasset_embeddings_classwise.py

Add scripts for sim-to-real feature extraction (DINO, Hash, Incpetion), similarity metrics such as cosine distance, KID, etc. And visualizations t-SNE, PCA and UMAP.

d73849c 22 days ago

Raw History Blame Contribute Delete

23 kB

#!/usr/bin/env python3
"""Class-wise, background-free domain-gap analysis for IASSET.
Every YOLO bounding box becomes one sample (the pixels under the box only).
Metrics are computed per class (real crops of class c vs. synthetic crops of
class c), pooled over all classes ("ALL"), and macro-averaged over classes
("MEAN_OVER_CLASSES"). Class ids are mapped to canonical names taken from the
real test set, so datasets with a different class order are handled by name.
This file reuses encoders, metrics and reduction code from
iasset_embeddings_analysis.py (keep both files in the same folder).
Stages
  index    : parse labels, map classes, write index.csv + class_stats.csv (run this first and check it)
  features : hashes + DINOv2 + Inception embeddings of the crops
  metrics  : per-class similarity metrics (metrics_long.csv)
  reduce   : offline PCA / t-SNE / UMAP projections
  verify   : class-mapping sanity check (needs dinov2 features; not part of "all")
  all      : features + metrics + reduce
Example
  python iasset_embeddings_classwise.py index --root /media/rtx5090/IASSET
  python iasset_embeddings_classwise.py all   --root /media/rtx5090/IASSET
"""
import argparse
import json
import os
import re
from concurrent.futures import ProcessPoolExecutor
from pathlib import Path

import numpy as np
import pandas as pd
import torch
from PIL import Image
from torch.utils.data import DataLoader, Dataset
from tqdm import tqdm

from iasset_embeddings_analysis import (HASH_SIZES, SKIP_DATASETS, THUMB, _compare, list_images,
                                        load_dinov2, load_inception, reduce_features, sample_ids)

DEFAULT_OUT = "/media/rtx5090/IASSET/Scripts/embeddings_results_classwise"
FILL = (124, 116, 104)  # ImageNet mean colour, used for letterbox padding
_norm = lambda s: re.sub(r"[^a-z0-9]", "", str(s).lower())


# --------------------------------------------------------------------------- crops
def get_crop(img, box, pad=0.0, mode="letterbox"):
    """Pixels under a (x1, y1, x2, y2) box. letterbox: pad to a square with FILL; squash: raw crop."""
    x1, y1, x2, y2 = box
    dx, dy = pad * (x2 - x1), pad * (y2 - y1)
    W, H = img.size
    l, t = max(0, int(np.floor(x1 - dx))), max(0, int(np.floor(y1 - dy)))
    r, b = min(W, int(np.ceil(x2 + dx))), min(H, int(np.ceil(y2 + dy)))
    c = img.crop((l, t, max(r, l + 1), max(b, t + 1)))
    if mode == "letterbox":
        s = max(c.size)
        canvas = Image.new("RGB", (s, s), FILL)
        canvas.paste(c, ((s - c.width) // 2, (s - c.height) // 2))
        return canvas
    return c


# --------------------------------------------------------------------------- labels and classes
def _walk(root, max_depth=3, skip=("images", "depth", "rgb")):
    root = Path(root)
    base = len(root.parts)
    for dp, dn, fn in os.walk(root):
        depth = len(Path(dp).parts) - base
        dn[:] = sorted(d for d in dn if d.lower() not in skip) if depth < max_depth else []
        for f in sorted(fn):
            yield Path(dp) / f


def read_names(root):
    """Class names (list, index = id) from a YOLO yaml with 'names' or from classes.txt."""
    import yaml
    files = list(_walk(root))
    yamls = sorted((f for f in files if f.suffix.lower() in (".yaml", ".yml")),
                   key=lambda f: f.name.lower() not in ("dataset.yaml", "data.yaml"))  # these first
    for f in yamls:
        try:
            d = yaml.safe_load(f.read_text())
        except Exception:
            continue
        if isinstance(d, dict) and "names" in d:
            n = d["names"]
            if isinstance(n, dict):
                n = [n[k] for k in sorted(n, key=int)]
            return [str(x) for x in n], f
    for f in files:
        if f.name.lower() in ("classes.txt", "names.txt", "obj.names"):
            return [l.strip() for l in f.read_text().splitlines() if l.strip()], f
    return None, None


def build_map(names, canon, override=None):
    """source class id -> canonical class id (matched by normalized name)."""
    cn = {_norm(n): i for i, n in enumerate(canon)}
    if override:
        return {int(k): cn[_norm(v)] for k, v in override.items()}
    if names is None:
        return {i: i for i in range(len(canon))}
    return {sid: cn[_norm(n)] for sid, n in enumerate(names) if _norm(n) in cn}


def parse_label(p):
    """YOLO txt -> [(class, x1, y1, x2, y2)] normalized. Handles boxes and polygons."""
    out = []
    for line in Path(p).read_text().splitlines():
        t = line.split()
        if len(t) < 5:
            continue
        try:
            cid, v = int(float(t[0])), [float(x) for x in t[1:]]
        except ValueError:
            continue
        if len(v) == 4:
            cx, cy, w, h = v
            out.append((cid, cx - w / 2, cy - h / 2, cx + w / 2, cy + h / 2))
        else:
            xs, ys = v[0::2], v[1::2]
            out.append((cid, min(xs), min(ys), max(xs), max(ys)))
    return out


class LabelFinder:
    """Find the YOLO label file of an image: images->labels swap, sibling folders, then unique-stem lookup."""

