#!/usr/bin/env python3 """Full exact-combination atlas.""" import argparse from collections import defaultdict from pathlib import Path import matplotlib matplotlib.use('Agg') import matplotlib.pyplot as plt from matplotlib.colors import ListedColormap, Normalize, TwoSlopeNorm import numpy as np import pandas as pd from plot_difficulty import load_annotations, load_scores FORBIDDEN={'box','permanent_mortise','nuisance_dent'} def find_coco(entries, pattern): out = [] for item in entries: p = Path(item) out.extend(p.rglob(pattern) if p.is_dir() else [p]) return sorted(set(str(p) for p in out if p.is_file())) def groups_from(memberships, min_count): g = defaultdict(list) for key, types in memberships.items(): combo = tuple(sorted(set(types) - FORBIDDEN)) if combo: g[combo].append(key) return {k: v for k, v in g.items() if len(v) >= min_count} def summarize(groups, m3dm, stfpm): diff = {k: m3dm[k] - stfpm[k] for k in set(m3dm) & set(stfpm)} rows = [] for combo, keys in groups.items(): def stat(scores): x = np.array([scores[k] for k in keys if k in scores], float) return (float(x.mean()) if len(x) else np.nan, float(x.std(ddof=1)) if len(x) > 1 else 0., len(x)) mm, ms, mn = stat(m3dm) sm, ss, sn = stat(stfpm) dm, ds, dn = stat(diff) rows.append(dict(combo=combo, n_types=len(combo), dataset_n=len(keys), m3dm_mean=mm, m3dm_std=ms, m3dm_n=mn, stfpm_mean=sm, stfpm_std=ss, stfpm_n=sn, diff_mean=dm, diff_std=ds, paired_n=dn)) rows.sort(key=lambda r: (r['n_types'], -r['dataset_n'], r['combo'])) return rows def short(s): return s.replace('nuisance_', '').replace('_', ' ') def save_csv(path, rows, categories): out = [] for rank, r in enumerate(rows, 1): d = {'rank': rank, 'combination': ' + '.join(r['combo']), 'n_types': r['n_types'], 'dataset_n': r['dataset_n'], 'm3dm_mean': r['m3dm_mean'], 'm3dm_std': r['m3dm_std'], 'm3dm_n': r['m3dm_n'], 'stfpm_mean': r['stfpm_mean'], 'stfpm_std': r['stfpm_std'], 'stfpm_n': r['stfpm_n'], 'm3dm_minus_stfpm': r['diff_mean'], 'difference_std': r['diff_std'], 'paired_n': r['paired_n']} d.update({f'has_{c}': int(c in r['combo']) for c in categories}) out.append(d) pd.DataFrame(out).to_csv(path, index=False) def draw(rows, categories, metric, stem): nr, nc = len(rows), len(categories) y = np.arange(nr) mem = np.array([[c in r['combo'] for c in categories] for r in rows], int) counts = np.array([r['dataset_n'] for r in rows]) m = np.array([r['m3dm_mean'] for r in rows]) s = np.array([r['stfpm_mean'] for r in rows]) d = np.array([r['diff_mean'] for r in rows]) pn = np.array([r['paired_n'] for r in rows]) absolute = np.r_[m[np.isfinite(m)], s[np.isfinite(s)]] anorm = Normalize(absolute.min(), absolute.max()) finite = d[np.isfinite(d)] bound = max(np.abs(finite).max(), 1e-9) dnorm = TwoSlopeNorm(vmin=-bound, vcenter=0, vmax=bound) fig = plt.figure(figsize=(11.7, max(12, 0.25 * nr + 2.8))) gs = fig.add_gridspec(1, 6, width_ratios=[4.8, 2.4, 1.15, 1.15, 1.3, .75], wspace=.07) axm = fig.add_subplot(gs[0]) axc = fig.add_subplot(gs[1], sharey=axm) ax1 = fig.add_subplot(gs[2], sharey=axm) ax2 = fig.add_subplot(gs[3], sharey=axm) axd = fig.add_subplot(gs[4], sharey=axm) axn = fig.add_subplot(gs[5], sharey=axm) axm.imshow(mem, aspect='auto', cmap=ListedColormap(['#f2f2f2', '#16324f']), vmin=0, vmax=1) axm.set_xticks(range(nc)) axm.set_xticklabels([short(c) for c in categories], rotation=55, ha='right', fontsize=7) axm.set_yticks(y) axm.set_yticklabels(range(1, nr + 1), fontsize=5.5) axm.set_ylabel('exact combination rank') axm.set_title('Exact anomaly membership', fontsize=9) axm.set_xticks(np.arange(-.5, nc, 1), minor=True) axm.set_yticks(np.arange(-.5, nr, 1), minor=True) axm.grid(which='minor', color='white', lw=.45) axm.tick_params(which='minor', length=0) axc.barh(y, counts, height=.72, color='#6c879f') axc.invert_yaxis() axc.set_title('Images', fontsize=9) axc.set_xlabel('count') axc.tick_params(axis='y', left=False, labelleft=False) axc.spines[['top', 'right']].set_visible(False) axc.set_xlim(0, counts.max() * 1.2) for i, v in enumerate(counts): axc.text(v, i, f' {v}', va='center', fontsize=5.2) def col(ax, vals, title, norm): cmap = plt.cm.RdYlBu.copy() cmap.set_bad('#e8e8e8') im = ax.imshow(np.ma.masked_invalid(vals[:, None]), aspect='auto', cmap=cmap, norm=norm) ax.set_xticks([0]) ax.set_xticklabels([title], rotation=55, ha='right', fontsize=7) ax.set_title(title, fontsize=9) ax.tick_params(axis='y', left=False, labelleft=False) for i, v in enumerate(vals): ax.text(0, i, f'{v:.2f}' if np.isfinite(v) else 'NA', ha='center', va='center', fontsize=5.1) return im im1 = col(ax1, m, 'M3DM', anorm) col(ax2, s, 'STFPM', anorm) imd = col(axd, d, 'M3DM - STFPM', dnorm) axn.imshow(pn[:, None], aspect='auto', cmap='Greys', vmin=0, vmax=max(pn.max(), 1)) axn.set_xticks([0]) axn.set_xticklabels(['paired n'], rotation=55, ha='right', fontsize=7) axn.set_title('Paired n', fontsize=9) axn.tick_params(axis='y', left=False, labelleft=False) for i, v in enumerate(pn): axn.text(0, i, str(v), ha='center', va='center', fontsize=5.1) starts, ends = {}, {} for i, r in enumerate(rows): starts.setdefault(r['n_types'], i) ends[r['n_types']] = i for k in sorted(starts): start = starts[k] if start: for ax in [axm, axc, ax1, ax2, axd, axn]: ax.axhline(start - .5, color='#b91c1c', lw=.8) axm.text(-1.05, (start + ends[k]) / 2, f'{k} type' if k == 1 else f'{k} types', ha='right', va='center', fontsize=6.2, fontweight='bold', color='#7f1d1d', clip_on=False) cb = fig.colorbar(im1, ax=[ax1, ax2], fraction=.035, pad=.02) cb.set_label(f'mean {metric}', fontsize=7) cb = fig.colorbar(imd, ax=axd, fraction=.09, pad=.04) cb.set_label(f'delta {metric}', fontsize=7) fig.suptitle('Full exact anomaly-combination atlas: prevalence and localization difficulty', fontsize=12, y=.998) fig.text(.5, .003, 'Rows grouped by multiplicity and sorted by frequency; positive difference means M3DM performs better.', ha='center', fontsize=6.5, style='italic', color='.35') fig.subplots_adjust(left=.07, right=.985, bottom=.035, top=.975) for ext in ['pdf', 'png']: fig.savefig(f'{stem}.{ext}', dpi=400, bbox_inches='tight') plt.close(fig) def default_coco_search(): candidates = [ Path("test"), Path("../test"), Path(__file__).resolve().parent.parent / "test", Path("train"), Path("../train"), Path(__file__).resolve().parent.parent / "train", Path("."), Path(".."), Path(__file__).resolve().parent.parent, ] for c in candidates: if c.is_dir() and list(c.rglob("annotations*.coco.json")): return [str(c)] return ["."] def find_files(names): resolved = [] for name in names: candidates = [ Path(f"checkpoint_data/{name}"), Path(f"../checkpoint_data/{name}"), Path(__file__).resolve().parent.parent / "checkpoint_data" / name, Path(__file__).resolve().parent / name, Path(name), ] found = next((str(c) for c in candidates if c.is_file()), None) if found: resolved.append(found) return resolved def default_out_dir(): if Path("checkpoint_data").is_dir(): return Path("checkpoint_data/combination_atlas") if (Path(__file__).resolve().parent.parent / "checkpoint_data").is_dir(): return Path(__file__).resolve().parent.parent / "checkpoint_data/combination_atlas" return Path("./combination_atlas") def main(): p = argparse.ArgumentParser(description="Full exact-combination atlas") p.add_argument('--coco', nargs='+', default=None, help="COCO annotation files or directory (auto-detected if omitted)") p.add_argument('--coco-pattern', default='annotations*.coco.json') p.add_argument('--m3dm', nargs='+', default=None, help="M3DM per-image CSVs (auto-detected from checkpoint_data/ if omitted)") p.add_argument('--stfpm', nargs='+', default=None, help="STFPM per-image CSVs (auto-detected from checkpoint_data/ if omitted)") p.add_argument('--metric', default='aupro_30') p.add_argument('--min-count', type=int, default=1) p.add_argument('--out', default=None, help="Output directory (default: checkpoint_data/combination_atlas or ./combination_atlas)") a = p.parse_args() coco_entries = a.coco if a.coco is not None else default_coco_search() coco = find_coco(coco_entries, a.coco_pattern) if not coco: raise SystemExit(f"No COCO annotation files found in {coco_entries}") m3dm_paths = a.m3dm if a.m3dm is not None else find_files(["per_image_aupro_zivid.csv", "per_image_aupro_realsense.csv"]) stfpm_paths = a.stfpm if a.stfpm is not None else find_files(["per_image_aupro_stfpm.csv"]) memberships = load_annotations(coco) memberships = {k: set(v) - FORBIDDEN for k, v in memberships.items()} groups = groups_from(memberships, a.min_count) m3dm = load_scores(m3dm_paths, a.metric, 'M3DM') if m3dm_paths else {} stfpm = load_scores(stfpm_paths, a.metric, 'STFPM') if stfpm_paths else {} rows = summarize(groups, m3dm, stfpm) cats = sorted({c for r in rows for c in r['combo']}) out = Path(a.out) if a.out is not None else default_out_dir() out.mkdir(parents=True, exist_ok=True) stem = out / 'combination_atlas_all' ann = {k for k, v in memberships.items() if v} print(f'COCO files: {len(coco)} | combinations: {len(rows)} | types: {len(cats)}') print(f'M3DM matched: {len(ann & set(m3dm))}/{len(m3dm)} | STFPM matched: {len(ann & set(stfpm))}/{len(stfpm)} | paired: {len(ann & set(m3dm) & set(stfpm))}') save_csv(f'{stem}_values.csv', rows, cats) draw(rows, cats, a.metric, stem) print(f'Wrote {stem}.pdf, .png, and _values.csv') if __name__ == '__main__': main()