"""Retriever tuning with a clean validation set, then ONE look at the 1,600 test queries. Validation queries: training-fold images that are in the seed-0 K=16 memory of a crop and fold but not in its K=4 memory and are never test queries. Each is scored against the K=4 memory of the same crop and fold (the memory MACE uses), so no test image or test label is touched while tuning. Validation images come from the memory collections, so tuning can only target the IID setting; held-out results are reported as they come. Search space (every config scored the same way, top 4 kept, ties by memory id): text BM25 over the whole caption with parameters k1, b and the organ repeat count (paper: 1.2, 0.75, 3) image DINOv2 cosine between test and memory image, min-max scaled within the memory, added with weight alpha cap at most `cap` entries per disease among the four (4 = no cap) Objective on validation: hit@4 (equal-crop mean), then mean number of true-disease entries among the four. python -m experiments.retriever_pilot.tune """ from __future__ import annotations import itertools import json import math from collections import Counter, defaultdict import numpy as np from agmem import packet from mace10 import jobs as JB from . import pilot as P OUT = P.OUT K1S, BS, ORGS = (0.6, 1.2, 2.0), (0.3, 0.75, 1.0), (0, 1, 3, 6) ALPHAS = (0, 0.25, 0.5, 1, 1.5, 2, 3, 5, 1e3) # 1e3 = image similarity alone (text breaks ties) CAPS = (4, 3, 2, 1) def bm25(q, docs, k1, b): N = len(docs); df = Counter(t for d in docs for t in set(d)); avg = sum(map(len, docs)) / N; qc = Counter(q); out = [] for d in docs: tf = Counter(d); s = 0.0 for t in qc: if t in tf: s += math.log(1 + (N - df[t] + 0.5) / (df[t] + 0.5)) * tf[t] * (k1 + 1) / (tf[t] + k1 * (1 - b + b * len(d) / avg)) out.append(s) m = max(out) or 1.0 return np.array(out) / m def items(P1, caps, E, idx): """Validation and test items: (crop, split, query id, true label, memory ids, memory labels).""" test = [(j["crop"], j["split"], j["image_id"], j["label"], [m["image_id"] for m in j["memory"]], [m["label"] for m in j["memory"]]) for j in P1] tq = set(JB.queries().image_id); val = [] for c, r in sorted({(j["crop"], j["run"]) for j in P1}): m4 = JB.mem(c, 0, r, 4); ids4 = {m["image_id"] for m in m4} for m in JB.mem(c, 0, r, 16): i = m["image_id"] if i not in ids4 and i not in tq and i in caps and i in idx: val.append((c, "val", i, m["label"], [x["image_id"] for x in m4], [x["label"] for x in m4])) return val, test def precompute(its, caps, E, idx): """Per item: text scores for every (k1, b, organ repeat) and the min-max scaled image similarity.""" toks = {} def tk(i): if i not in toks: toks[i] = packet.tok(packet.caption_text(caps[i])) return toks[i] pre = [] for c, sp, q, lab, mids, mlabs in its: docs = [tk(m) for m in mids]; organ = packet.tok(str(caps[q].get("organ", ""))) T = {(k1, b, o): bm25(tk(q) + organ * o, docs, k1, b) for k1 in K1S for b in BS for o in ORGS} cos = E[[idx[m] for m in mids]] @ E[idx[q]]; img = (cos - cos.min()) / ((cos.max() - cos.min()) or 1.0) pre.append((T, img)) return pre def pick(score, mids, mlabs, cap): order = sorted(range(len(mids)), key=lambda i: (-score[i], "m_" + mids[i])); out, n = [], Counter() for i in order: if n[mlabs[i]] < cap: out.append(i); n[mlabs[i]] += 1 if len(out) == 4: break return out def evaluate(its, pre, cfg): (k1, b, o), a, cap = cfg; per = defaultdict(lambda: defaultdict(list)) for (c, sp, q, lab, mids, mlabs), (T, img) in zip(its, pre): sel = pick(T[(k1, b, o)] + a * img, mids, mlabs, cap); labs = [mlabs[i] for i in sel] cnt = Counter(labs); best = max(cnt.values()); vote = next(l for l in labs if cnt[l] == best) r = per[(sp, c)]; r["hit4"].append(lab in