    def __init__(self, ds_root):
        self.root, self.stems = Path(ds_root), None

    def __call__(self, img):
        parts = list(img.relative_to(self.root).with_suffix(".txt").parts)
        for i in range(len(parts) - 2, -1, -1):
            if parts[i].lower() == "images":
                c = self.root.joinpath(*(parts[:i] + ["labels"] + parts[i + 1:]))
                if c.exists():
                    return c
        for c in (img.with_suffix(".txt"), img.parent.parent / "labels" / (img.stem + ".txt"),
                  img.parent.parent / "labels" / img.parent.name / (img.stem + ".txt")):
            if c.exists():
                return c
        if self.stems is None:
            seen, self.stems = {}, {}
            for f in _walk(self.root, max_depth=5):
                if f.suffix == ".txt" and f.name.lower() not in ("classes.txt", "names.txt"):
                    seen[f.stem] = None if f.stem in seen else f
            self.stems = {k: v for k, v in seen.items() if v is not None}
        return self.stems.get(img.stem)


def build_crop_index(a):
    root = Path(a.root)
    real_root = root / "Real_Test_Set"
    sets = [("REAL", "Real", "real", real_root / "images" / "full" / "rgb", real_root)]
    synth = root / "Synthetic_Train_Sets"
    for g in sorted(p for p in synth.iterdir() if p.is_dir() and p.name not in a.skip_groups):
        for d in sorted(p for p in g.iterdir() if p.is_dir() and p.name not in SKIP_DATASETS):
            sets.append((f"{g.name}/{d.name}", g.name, "synthetic", d, d))
    canon, src = read_names(real_root)
    if canon is None:
        raise SystemExit(f"No class names found under {real_root} (dataset.yaml 'names' or classes.txt).")
    print(f"Canonical classes from {src}: {dict(enumerate(canon))}")
    override = json.load(open(a.class_map)) if a.class_map else {}
    rng = np.random.default_rng(a.seed)
    cols = ["dataset", "group", "domain", "path", "class_id", "class_name", "src_class_id",
            "x1", "y1", "x2", "y2", "img_w", "img_h"]
    frames = []
    for name, group, domain, img_dir, ds_root in sets:
        files = list_images(img_dir)
        if a.max_images_per_set and len(files) > a.max_images_per_set:
            files = [files[i] for i in sorted(rng.choice(len(files), a.max_images_per_set, replace=False))]
        names, nsrc = (canon, src) if domain == "real" else read_names(ds_root)
        if names is None and name not in override:
            print(f"[warn] {name}: no class names file, assuming the real-set class order")
        cmap = build_map(names, canon, override.get(name))
        if names is not None and [_norm(n) for n in names] != [_norm(n) for n in canon]:
            print(f"[info] {name}: class list differs from the real set; remapped by name "
                  f"({len(cmap)}/{len(names)} matched)")
        find, rows, n_lab, unmatched = LabelFinder(ds_root), [], 0, set()
        for f in tqdm(files, desc=name[:60], leave=False):
            lp = find(f)
            boxes = parse_label(lp) if lp else []
            if not boxes:
                continue
            n_lab += 1
            with Image.open(f) as im:
                W, H = im.size
            for cid, x1, y1, x2, y2 in boxes:
                if cid not in cmap:
                    unmatched.add(cid)
                    continue
                px = [max(0.0, x1 * W), max(0.0, y1 * H), min(float(W), x2 * W), min(float(H), y2 * H)]
                if min(px[2] - px[0], px[3] - px[1]) < a.min_box_px:
                    continue
                rows.append((name, group, domain, str(f), cmap[cid], canon[cmap[cid]], cid, *px, W, H))
        df = pd.DataFrame(rows, columns=cols)
        if a.max_per_class and len(df):
            keep = []
            for _, g in df.groupby("class_id"):
                ii = g.index.values
                keep += list(rng.choice(ii, a.max_per_class, replace=False) if len(ii) > a.max_per_class else ii)
            df = df.loc[sorted(keep)]
        print(f"{name:75s} images {len(files):6d} | labelled {n_lab:6d} | crops {len(df):6d}"
              + (f" | unmatched src class ids {sorted(unmatched)}" if unmatched else ""))
        frames.append(df)
    idx = pd.concat(frames, ignore_index=True)
    idx.insert(0, "crop_id", np.arange(len(idx)))
    return idx


def get_index(a):
    out = Path(a.out)
    out.mkdir(parents=True, exist_ok=True)
    f = out / "index.csv"
    if f.exists() and not a.overwrite:
        return pd.read_csv(f)
    idx = build_crop_index(a)
    idx.to_csv(f, index=False)
    stats = idx.groupby(["dataset", "class_name"]).size().unstack(fill_value=0)
    stats.to_csv(out / "class_stats.csv")
    with pd.option_context("display.width", 250, "display.max_columns", 50, "display.max_rows", 200):
        print("\nCrops per dataset and class:\n", stats)
    return idx


# --------------------------------------------------------------------------- hashes (raw crops)
def _hash_img(img):
    import imagehash
    img.thumbnail((THUMB, THUMB), Image.LANCZOS)
    fn = {"ahash": imagehash.average_hash, "phash": imagehash.phash,
          "dhash": imagehash.dhash, "whash": imagehash.whash}
    out = {}
    for t, sizes in HASH_SIZES.items():
        for s in sizes:
            h = imagehash.colorhash(img, binbits=s) if t == "colorhash" else fn[t](img, hash_size=s)
            out[f"hash_{t}_{s}"] = h.hash.flatten().astype(np.uint8)
    return out


def _hash_group(task):
    path, boxes, pad = task
    with Image.open(path) as im:
        img = im.convert("RGB")
    return [_hash_img(get_crop(img, b, pad, "squash")) for b in boxes]


def run_hashes(idx, fd, a):
    groups = list(idx.groupby("path", sort=False))
    tasks = [(p, g[["x1", "y1", "x2", "y2"]].values.tolist(), a.pad) for p, g in groups]
    order = np.concatenate([g.index.values for _, g in groups])
    with ProcessPoolExecutor(a.workers) as ex:
        res = [r for rs in tqdm(ex.map(_hash_group, tasks, chunksize=8), total=len(tasks), desc="hashes")
               for r in rs]
    inv = np.argsort(order)
    hex_cols = {}
    for key in res[0]:
        arr = np.stack([r[key] for r in res])[inv]  # back to index order
        np.save(fd / f"{key}.npy", arr)
        hex_cols[key] = [np.packbits(r).tobytes().hex() for r in arr]
    df = pd.DataFrame(hex_cols)
    df.insert(0, "crop_id", idx.crop_id.values)
    df.to_csv(fd.parent / "hashes.csv", index=False)


