v802-classifier / src /build_v802.py
sneo's picture
Add v802 classifier head and source code
59839e3 verified
Raw History Blame Contribute Delete
11.3 kB
#!/usr/bin/env python3
"""Fit the final v802 on development736 and evaluate it once on both holdouts.
The config JSON lists weighted components and the decoder weights chosen on
development OOF (see experiment_v802_ensemble.py):
{"components": [{"name": "v801", "weight": 0.3},
{"name": "probe", "weight": 0.7, "oof_prefix": "...",
"features": [...], "aug_features": [...], "layers": [...],
"stats": [...], "c": 0.03}],
"prior": 0.5, "card": 0.25}
Component probabilities are fitted on all of development736. The structured
decoder (label-set prior + cardinality model) is fitted on the 296 target
images using stacked OOF probabilities from every dev/check seed, matching how
it was selected.
"""
from __future__ import annotations
import argparse
import json
from pathlib import Path
import joblib
import numpy as np
from scipy.stats import binomtest
from sklearn.linear_model import LogisticRegression
from experiment_v802_probe import l2, pooled_features
from train_v801 import (
EPS, Spec, all_labelsets, binary_probabilities, fit_binary, labelset_ids, make_variant,
meta_features, report_metrics,
)
from v802_common import CHECK_SEEDS, DEV_SEEDS, LEGACY_FEATURES, ROOT, dump, load_data, logit
TRIALS = ROOT / "artifacts" / "models" / "v802_trials"
OUTPUT = ROOT / "artifacts" / "models" / "v802"
def knn_predict(x_train, y_train, x_query, k=11, temperature=0.1):
similarity = x_query @ x_train.T
neighbours = np.argpartition(similarity, -k, axis=1)[:, -k:]
neighbour_similarity = np.take_along_axis(similarity, neighbours, axis=1)
neighbour_similarity -= neighbour_similarity.max(axis=1, keepdims=True)
weights = np.exp(neighbour_similarity / temperature)
weights /= weights.sum(axis=1, keepdims=True)
return np.einsum("nk,nkc->nc", weights, y_train[neighbours])
def fit_component(spec: dict, data, train_idx, query_idx):
"""Return (fitted state, query probabilities) for one component."""
if spec["name"] == "mil":
from experiment_v802_mil import load_projected
from v802_mil import MilConfig, fit_projection, project, train_predict
base = Path(spec["features"])
x = load_projected(base, spec["layers"], spec["dims"], data)
x_aug = (load_projected(Path(spec["aug_features"]), spec["layers"], spec["dims"], data,
basis_dir=base) if spec.get("aug_features") else None)
# Recompute the (deterministic) PCA bases behind the cached projection so they can be saved.
pca = []
for slot, layer in enumerate(spec["layers"]):
patches = np.load(base / f"patches_l{layer}.npy", mmap_mode="r")
mean, std, basis = fit_projection(patches, data.anchor_idx, spec["dims"])
check = project(patches[:4], mean, std, basis)
cached = x[:4, :, slot * spec["dims"] : (slot + 1) * spec["dims"]]
assert np.allclose(check, cached, atol=1e-5), f"PCA mismatch for layer {layer}"
pca.append({"layer": layer, "mean": mean, "std": std, "basis": basis})
heads: list = []
probs, _ = train_predict(x, data.y, train_idx, query_idx, MilConfig(**spec["config"]),
spec.get("seed", 802), x_aug, models_out=heads)
return {"mil_config": spec["config"], "layers": spec["layers"], "dims": spec["dims"],
"pca": pca, "heads": heads}, probs
if spec["name"] == "v801":
bundle = np.load(LEGACY_FEATURES)
lookup = {n: i for i, n in enumerate(bundle["file_names"].astype(str))}
order = np.asarray([lookup[n] for n in data.names])
legacy_x = make_variant(bundle, "legacy")[order]
patch_x = make_variant(bundle, "meanmax_clspatchmax")[order]
probe = fit_binary(legacy_x[train_idx], data.y[train_idx],
Spec("legacy", 0.003, class_weight=None), np.ones(len(train_idx)))
legacy = binary_probabilities(probe, legacy_x[query_idx])
knn = knn_predict(patch_x[train_idx], data.y[train_idx], patch_x[query_idx])
state = {"legacy_probe": probe, "knn_train_names": data.names[train_idx]}
return state, 0.8 * legacy + 0.2 * knn
dirs = [Path(d) for d in spec["features"]]
x = l2(np.mean([pooled_features(d, spec["layers"], spec["stats"]) for d in dirs], axis=0))
views = [x] + [pooled_features(Path(d), spec["layers"], spec["stats"])
for d in spec.get("aug_features", [])]
probe = fit_binary(np.concatenate([v[train_idx] for v in views]),
np.tile(data.y[train_idx], (len(views), 1)),
Spec("v802probe", spec["c"], class_weight=None),
np.ones(len(train_idx) * len(views)))
probs = np.mean([binary_probabilities(probe, v[query_idx]) for v in views], axis=0)
return {"probe": probe}, probs
def component_oof(spec: dict, data, seed: int) -> np.ndarray:
if spec["name"] == "v801":
stored = np.load(TRIALS / "cache" / f"v801_oof_seed{seed}.npz")
probs = 0.8 * stored["legacy"] + 0.2 * stored["knn"]
else:
stored = np.load(TRIALS / "oof" / f"{spec['oof_prefix']}_seed{seed}.npz")
probs = stored["probs"]
assert (stored["names"] == data.names[data.target_idx]).all()
return probs
def fit_decoder(data, config):
y_target = data.y[data.target_idx]
labelsets = all_labelsets(y_target.shape[1])
