#!/usr/bin/env python3 """Logistic probes on v802 pooled descriptors (native-aspect DINOv3 / SigLIP2).""" from __future__ import annotations import argparse import json import time from pathlib import Path import numpy as np from train_v801 import Spec, binary_probabilities, fit_binary, oof_probabilities from v802_common import ( CHECK_SEEDS, DEV_SEEDS, ROOT, dump, evaluate, load_data, make_folds, summarize, v801_oof, ) TRIALS = ROOT / "artifacts" / "models" / "v802_trials" DINO_STATS = {"cls": 0, "reg": 1, "mean": 2, "max": 3, "std": 4} SIGLIP_STATS = {"emb": 0, "mean": 1, "max": 2, "std": 3} def l2(x: np.ndarray) -> np.ndarray: return x / np.maximum(np.linalg.norm(x, axis=-1, keepdims=True), 1e-12) def pooled_features(feature_dir: Path, layers: list[int] | None, stats: list[str]) -> np.ndarray: """Concatenate per-block L2-normalized descriptors; averages with a sibling -flip dir if given.""" meta = json.loads((feature_dir / "meta.json").read_text()) pooled = np.load(feature_dir / "pooled.npy") if pooled.ndim == 4: # DINOv3: (N, L, S, D) slots = [meta["pooled_layers"].index(layer) for layer in layers] blocks = [pooled[:, s, DINO_STATS[k]] for s in slots for k in stats] else: # SigLIP2: (N, S, D) blocks = [pooled[:, SIGLIP_STATS[k]] for k in stats] return l2(np.concatenate([l2(b) for b in blocks], axis=1)) def aug_oof_probabilities(views, y, anchor_idx, target_idx, folds, spec: Spec) -> np.ndarray: """Train on every view (e.g. raw + flipped) as separate samples; average view probabilities.""" result = np.zeros((len(target_idx), y.shape[1])) for train_local, valid_local in folds: train_idx = np.concatenate((anchor_idx, target_idx[train_local])) x_train = np.concatenate([v[train_idx] for v in views]) y_train = np.tile(y[train_idx], (len(views), 1)) probe = fit_binary(x_train, y_train, spec, np.ones(len(x_train))) valid_idx = target_idx[valid_local] result[valid_local] = np.mean([binary_probabilities(probe, v[valid_idx]) for v in views], axis=0) return result def main() -> None: parser = argparse.ArgumentParser() parser.add_argument("--features", type=Path, nargs="+", required=True, help="feature dirs; multiple dirs (e.g. raw + flip) are averaged") parser.add_argument("--aug-features", type=Path, nargs="*", default=[], help="extra view dirs used as training augmentation + probability TTA") parser.add_argument("--layers", type=int, nargs="+", default=[24]) parser.add_argument("--stats", nargs="+", default=["cls", "mean"]) parser.add_argument("--c", type=float, nargs="+", default=[0.001, 0.003, 0.01, 0.03]) parser.add_argument("--name", required=True) parser.add_argument("--seeds", choices=("dev", "check", "all"), default="dev") args = parser.parse_args() started = time.time() meta = json.loads((args.features[0] / "meta.json").read_text()) data = load_data(meta["file_names"]) x = l2(np.mean([pooled_features(d, args.layers, args.stats) for d in args.features], axis=0)) views = [x] + [pooled_features(d, args.layers, args.stats) for d in args.aug_features] seeds = {"dev": DEV_SEEDS, "check": CHECK_SEEDS, "all": DEV_SEEDS + CHECK_SEEDS}[args.seeds] report = {"name": args.name, "features": [str(d) for d in args.features], "aug_features": [str(d) for d in args.aug_features], "layers": args.layers, "stats": args.stats, "dimension": int(x.shape[1]), "seeds": list(seeds), "by_c": {}} for c in args.c: rows = {"probe": [], "mix": []} for seed in seeds: folds = make_folds(data, seed) spec = Spec("v802probe", c, class_weight=None) probs = ( aug_oof_probabilities(views, data.y, data.anchor_idx, data.target_idx, folds, spec) if args.aug_features else oof_probabilities(x, data.y, data.origins, data.anchor_idx, data.target_idx, folds, spec) ) base = v801_oof(data, seed, TRIALS / "cache") v801 = 0.8 * base["legacy"] + 0.2 * base["knn"] rows["probe"].append(evaluate(data, probs, folds)) rows["mix"].append(evaluate(data, 0.5 * probs + 0.5 * v801, folds)) (TRIALS / "oof").mkdir(parents=True, exist_ok=True) np.savez(TRIALS / "oof" / f"{args.name}_c{c}_seed{seed}.npz", names=data.names[data.target_idx], probs=probs) report["by_c"][str(c)] = {k: summarize(v) for k, v in rows.items()} print(c, json.dumps(report["by_c"][str(c)]), flush=True) report["seconds"] = round(time.time() - started, 1) dump(TRIALS / f"{args.name}.json", report) if __name__ == "__main__": main()