# --------------------------------------------------------------------------- deep features (crops)
class CropDataset(Dataset):
    def __init__(self, idx, tf, pad, mode):
        self.p = idx.path.values
        self.b = idx[["x1", "y1", "x2", "y2"]].values
        self.tf, self.pad, self.mode = tf, pad, mode
        self._p, self._img = None, None  # per-worker cache: consecutive crops share one image

    def __len__(self):
        return len(self.p)

    def __getitem__(self, i):
        if self.p[i] != self._p:
            with Image.open(self.p[i]) as im:
                self._img = im.convert("RGB")
            self._p = self.p[i]
        return self.tf(get_crop(self._img, self.b[i], self.pad, self.mode))


def extract_crops(idx, fwd, tf, a):
    dl = DataLoader(CropDataset(idx, tf, a.pad, a.crop_mode), batch_size=a.batch_size,
                    num_workers=a.workers, pin_memory=a.device.startswith("cuda"),
                    persistent_workers=a.workers > 0)
    acc = {}
    with torch.no_grad(), torch.autocast("cuda", dtype=torch.float16,
                                         enabled=(a.fp16 and a.device.startswith("cuda"))):
        for x in tqdm(dl, desc="encode crops"):
            for k, v in fwd(x.to(a.device, non_blocking=True)).items():
                acc.setdefault(k, []).append(v.float().cpu())
    return {k: torch.cat(v).numpy() for k, v in acc.items()}


def stage_features(a):
    idx = get_index(a)
    fd = Path(a.out) / "features"
    fd.mkdir(exist_ok=True)
    if "hash" in a.encoders:
        keys = [f"hash_{t}_{s}" for t, ss in HASH_SIZES.items() for s in ss]
        if a.overwrite or not all((fd / f"{k}.npy").exists() for k in keys):
            run_hashes(idx, fd, a)
    jobs = []
    if "dinov2" in a.encoders:
        for size in a.dinov2_sizes:
            jobs.append(([f"dinov2_{size}_cls", f"dinov2_{size}_patchmean"], lambda s=size: load_dinov2(s, a)))
    if "inception" in a.encoders:
        jobs.append((["inception_pool3"], lambda: load_inception(a)))
    for names, loader in jobs:
        if not a.overwrite and all((fd / f"{n}.npy").exists() for n in names):
            continue
        fwd, tf = loader()
        for k, arr in extract_crops(idx, fwd, tf, a).items():
            np.save(fd / f"{k}.npy", arr)
            if a.csv_deep:
                df = pd.DataFrame(arr, columns=[f"d{i}" for i in range(arr.shape[1])])
                df.insert(0, "crop_id", idx.crop_id.values)
                df.to_csv(fd / f"{k}.csv", index=False)


# --------------------------------------------------------------------------- metrics
def stage_metrics(a):
    out = Path(a.out)
    idx = pd.read_csv(out / "index.csv")
    rng = np.random.default_rng(a.seed)
    pairs, skipped = [], 0
    for cls in ["ALL"] + sorted(idx.class_name.unique()):
        sel = idx if cls == "ALL" else idx[idx.class_name == cls]
        real = sel.index[sel.domain == "real"].to_numpy()
        half = rng.permutation(real)
        cands = [("REAL_vs_REAL_split", "Baseline", half[: len(half) // 2], half[len(half) // 2:])]
        cands += [(n, g_, real, d.index.to_numpy())
                  for (n, g_), d in sel[sel.domain == "synthetic"].groupby(["dataset", "group"])]
        for n, g_, xi, yi in cands:
            if min(len(xi), len(yi)) >= a.min_crops:
                pairs.append((cls, n, g_, xi, yi))
            else:
                skipped += 1
    print(f"{len(pairs)} comparisons ({skipped} skipped: fewer than {a.min_crops} crops on one side)")
    rows = []
    for f in sorted((out / "features").glob("*.npy")):
        if a.only and not any(o in f.stem for o in a.only):
            continue
        kind = "hash" if f.stem.startswith("hash_") else "deep"
        F = torch.tensor(np.load(f), dtype=torch.float32, device=a.device)
        for cls, name, group, xi, yi in tqdm(pairs, desc=f.stem):
            for metric, (mu, sd) in _compare(F, kind, xi, yi, a, rng).items():
                if metric == "fd" and min(len(xi), len(yi)) < a.fd_min_n:
                    continue  # covariance estimate dominated by sample size
                rows.append((f.stem, cls, name, group, len(xi), len(yi), metric, mu, sd, "subsample"))
    cols = ["feature", "class_name", "dataset", "group", "n_ref", "n_cand", "metric", "mean", "std", "std_kind"]
    df = pd.DataFrame(rows, columns=cols)
    cl = df[df.class_name != "ALL"]
    macro = (cl.groupby(["feature", "dataset", "group", "metric"])
               .agg(n_ref=("n_ref", "sum"), n_cand=("n_cand", "sum"), std=("mean", "std"), mean=("mean", "mean"))
               .reset_index())
    macro["class_name"], macro["std_kind"] = "MEAN_OVER_CLASSES", "across_classes"
    pd.concat([df, macro[cols]], ignore_index=True).to_csv(out / "metrics_long.csv", index=False)


# --------------------------------------------------------------------------- reduction
def stage_reduce(a):
    out = Path(a.out)
    rd = out / "reductions"
    rd.mkdir(exist_ok=True)
    idx = pd.read_csv(out / "index.csv")
    ids = sample_ids(idx, idx.dataset.unique(), a.reduce_per_set, a.seed)
    for f in sorted((out / "features").glob("*.npy")):
        if a.only and not any(o in f.stem for o in a.only):
            continue
        kind = "hash" if f.stem.startswith("hash_") else "deep"
        X = np.load(f, mmap_mode="r")[ids]
        for method in a.methods:
            for nc in a.dims:
                Z = reduce_features(X, kind, method, nc, a.pre_dims, seed=a.seed)
                df = pd.DataFrame(Z, columns=[f"c{i + 1}" for i in range(nc)])
                df.insert(0, "crop_id", ids)
                df.to_csv(rd / f"{f.stem}__{method}_{nc}d.csv", index=False)
                print("saved", f.stem, method, nc)


