"""Shreyan's retriever with the image feature: Sonnet writes the weights of a small linear scoring network over the 16 caption features of pilot.py plus DINOv2 image similarity (min-max scaled within the memory). Two contexts, one Sonnet call per crop each: 'schema' (feature definitions, crop, disease list) and 'memory' (the same plus the crop's memory captions and labels). The context is chosen on the validation set of tune.py (hit@4, then true entries); the test queries are scored once, and the chosen retriever's test retrieval is written for full.py. python -m experiments.retriever_pilot.llm_image """ from __future__ import annotations import json from collections import Counter, defaultdict import numpy as np from agmem import llm, packet from . import pilot as P, tune as T FEATS = P.FEATS + ["image_sim"] DOC = {**P.FEAT_DOC, "image_sim": "Visual similarity of the two images (DINOv2 image-embedding cosine), scaled so the most similar memory image is 1 and the least similar is 0. It does not use the captions."} def prompt(crop, diseases, bank, caps, with_memory): lines = [f"Task: a frozen vision-language model diagnoses {crop.lower()} diseases. Each reference image in memory is stored as a", "label-free text caption plus its disease label. For a test image we caption it the same way, score every memory entry with", "score = sum_i w_i * feature_i, and give the 4 highest-scoring entries (caption + label) to the diagnosis model.", "Good weights rank entries of the test image's true disease highest; the label is unknown at test time.", "", f"Possible diseases: {', '.join(diseases)}.", "", "Features (each computed between the test image/caption and one memory entry):"] lines += [f"- {k}: {DOC[k]}" for k in FEATS] if with_memory: lines += ["", f"The memory for this crop ({len(bank)} entries), as 'label | caption':"] lines += [f"- {lab} | {packet.caption_text(caps[i])}" for i, lab in bank] lines += ["", "Write the weights."] return "\n".join(lines) def weights(P1, caps, with_memory): out = {} for crop in sorted({j["crop"] for j in P1}): J = [j for j in P1 if j["crop"] == crop] bank = sorted({(m["image_id"], m["label"]) for m in J[0]["memory"]}, key=lambda x: (x[1], x[0])) r = llm.call(P.LLM_SYS, [{"type": "text", "text": prompt(crop, sorted({l for _, l in bank}), bank, caps, with_memory)}], stage="retriever_pilot_img", tag=f"{'memory' if with_memory else 'schema'}|{crop}", reserve_usd=0.15) p = P.last_json(r.get("text", "") or "") or {}; w = p.get("weights", {}) out[crop] = {"w": [float(w.get(k, 0.0)) for k in FEATS], "missing": [k for k in FEATS if k not in w], "rationale": p.get("rationale"), "cost_usd": r.get("cost_usd")} return out def score_items(its, caps, E, idx, W): """Per item: top-4 indices under the crop's weights; plus BM25 and tuned text+image for reference.""" res = defaultdict(lambda: defaultdict(lambda: defaultdict(list))) picks = defaultdict(dict) for c, sp, q, lab, mids, mlabs in its: F = P.features(caps[q], mids, caps) cos = E[[idx[m] for m in mids]] @ E[idx[q]]; img = (cos - cos.min()) / ((cos.max() - cos.min()) or 1.0) F = np.hstack([F, img[:, None]]) for name, w in W.items(): sel = T.pick(F @ np.array(w[c]["w"]), mids, mlabs, 4); 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 = res[name][(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) picks[name][q] = [mids[i] for i in sel] return res, picks 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 = T.items(P1, caps, E, idx) llm.set_budget(llm.spent() + 2.0) W = {"llm_img_schema": weights(P1, caps, False), "llm_img_memory": weights(P1, caps, True)} print("missing weights:", {k: sum(len(x["missing"]) for x in v.values()) for k, v in W.items()}, "USD:", round(sum((x["cost_usd"] or 0) for v in W.values() for x in v.values()), 3)) (P.OUT / "llm_image_weights.json").write_text(json.dumps({k: {c: {**x, "features": FEATS} for c, x in v.items()} for k, v in W.items()}, indent=1)) vres, _ = score_items(val, caps, E, idx, W) pv = T.precompute(val, caps, E, idx) ref = {"BM25 (paper)": ((1.2, 0.75, 3), 0, 4), "Tuned text + image": ((2.0, 0.3, 0), 1.5, 4)} print("validation (equal-crop):") for n, cfg in ref.items(): s = T.summary(T.evaluate(val, pv, cfg), "val"); print(f" {n:<22} hit@4 {s['hit4']:.1f} true entries {s['ntrue']:.2f} vote {s['vote']:.1f}") vs = {n: T.summary(vres[n], "val") for n in W} for n, s in vs.items(): print(f" {n:<22} hit@4 {s['hit4']:.1f} true entries {s['ntrue']:.2f} vote {s['vote']:.1f}") chosen = max(vs, key=lambda n: (round(vs[n]["hit4"], 6), vs[n]["ntrue"])) print("chosen on validation:", chosen) tres, picks = score_items(test, caps, E, idx, {chosen: W[chosen]}) for sp in ("iid", "ood"): s = T.summary(tres[chosen], sp); print(f"test {sp}: hit@4 {s['hit4']:.1f} true entries {s['ntrue']:.2f} vote {s['vote']:.1f}") (P.OUT / "llm_image_retrieved.json").write_text(json.dumps({"chosen": chosen, "retrieved": picks[chosen]})) (P.OUT / "llm_image_results.json").write_text(json.dumps({"validation": vs, "chosen": chosen, "test": {sp: T.summary(tres[chosen], sp) for sp in ("iid", "ood")}}, indent=1)) if __name__ == "__main__": main()