labs); r["ntrue"].append(labs.count(lab)); r["vote"].append(vote == lab); r["top1"].append(labs[0] == lab); r["q"].append(q) return per def summary(per, sp): crops = sorted(c for s, c in per if s == sp) return {m: 100 * float(np.mean([np.mean(per[(sp, c)][m]) for c in crops])) if m != "ntrue" else float(np.mean([np.mean(per[(sp, c)][m]) for c in crops])) for m in ("hit4", "ntrue", "vote", "top1")} def paired_ci(pa, pb, sp, m="hit4", n=2000): rng = np.random.default_rng(0); crops = sorted(c for s, c in pa if s == sp); ds = [] for _ in range(n): v = [] for c in crops: a, b = np.array(pa[(sp, c)][m], float), np.array(pb[(sp, c)][m], float); k = rng.integers(0, len(a), len(a)); v.append((a[k] - b[k]).mean()) ds.append(100 * np.mean(v)) return [round(float(np.percentile(ds, 2.5)), 1), round(float(np.percentile(ds, 97.5)), 1)] def main(): caps, P1, _ = P.load() d = np.load(P.M10 / "dinov2_embeddings.npz"); E = d["E"].astype(np.float32); E /= np.linalg.norm(E, axis=1, keepdims=True); idx = {k: i for i, k in enumerate(d["ids"])} val, test = items(P1, caps, E, idx) print(f"validation queries {len(val):,} ({len({v[0] for v in val})} crops); test queries {len(test):,}") pv = precompute(val, caps, E, idx) grid = list(itertools.product([(k1, b, o) for k1 in K1S for b in BS for o in ORGS], ALPHAS, CAPS)) rows = [] for cfg in grid: s = summary(evaluate(val, pv, cfg), "val"); rows.append((cfg, s)) rows.sort(key=lambda r: (-round(r[1]["hit4"], 6), -r[1]["ntrue"])) base = ((1.2, 0.75, 3), 0, 4) sb = next(s for cfg, s in rows if cfg == base) print(f"validation: paper BM25 hit@4 {sb['hit4']:.1f}, true entries {sb['ntrue']:.2f}; {len(grid)} configs") for cfg, s in rows[:8]: print(f" {cfg} hit@4 {s['hit4']:.1f} true entries {s['ntrue']:.2f} vote {s['vote']:.1f}") best = rows[0][0] # best without the per-disease cap, so we can see what the image similarity alone adds best_nocap = next(cfg for cfg, s in rows if cfg[2] == 4) img_only = ((1.2, 0.75, 3), 1e3, 4) pt = precompute(test, caps, E, idx) named = {"BM25 (paper)": base, "Image only (DINOv2)": img_only, "Tuned, no cap": best_nocap, "Tuned (best on validation)": best} res = {n: evaluate(test, pt, c) for n, c in named.items()} lines = ["| Retriever | config (k1, b, organ x; alpha; cap) | hit@4 IID | hit@4 held-out | true entries IID | true entries held-out | vote IID | vote held-out | hit@4 vs BM25 IID [95% CI] | held-out |", "|---|---|" + "---:|" * 8] out = {"validation_n": len(val), "test_n": len(test), "selected": {n: list(map(str, c)) for n, c in named.items()}, "test": {}, "validation_top": [[str(c), s] for c, s in rows[:20]]} for n, c in named.items(): si, so = summary(res[n], "iid"), summary(res[n], "ood"); out["test"][n] = {"iid": si, "ood": so} ci = "" if n == "BM25 (paper)" else f"{si['hit4'] - summary(res['BM25 (paper)'], 'iid')['hit4']:+.1f} {paired_ci(res[n], res['BM25 (paper)'], 'iid')}" co = "" if n == "BM25 (paper)" else f"{so['hit4'] - summary(res['BM25 (paper)'], 'ood')['hit4']:+.1f} {paired_ci(res[n], res['BM25 (paper)'], 'ood')}" lines.append(f"| {n} | {c[0]}; {c[1]:g}; {c[2]} | {si['hit4']:.1f} | {so['hit4']:.1f} | {si['ntrue']:.2f} | {so['ntrue']:.2f} | {si['vote']:.1f} | {so['vote']:.1f} | {ci} | {co} |") (OUT / "tune_table.md").write_text("\n".join(lines) + "\n"); (OUT / "tune_results.json").write_text(json.dumps(out, indent=1)) # retrieved ids of the selected configs, so a diagnosis run can use them directly sel = {} for n, (cfgT, a, cap) in named.items(): for (c, sp, q, lab, mids, mlabs), (T, img) in zip(test, pt): sel.setdefault(n, {})[q] = [mids[i] for i in pick(T[cfgT] + a * img, mids, mlabs, cap)] (OUT / "tune_retrieved.json").write_text(json.dumps(sel)) print("\n".join(lines)) if __name__ == "__main__": main()