# --------------------------------------------------------------------------- class-mapping check
def _centroids(X, sel, min_n=5):
    out = {}
    for c, g in sel.groupby("class_name"):
        if len(g) >= min_n:
            m = X[g.index.values].mean(0)
            out[c] = m / np.linalg.norm(m)
    return out


def stage_verify(a):
    """Sanity check of the class mapping with DINOv2 CLS centroids: a synthetic class should be nearest
    to the real class of the same name. Perfect agreement is not expected (similar parts, e.g. O-ring
    sizes), so read every dataset against the REAL_split row (real half B vs. real half A)."""
    out = Path(a.out)
    idx = pd.read_csv(out / "index.csv")
    fs = sorted((out / "features").glob("dinov2_*_cls.npy"))
    if not fs:
        raise SystemExit("Run the features stage with the dinov2 encoder first.")
    X = np.load(fs[0])
    X = X / np.linalg.norm(X, axis=1, keepdims=True)
    rng = np.random.default_rng(a.seed)
    real = idx[idx.domain == "real"]
    perm = rng.permutation(real.index.values)
    ha, hb = perm[: len(perm) // 2], perm[len(perm) // 2:]
    jobs = [("REAL_split", idx.loc[hb], _centroids(X, idx.loc[ha]))]
    R = _centroids(X, real)
    jobs += [(n, g, R) for n, g in idx[idx.domain == "synthetic"].groupby("dataset")]
    rows = []
    for name, sel, ref in jobs:
        keys = list(ref)
        RM = np.stack([ref[k] for k in keys])
        for c, cen in _centroids(X, sel).items():
            if c not in ref:
                continue
            s = RM @ cen
            best = keys[int(s.argmax())]
            rows.append((name, c, best, best == c, float(s[keys.index(c)]), float(s.max())))
    df = pd.DataFrame(rows, columns=["dataset", "class_name", "nearest_real_class", "match", "cos_same", "cos_best"])
    df.to_csv(out / "class_mapping_check.csv", index=False)
    with pd.option_context("display.width", 250, "display.max_rows", 200):
        print("Top-1 centroid agreement per dataset (lowest first):")
        print(df.groupby("dataset").match.agg(["mean", "size"]).rename(columns={"mean": "agreement", "size": "n_classes"})
              .sort_values("agreement").to_string())
    print(f"Details: {out / 'class_mapping_check.csv'}")


# --------------------------------------------------------------------------- CLI
def main():
    p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
    p.add_argument("stage", choices=["index", "features", "metrics", "reduce", "verify", "all"])
    p.add_argument("--root", default="/media/rtx5090/IASSET")
    p.add_argument("--out", default=DEFAULT_OUT)
    p.add_argument("--seed", type=int, default=0)
    p.add_argument("--overwrite", action="store_true")
    p.add_argument("--only", nargs="*", help="restrict metrics/reduce to features containing these substrings")
    g = p.add_argument_group("crops")
    g.add_argument("--min-box-px", type=float, default=16, help="drop boxes whose shorter side is below this")
    g.add_argument("--pad", type=float, default=0.0, help="box margin as a fraction of box size (0 = strictly inside)")
    g.add_argument("--crop-mode", choices=["letterbox", "squash"], default="letterbox",
                   help="deep encoders: pad crop to a square (letterbox) or resize directly (squash)")
    g.add_argument("--max-per-class", type=int, default=1000,
                   help="cap crops per (dataset, class); 0 = all. Subsample metrics use at most min(n_real, "
                        "n_synthetic, --subset-size) crops, so large synthetic classes add cost, not information")
    g.add_argument("--max-images-per-set", type=int, default=0,
                   help="randomly keep at most this many images per dataset (0 = all); for quick test runs")
    g.add_argument("--skip-groups", nargs="*", default=["Training_Set_Size_Ablation"],
                   help="folders under Synthetic_Train_Sets to ignore")
    g.add_argument("--class-map", help='JSON {"dataset": {"src_id": "canonical class name"}} overriding name matching')
    g = p.add_argument_group("features")
    g.add_argument("--encoders", nargs="+", default=["hash", "dinov2", "inception"])
    g.add_argument("--dinov2-sizes", nargs="+", default=["base"], choices=["small", "base", "large", "giant"])
    g.add_argument("--batch-size", type=int, default=128)
    g.add_argument("--workers", type=int, default=8)
    g.add_argument("--device", default="cuda")
    g.add_argument("--fp16", action="store_true")
    g.add_argument("--csv-deep", action="store_true", help="also write deep features as CSV (large)")
    g = p.add_argument_group("metrics")
    g.add_argument("--subset-size", type=int, default=1000)
    g.add_argument("--repeats", type=int, default=10)
    g.add_argument("--k", type=int, default=5)
    g.add_argument("--min-crops", type=int, default=20, help="skip comparisons with fewer crops on either side")
    g.add_argument("--fd-min-n", type=int, default=100, help="report FD only if both sides have at least this many crops")
    g = p.add_argument_group("reduce")
    g.add_argument("--methods", nargs="+", default=["pca", "tsne", "umap"])
    g.add_argument("--dims", nargs="+", type=int, default=[2, 3])
    g.add_argument("--pre-dims", type=int, default=50)
    g.add_argument("--reduce-per-set", type=int, default=500)
    a = p.parse_args()
    a.crop = "none"  # crops are already square (letterbox) or deliberately squashed: no centre-crop
    if a.stage == "index":
        get_index(a)
    if a.stage == "verify":
        stage_verify(a)
    if a.stage in ("features", "all"):
        stage_features(a)
    if a.stage in ("metrics", "all"):
        stage_metrics(a)
    if a.stage in ("reduce", "all"):
        stage_reduce(a)


if __name__ == "__main__":
    main()

	#!/usr/bin/env python3
	"""Class-wise, background-free domain-gap analysis for IASSET.

