Download scripts/diversity_analysis.py from vineet-datasets/probe-benchmark-600: direct link, hf CLI and curl.
- Browser
- Download file 4.59 kB
-
https://huggingface.co/datasets/vineet-datasets/probe-benchmark-600/resolve/main/scripts/diversity_analysis.py
- Command line
-
hf download hf://datasets/vineet-datasets/probe-benchmark-600/scripts/diversity_analysis.py
-
curl -L -o diversity_analysis.py https://huggingface.co/datasets/vineet-datasets/probe-benchmark-600/resolve/main/scripts/diversity_analysis.py
4.59 kB
| #!/usr/bin/env python3 | |
| """Object-diversity analysis of the 783 accepted-question pool. | |
| Per-type diversity SIGNATURE (the object(s) a question is fundamentally about): | |
| count -> singleton: target_ref_folder (the counted object) | |
| find -> singleton: queried_ref_folder (the searched object) | |
| beneath -> duplet: (cover_folder, buried_target_folder) | |
| compare -> triplet: frozenset(3 marked obj_folders) | |
| Reports how concentrated the pool is on repeated signatures, and the union of | |
| unique object folders — the raw material for a diversity-maximizing selection. | |
| """ | |
| import json, glob, os | |
| from collections import Counter, defaultdict | |
| BANK="approved-questions/bank" | |
| def snap_folder_map(ep_dir): | |
| """name -> folder from the snapshot objects (authoritative folder ids).""" | |
| sp=os.path.join(ep_dir,"scene_snapshot","scene_snapshot.json") | |
| m={} | |
| if os.path.isfile(sp): | |
| s=json.load(open(sp)) | |
| for o in (s.get("objects") or []): | |
| if o.get("name") and o.get("folder"): | |
| m[o["name"]]=o["folder"] | |
| return m | |
| def episodes(ct): | |
| for rj in sorted(glob.glob(f"{BANK}/ep_*_{ct}/review.json")): | |
| yield os.path.dirname(rj), json.load(open(rj)) | |
| # ── build signatures ────────────────────────────────────────────────── | |
| sig = defaultdict(dict) # ct -> {ep_name: signature} | |
| all_folders_by_type = defaultdict(set) | |
| all_folders = set() | |
| # count / find : singleton folder | |
| for ct, key in [("count","target_ref_folder"), ("find","queried_ref_folder")]: | |
| for ep_dir, d in episodes(ct): | |
| q=d.get("question") or {} | |
| fol=q.get(key) | |
| sig[ct][os.path.basename(ep_dir)]=fol | |
| if fol: all_folders_by_type[ct].add(fol); all_folders.add(fol) | |
| # beneath : (cover_folder, buried_target_folder) | |
| for ep_dir, d in episodes("beneath"): | |
| q=d.get("question") or {} | |
| fmap=snap_folder_map(ep_dir) | |
| cover=fmap.get(q.get("cover_object")) | |
| tgt=fmap.get(q.get("target_name")) | |
| sig["beneath"][os.path.basename(ep_dir)]=(cover,tgt) | |
| for f in (cover,tgt): | |
| if f: all_folders_by_type["beneath"].add(f); all_folders.add(f) | |
| # compare : triplet of marked folders | |
| for ep_dir, d in episodes("compare"): | |
| q=d.get("question") or {} | |
| ml=q.get("marked_labels") or {} | |
| folders=tuple(sorted(v.get("obj_folder") for v in ml.values() if v.get("obj_folder"))) | |
| sig["compare"][os.path.basename(ep_dir)]=folders | |
| for f in folders: | |
| if f: all_folders_by_type["compare"].add(f); all_folders.add(f) | |
| # ── report ──────────────────────────────────────────────────────────── | |
| print("="*72) | |
| print("PER-TYPE SIGNATURE REUSE (how concentrated is the pool?)") | |
| print("="*72) | |
| for ct in ["count","find","beneath","compare"]: | |
| sigs=list(sig[ct].values()) | |
| n=len(sigs) | |
| c=Counter(s for s in sigs if s not in (None,(None,None),())) | |
| n_unique=len(c) | |
| # histogram: how many signatures appear k times | |
| hist=Counter(c.values()) | |
| max_reuse=max(c.values()) if c else 0 | |
| # how many episodes are "unique-signature" (appear exactly once) | |
| n_once=sum(v for k,v in hist.items() if k==1) | |
| print(f"\n{ct.upper()} ({n} episodes)") | |
| print(f" distinct signatures : {n_unique}") | |
| print(f" episodes w/ unique sig : {n_once} ({100*n_once//max(n,1)}%)") | |
| print(f" max reuse of one signature: {max_reuse}") | |
| print(f" reuse histogram (k:sigs) : {dict(sorted(hist.items()))}") | |
| # top offenders | |
| top=c.most_common(5) | |
| def short(s): | |
| if isinstance(s,tuple): return " + ".join(os.path.basename(x or "?") for x in s) | |
| return os.path.basename(s or "?") | |
| print(f" most-reused signatures :") | |
| for s,k in top: | |
| if k>1: print(f" {k:>2}x {short(s)}") | |
| print("\n"+"="*72) | |
| print("UNIQUE OBJECT FOLDERS") | |
| print("="*72) | |
| for ct in ["count","find","beneath","compare"]: | |
| print(f" {ct:9s}: {len(all_folders_by_type[ct])} unique object folders") | |
| print(f" {'UNION':9s}: {len(all_folders)} unique object folders across all types") | |
| # cross-type sharing | |
| print("\n"+"="*72) | |
| print("OBJECT SHARING ACROSS TYPES (folders used in N types)") | |
| print("="*72) | |
| fol_types=defaultdict(set) | |
| for ct in ["count","find","beneath","compare"]: | |
| for f in all_folders_by_type[ct]: fol_types[f].add(ct) | |
| byn=Counter(len(v) for v in fol_types.values()) | |
| print(" #folders used in exactly N types:", dict(sorted(byn.items()))) | |