	Every YOLO bounding box becomes one sample (the pixels under the box only).
	Metrics are computed per class (real crops of class c vs. synthetic crops of
	class c), pooled over all classes ("ALL"), and macro-averaged over classes
	("MEAN_OVER_CLASSES"). Class ids are mapped to canonical names taken from the
	real test set, so datasets with a different class order are handled by name.

	This file reuses encoders, metrics and reduction code from
	iasset_embeddings_analysis.py (keep both files in the same folder).

	Stages
	index : parse labels, map classes, write index.csv + class_stats.csv (run this first and check it)
	features : hashes + DINOv2 + Inception embeddings of the crops
	metrics : per-class similarity metrics (metrics_long.csv)
	reduce : offline PCA / t-SNE / UMAP projections
	verify : class-mapping sanity check (needs dinov2 features; not part of "all")
	all : features + metrics + reduce

	Example
	python iasset_embeddings_classwise.py index --root /media/rtx5090/IASSET
	python iasset_embeddings_classwise.py all --root /media/rtx5090/IASSET
	"""
	import argparse
	import json
	import os
	import re
	from concurrent.futures import ProcessPoolExecutor
	from pathlib import Path

	import numpy as np
	import pandas as pd
	import torch
	from PIL import Image
	from torch.utils.data import DataLoader, Dataset
	from tqdm import tqdm

	from iasset_embeddings_analysis import (HASH_SIZES, SKIP_DATASETS, THUMB, _compare, list_images,
	load_dinov2, load_inception, reduce_features, sample_ids)

	DEFAULT_OUT = "/media/rtx5090/IASSET/Scripts/embeddings_results_classwise"
	FILL = (124, 116, 104) # ImageNet mean colour, used for letterbox padding
	_norm = lambda s: re.sub(r"[^a-z0-9]", "", str(s).lower())


	# --------------------------------------------------------------------------- crops
	def get_crop(img, box, pad=0.0, mode="letterbox"):
	"""Pixels under a (x1, y1, x2, y2) box. letterbox: pad to a square with FILL; squash: raw crop."""
	x1, y1, x2, y2 = box
	dx, dy = pad * (x2 - x1), pad * (y2 - y1)
	W, H = img.size
	l, t = max(0, int(np.floor(x1 - dx))), max(0, int(np.floor(y1 - dy)))
	r, b = min(W, int(np.ceil(x2 + dx))), min(H, int(np.ceil(y2 + dy)))
	c = img.crop((l, t, max(r, l + 1), max(b, t + 1)))
	if mode == "letterbox":
	s = max(c.size)
	canvas = Image.new("RGB", (s, s), FILL)
	canvas.paste(c, ((s - c.width) // 2, (s - c.height) // 2))
	return canvas
	return c


	# --------------------------------------------------------------------------- labels and classes
	def _walk(root, max_depth=3, skip=("images", "depth", "rgb")):
	root = Path(root)
	base = len(root.parts)
	for dp, dn, fn in os.walk(root):
	depth = len(Path(dp).parts) - base
	dn[:] = sorted(d for d in dn if d.lower() not in skip) if depth < max_depth else []
	for f in sorted(fn):
	yield Path(dp) / f


	def read_names(root):
	"""Class names (list, index = id) from a YOLO yaml with 'names' or from classes.txt."""
	import yaml
	files = list(_walk(root))
	yamls = sorted((f for f in files if f.suffix.lower() in (".yaml", ".yml")),
	key=lambda f: f.name.lower() not in ("dataset.yaml", "data.yaml")) # these first
	for f in yamls:
	try:
	d = yaml.safe_load(f.read_text())
	except Exception:
	continue
	if isinstance(d, dict) and "names" in d:
	n = d["names"]
	if isinstance(n, dict):
	n = [n[k] for k in sorted(n, key=int)]
	return [str(x) for x in n], f
	for f in files:
	if f.name.lower() in ("classes.txt", "names.txt", "obj.names"):
	return [l.strip() for l in f.read_text().splitlines() if l.strip()], f
	return None, None


	def build_map(names, canon, override=None):
	"""source class id -> canonical class id (matched by normalized name)."""
	cn = {_norm(n): i for i, n in enumerate(canon)}
	if override:
	return {int(k): cn[_norm(v)] for k, v in override.items()}
	if names is None:
	return {i: i for i in range(len(canon))}
	return {sid: cn[_norm(n)] for sid, n in enumerate(names) if _norm(n) in cn}


	def parse_label(p):
	"""YOLO txt -> [(class, x1, y1, x2, y2)] normalized. Handles boxes and polygons."""
	out = []
	for line in Path(p).read_text().splitlines():
	t = line.split()
	if len(t) < 5:
	continue
	try:
	cid, v = int(float(t[0])), [float(x) for x in t[1:]]
	except ValueError:
	continue
	if len(v) == 4:
	cx, cy, w, h = v
	out.append((cid, cx - w / 2, cy - h / 2, cx + w / 2, cy + h / 2))
	else:
	xs, ys = v[0::2], v[1::2]
	out.append((cid, min(xs), min(ys), max(xs), max(ys)))
	return out


	class LabelFinder:
	"""Find the YOLO label file of an image: images->labels swap, sibling folders, then unique-stem lookup."""

	def __init__(self, ds_root):
	self.root, self.stems = Path(ds_root), None

	def __call__(self, img):
	parts = list(img.relative_to(self.root).with_suffix(".txt").parts)
	for i in range(len(parts) - 2, -1, -1):
	if parts[i].lower() == "images":
	c = self.root.joinpath(*(parts[:i] + ["labels"] + parts[i + 1:]))
	if c.exists():
	return c
	for c in (img.with_suffix(".txt"), img.parent.parent / "labels" / (img.stem + ".txt"),
	img.parent.parent / "labels" / img.parent.name / (img.stem + ".txt")):
	if c.exists():
	return c
	if self.stems is None:
	seen, self.stems = {}, {}
	for f in _walk(self.root, max_depth=5):
	if f.suffix == ".txt" and f.name.lower() not in ("classes.txt", "names.txt"):
	seen[f.stem] = None if f.stem in seen else f
	self.stems = {k: v for k, v in seen.items() if v is not None}
	return self.stems.get(img.stem)


	def build_crop_index(a):
	root = Path(a.root)
	real_root = root / "Real_Test_Set"
	sets = [("REAL", "Real", "real", real_root / "images" / "full" / "rgb", real_root)]
	synth = root / "Synthetic_Train_Sets"
	for g in sorted(p for p in synth.iterdir() if p.is_dir() and p.name not in a.skip_groups):
	for d in sorted(p for p in g.iterdir() if p.is_dir() and p.name not in SKIP_DATASETS):
	sets.append((f"{g.name}/{d.name}", g.name, "synthetic", d, d))
	canon, src = read_names(real_root)
	if canon is None:
	raise SystemExit(f"No class names found under {real_root} (dataset.yaml 'names' or classes.txt).")
	print(f"Canonical classes from {src}: {dict(enumerate(canon))}")
	override = json.load(open(a.class_map)) if a.class_map else {}
	rng = np.random.default_rng(a.seed)
	cols = ["dataset", "group", "domain", "path", "class_id", "class_name", "src_class_id",
	"x1", "y1", "x2", "y2", "img_w", "img_h"]
	frames = []
	for name, group, domain, img_dir, ds_root in sets:
	files = list_images(img_dir)
	if a.max_images_per_set and len(files) > a.max_images_per_set:
	files = [files[i] for i in sorted(rng.choice(len(files), a.max_images_per_set, replace=False))]
	names, nsrc = (canon, src) if domain == "real" else read_names(ds_root)
	if names is None and name not in override:
	print(f"[warn] {name}: no class names file, assuming the real-set class order")
	cmap = build_map(names, canon, override.get(name))
	if names is not None and [_norm(n) for n in names] != [_norm(n) for n in canon]:
	print(f"[info] {name}: class list differs from the real set; remapped by name "
	f"({len(cmap)}/{len(names)} matched)")
	find, rows, n_lab, unmatched = LabelFinder(ds_root), [], 0, set()
	for f in tqdm(files, desc=name[:60], leave=False):
	lp = find(f)
	boxes = parse_label(lp) if lp else []
	if not boxes:
	continue
	n_lab += 1
	with Image.open(f) as im:
	W, H = im.size
	for cid, x1, y1, x2, y2 in boxes:
	if cid not in cmap:
	unmatched.add(cid)
	continue
	px = [max(0.0, x1 * W), max(0.0, y1 * H), min(float(W), x2 * W), min(float(H), y2 * H)]
	if min(px[2] - px[0], px[3] - px[1]) < a.min_box_px:
	continue
	rows.append((name, group, domain, str(f), cmap[cid], canon[cmap[cid]], cid, *px, W, H))
	df = pd.DataFrame(rows, columns=cols)
	if a.max_per_class and len(df):
	keep = []
	for _, g in df.groupby("class_id"):
	ii = g.index.values
	keep += list(rng.choice(ii, a.max_per_class, replace=False) if len(ii) > a.max_per_class else ii)
	df = df.loc[sorted(keep)]
	print(f"{name:75s} images {len(files):6d} \| labelled {n_lab:6d} \| crops {len(df):6d}"
	+ (f" \| unmatched src class ids {sorted(unmatched)}" if unmatched else ""))
	frames.append(df)
	idx = pd.concat(frames, ignore_index=True)
	idx.insert(0, "crop_id", np.arange(len(idx)))
	return idx


	def get_index(a):
	out = Path(a.out)
	out.mkdir(parents=True, exist_ok=True)
	f = out / "index.csv"
	if f.exists() and not a.overwrite:
	return pd.read_csv(f)
	idx = build_crop_index(a)
	idx.to_csv(f, index=False)
	stats = idx.groupby(["dataset", "class_name"]).size().unstack(fill_value=0)
	stats.to_csv(out / "class_stats.csv")
	with pd.option_context("display.width", 250, "display.max_columns", 50, "display.max_rows", 200):
	print("\nCrops per dataset and class:\n", stats)
	return idx


	# --------------------------------------------------------------------------- hashes (raw crops)
	def _hash_img(img):
	import imagehash
	img.thumbnail((THUMB, THUMB), Image.LANCZOS)
	fn = {"ahash": imagehash.average_hash, "phash": imagehash.phash,
	"dhash": imagehash.dhash, "whash": imagehash.whash}
	out = {}
	for t, sizes in HASH_SIZES.items():
	for s in sizes:
	h = imagehash.colorhash(img, binbits=s) if t == "colorhash" else fn[t](img, hash_size=s)
	out[f"hash_{t}_{s}"] = h.hash.flatten().astype(np.uint8)
	return out


	def _hash_group(task):
	path, boxes, pad = task
	with Image.open(path) as im:
	img = im.convert("RGB")
	return [_hash_img(get_crop(img, b, pad, "squash")) for b in boxes]


	def run_hashes(idx, fd, a):
	groups = list(idx.groupby("path", sort=False))
	tasks = [(p, g[["x1", "y1", "x2", "y2"]].values.tolist(), a.pad) for p, g in groups]
	order = np.concatenate([g.index.values for _, g in groups])
	with ProcessPoolExecutor(a.workers) as ex:
	res = [r for rs in tqdm(ex.map(_hash_group, tasks, chunksize=8), total=len(tasks), desc="hashes")
	for r in rs]
	inv = np.argsort(order)
	hex_cols = {}
	for key in res[0]:
	arr = np.stack([r[key] for r in res])[inv] # back to index order
	np.save(fd / f"{key}.npy", arr)
	hex_cols[key] = [np.packbits(r).tobytes().hex() for r in arr]
	df = pd.DataFrame(hex_cols)
	df.insert(0, "crop_id", idx.crop_id.values)
	df.to_csv(fd.parent / "hashes.csv", index=False)


	# --------------------------------------------------------------------------- deep features (crops)
	class CropDataset(Dataset):
	def __init__(self, idx, tf, pad, mode):
	self.p = idx.path.values
	self.b = idx[["x1", "y1", "x2", "y2"]].values
	self.tf, self.pad, self.mode = tf, pad, mode
	self._p, self._img = None, None # per-worker cache: consecutive crops share one image

	def __len__(self):
	return len(self.p)

	def __getitem__(self, i):
	if self.p[i] != self._p:
	with Image.open(self.p[i]) as im:
	self._img = im.convert("RGB")
	self._p = self.p[i]
	return self.tf(get_crop(self._img, self.b[i], self.pad, self.mode))


	def extract_crops(idx, fwd, tf, a):
	dl = DataLoader(CropDataset(idx, tf, a.pad, a.crop_mode), batch_size=a.batch_size,
	num_workers=a.workers, pin_memory=a.device.startswith("cuda"),
	persistent_workers=a.workers > 0)
	acc = {}
	with torch.no_grad(), torch.autocast("cuda", dtype=torch.float16,
	enabled=(a.fp16 and a.device.startswith("cuda"))):
	for x in tqdm(dl, desc="encode crops"):
	for k, v in fwd(x.to(a.device, non_blocking=True)).items():
	acc.setdefault(k, []).append(v.float().cpu())
	return {k: torch.cat(v).numpy() for k, v in acc.items()}


	def stage_features(a):
	idx = get_index(a)
	fd = Path(a.out) / "features"
	fd.mkdir(exist_ok=True)
	if "hash" in a.encoders:
	keys = [f"hash_{t}_{s}" for t, ss in HASH_SIZES.items() for s in ss]
	if a.overwrite or not all((fd / f"{k}.npy").exists() for k in keys):
	run_hashes(idx, fd, a)
	jobs = []
	if "dinov2" in a.encoders:
	for size in a.dinov2_sizes:
	jobs.append(([f"dinov2_{size}_cls", f"dinov2_{size}_patchmean"], lambda s=size: load_dinov2(s, a)))
	if "inception" in a.encoders:
	jobs.append((["inception_pool3"], lambda: load_inception(a)))
	for names, loader in jobs:
	if not a.overwrite and all((fd / f"{n}.npy").exists() for n in names):
	continue
	fwd, tf = loader()
	for k, arr in extract_crops(idx, fwd, tf, a).items():
	np.save(fd / f"{k}.npy", arr)
	if a.csv_deep:
	df = pd.DataFrame(arr, columns=[f"d{i}" for i in range(arr.shape[1])])
	df.insert(0, "crop_id", idx.crop_id.values)
	df.to_csv(fd / f"{k}.csv", index=False)


	# --------------------------------------------------------------------------- metrics
	def stage_metrics(a):
	out = Path(a.out)
	idx = pd.read_csv(out / "index.csv")
	rng = np.random.default_rng(a.seed)
	pairs, skipped = [], 0
	for cls in ["ALL"] + sorted(idx.class_name.unique()):
	sel = idx if cls == "ALL" else idx[idx.class_name == cls]
	real = sel.index[sel.domain == "real"].to_numpy()
	half = rng.permutation(real)
	cands = [("REAL_vs_REAL_split", "Baseline", half[: len(half) // 2], half[len(half) // 2:])]
	cands += [(n, g_, real, d.index.to_numpy())
	for (n, g_), d in sel[sel.domain == "synthetic"].groupby(["dataset", "group"])]
	for n, g_, xi, yi in cands:
	if min(len(xi), len(yi)) >= a.min_crops:
	pairs.append((cls, n, g_, xi, yi))
	else:
	skipped += 1
	print(f"{len(pairs)} comparisons ({skipped} skipped: fewer than {a.min_crops} crops on one side)")
	rows = []
	for f in sorted((out / "features").glob("*.npy")):
	if a.only and not any(o in f.stem for o in a.only):
	continue
	kind = "hash" if f.stem.startswith("hash_") else "deep"
	F = torch.tensor(np.load(f), dtype=torch.float32, device=a.device)
	for cls, name, group, xi, yi in tqdm(pairs, desc=f.stem):
	for metric, (mu, sd) in _compare(F, kind, xi, yi, a, rng).items():
	if metric == "fd" and min(len(xi), len(yi)) < a.fd_min_n:
	continue # covariance estimate dominated by sample size
	rows.append((f.stem, cls, name, group, len(xi), len(yi), metric, mu, sd, "subsample"))
	cols = ["feature", "class_name", "dataset", "group", "n_ref", "n_cand", "metric", "mean", "std", "std_kind"]
	df = pd.DataFrame(rows, columns=cols)
	cl = df[df.class_name != "ALL"]
	macro = (cl.groupby(["feature", "dataset", "group", "metric"])
	.agg(n_ref=("n_ref", "sum"), n_cand=("n_cand", "sum"), std=("mean", "std"), mean=("mean", "mean"))
	.reset_index())
	macro["class_name"], macro["std_kind"] = "MEAN_OVER_CLASSES", "across_classes"
	pd.concat([df, macro[cols]], ignore_index=True).to_csv(out / "metrics_long.csv", index=False)


	# --------------------------------------------------------------------------- reduction
	def stage_reduce(a):
	out = Path(a.out)
	rd = out / "reductions"
	rd.mkdir(exist_ok=True)
	idx = pd.read_csv(out / "index.csv")
	ids = sample_ids(idx, idx.dataset.unique(), a.reduce_per_set, a.seed)
	for f in sorted((out / "features").glob("*.npy")):
	if a.only and not any(o in f.stem for o in a.only):
	continue
	kind = "hash" if f.stem.startswith("hash_") else "deep"
	X = np.load(f, mmap_mode="r")[ids]
	for method in a.methods:
	for nc in a.dims:
	Z = reduce_features(X, kind, method, nc, a.pre_dims, seed=a.seed)
	df = pd.DataFrame(Z, columns=[f"c{i + 1}" for i in range(nc)])
	df.insert(0, "crop_id", ids)
	df.to_csv(rd / f"{f.stem}__{method}_{nc}d.csv", index=False)
	print("saved", f.stem, method, nc)


	# --------------------------------------------------------------------------- class-mapping check
	def _centroids(X, sel, min_n=5):
	out = {}
	for c, g in sel.groupby("class_name"):
	if len(g) >= min_n:
	m = X[g.index.values].mean(0)
	out[c] = m / np.linalg.norm(m)
	return out


	def stage_verify(a):
	"""Sanity check of the class mapping with DINOv2 CLS centroids: a synthetic class should be nearest
	to the real class of the same name. Perfect agreement is not expected (similar parts, e.g. O-ring
	sizes), so read every dataset against the REAL_split row (real half B vs. real half A)."""
	out = Path(a.out)
	idx = pd.read_csv(out / "index.csv")
	fs = sorted((out / "features").glob("dinov2_*_cls.npy"))
	if not fs:
	raise SystemExit("Run the features stage with the dinov2 encoder first.")
	X = np.load(fs[0])
	X = X / np.linalg.norm(X, axis=1, keepdims=True)
	rng = np.random.default_rng(a.seed)
	real = idx[idx.domain == "real"]
	perm = rng.permutation(real.index.values)
	ha, hb = perm[: len(perm) // 2], perm[len(perm) // 2:]
	jobs = [("REAL_split", idx.loc[hb], _centroids(X, idx.loc[ha]))]
	R = _centroids(X, real)
	jobs += [(n, g, R) for n, g in idx[idx.domain == "synthetic"].groupby("dataset")]
	rows = []
	for name, sel, ref in jobs:
	keys = list(ref)
	RM = np.stack([ref[k] for k in keys])
	for c, cen in _centroids(X, sel).items():
	if c not in ref:
	continue
	s = RM @ cen
	best = keys[int(s.argmax())]
	rows.append((name, c, best, best == c, float(s[keys.index(c)]), float(s.max())))
	df = pd.DataFrame(rows, columns=["dataset", "class_name", "nearest_real_class", "match", "cos_same", "cos_best"])
	df.to_csv(out / "class_mapping_check.csv", index=False)
	with pd.option_context("display.width", 250, "display.max_rows", 200):
	print("Top-1 centroid agreement per dataset (lowest first):")
	print(df.groupby("dataset").match.agg(["mean", "size"]).rename(columns={"mean": "agreement", "size": "n_classes"})
	.sort_values("agreement").to_string())
	print(f"Details: {out / 'class_mapping_check.csv'}")


	# --------------------------------------------------------------------------- CLI
	def main():
	p = argparse.ArgumentParser(description=__doc__, formatter_class=argparse.RawDescriptionHelpFormatter)
	p.add_argument("stage", choices=["index", "features", "metrics", "reduce", "verify", "all"])
	p.add_argument("--root", default="/media/rtx5090/IASSET")
	p.add_argument("--out", default=DEFAULT_OUT)
	p.add_argument("--seed", type=int, default=0)
	p.add_argument("--overwrite", action="store_true")
	p.add_argument("--only", nargs="*", help="restrict metrics/reduce to features containing these substrings")
	g = p.add_argument_group("crops")
	g.add_argument("--min-box-px", type=float, default=16, help="drop boxes whose shorter side is below this")
	g.add_argument("--pad", type=float, default=0.0, help="box margin as a fraction of box size (0 = strictly inside)")
	g.add_argument("--crop-mode", choices=["letterbox", "squash"], default="letterbox",
	help="deep encoders: pad crop to a square (letterbox) or resize directly (squash)")
	g.add_argument("--max-per-class", type=int, default=1000,
	help="cap crops per (dataset, class); 0 = all. Subsample metrics use at most min(n_real, "
	"n_synthetic, --subset-size) crops, so large synthetic classes add cost, not information")
	g.add_argument("--max-images-per-set", type=int, default=0,
	help="randomly keep at most this many images per dataset (0 = all); for quick test runs")
	g.add_argument("--skip-groups", nargs="*", default=["Training_Set_Size_Ablation"],
	help="folders under Synthetic_Train_Sets to ignore")
	g.add_argument("--class-map", help='JSON {"dataset": {"src_id": "canonical class name"}} overriding name matching')
	g = p.add_argument_group("features")
	g.add_argument("--encoders", nargs="+", default=["hash", "dinov2", "inception"])
	g.add_argument("--dinov2-sizes", nargs="+", default=["base"], choices=["small", "base", "large", "giant"])
	g.add_argument("--batch-size", type=int, default=128)
	g.add_argument("--workers", type=int, default=8)
	g.add_argument("--device", default="cuda")
	g.add_argument("--fp16", action="store_true")
	g.add_argument("--csv-deep", action="store_true", help="also write deep features as CSV (large)")
	g = p.add_argument_group("metrics")
	g.add_argument("--subset-size", type=int, default=1000)
	g.add_argument("--repeats", type=int, default=10)
	g.add_argument("--k", type=int, default=5)
	g.add_argument("--min-crops", type=int, default=20, help="skip comparisons with fewer crops on either side")
	g.add_argument("--fd-min-n", type=int, default=100, help="report FD only if both sides have at least this many crops")
	g = p.add_argument_group("reduce")
	g.add_argument("--methods", nargs="+", default=["pca", "tsne", "umap"])
	g.add_argument("--dims", nargs="+", type=int, default=[2, 3])
	g.add_argument("--pre-dims", type=int, default=50)
	g.add_argument("--reduce-per-set", type=int, default=500)
	a = p.parse_args()
	a.crop = "none" # crops are already square (letterbox) or deliberately squashed: no centre-crop
	if a.stage == "index":
	get_index(a)
	if a.stage == "verify":
	stage_verify(a)
	if a.stage in ("features", "all"):
	stage_features(a)
	if a.stage in ("metrics", "all"):
	stage_metrics(a)
	if a.stage in ("reduce", "all"):
	stage_reduce(a)


	if __name__ == "__main__":
	main()