counts = np.bincount(labelset_ids(y_target, labelsets), minlength=len(labelsets))
log_prior = np.log((counts + 1.0) / (len(y_target) + len(labelsets)))
seeds = [s for s in DEV_SEEDS + CHECK_SEEDS
if all(c["name"] == "v801" or (TRIALS / "oof" / f"{c['oof_prefix']}_seed{s}.npz").exists()
for c in config["components"])]
oof = [sum(c["weight"] * component_oof(c, data, s) for c in config["components"]) for s in seeds]
cardinality = LogisticRegression(C=1.0, max_iter=3000, class_weight="balanced", random_state=801)
cardinality.fit(np.concatenate([meta_features(p) for p in oof]),
np.tile(y_target.sum(axis=1), len(oof)))
return {"labelsets": labelsets, "log_prior": log_prior, "cardinality": cardinality,
"decoder_seeds": seeds}
def decode(decoder, probs, prior, card):
labelsets = decoder["labelsets"]
card_log = np.log(np.clip(decoder["cardinality"].predict_proba(meta_features(probs)), EPS, 1))
score = logit(probs) @ labelsets.T + prior * decoder["log_prior"]
score += card * card_log[:, labelsets.sum(axis=1).astype(int) - 1]
return labelsets[score.argmax(axis=1)]
def metrics(y, probs, pred):
m = report_metrics(y, probs, pred)
return {"exact_match_accuracy": float(m["exact_match_accuracy"]),
"micro_f1": float(m["micro_f1"]), "macro_f1": float(m["macro_f1"]),
"mean_average_precision": float(m["mean_average_precision"]),
"cardinality_accuracy": float((pred.sum(1) == y.sum(1)).mean()),
"samples": int(len(y))}
def paired(y, pred_new, pred_old, seed=802000, resamples=10000):
new_ok = (pred_new == y).all(1)
old_ok = (pred_old == y).all(1)
gained, lost = int((new_ok & ~old_ok).sum()), int((~new_ok & old_ok).sum())
rng = np.random.default_rng(seed)
idx = rng.integers(0, len(y), (resamples, len(y)))
deltas = new_ok[idx].mean(1) - old_ok[idx].mean(1)
return {"delta_exact": float(new_ok.mean() - old_ok.mean()), "gained": gained, "lost": lost,
"mcnemar_exact_p": float(binomtest(gained, gained + lost).pvalue) if gained + lost else 1.0,
"bootstrap_delta_95pct": [float(np.quantile(deltas, 0.025)),
float(np.quantile(deltas, 0.975))]}
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--config", type=Path, required=True)
parser.add_argument("--output", type=Path, default=OUTPUT)
args = parser.parse_args()
config = json.loads(args.config.read_text())
dirs = []
for spec in config["components"]:
for key in ("features", "aug_features"):
value = spec.get(key) or []
dirs += [value] if isinstance(value, str) else value
names = json.loads((Path(dirs[0]) / "meta.json").read_text())["file_names"]
for d in dirs:
assert json.loads((Path(d) / "meta.json").read_text())["file_names"] == names, d
data = load_data(names)
train_idx = data.development_idx
holdout_idx = np.concatenate((data.new_idx, data.historical_idx))
decoder = fit_decoder(data, config) # cheap; fails fast on missing/corrupt OOF files
states, holdout_probs, baseline_probs = [], 0.0, None
for spec in config["components"]:
state, probs = fit_component(spec, data, train_idx, holdout_idx)
states.append(state)
holdout_probs = holdout_probs + spec["weight"] * probs
if spec["name"] == "v801":
baseline_probs = probs
pred = decode(decoder, holdout_probs, config["prior"], config["card"])
if baseline_probs is None:
_, baseline_probs = fit_component({"name": "v801"}, data, train_idx, holdout_idx)
baseline_decoder = fit_decoder(data, {"components": [{"name": "v801", "weight": 1.0}]})
baseline_pred = decode(baseline_decoder, baseline_probs, 0.5, 0.25)
y = data.y[holdout_idx]
groups = {"new_holdout32": np.arange(len(data.new_idx)),
"historical_holdout32": np.arange(len(data.new_idx), len(holdout_idx)),
"combined64": np.arange(len(holdout_idx))}
report = {"protocol": "v802 components fitted on development736; decoder fitted on stacked "
"target OOF probabilities; each holdout evaluated once. v801 baseline "
"is refitted here with the same protocol (legacy probe + kNN, prior 0.5, "
"card 0.25).",
"config": config, "decoder_seeds": decoder["decoder_seeds"], "holdouts": {}}
for group, rows in groups.items():
report["holdouts"][group] = {
"v802": metrics(y[rows], holdout_probs[rows], pred[rows]),
"v801_refit": metrics(y[rows], baseline_probs[rows], baseline_pred[rows]),
"paired_v802_vs_v801_refit": paired(y[rows], pred[rows], baseline_pred[rows]),
}
print(group, json.dumps(report["holdouts"][group]["v802"]),
"v801_refit", report["holdouts"][group]["v801_refit"]["exact_match_accuracy"],
flush=True)
classes = list(data.classes)
report["predictions"] = [
{"file_name": str(data.names[i]), "true": [classes[j] for j in np.flatnonzero(data.y[i])],
"v802": [classes[j] for j in np.flatnonzero(p)],
"v801_refit": [classes[j] for j in np.flatnonzero(b)],
"v802_probabilities": [round(float(v), 6) for v in q]}
for i, p, b, q in zip(holdout_idx, pred, baseline_pred, holdout_probs)
]
args.output.mkdir(parents=True, exist_ok=True)
dump(args.output / "holdout_metrics.json", report)
joblib.dump({"config": config, "classes": classes, "components": states, "decoder": decoder},
args.output / "classifier.joblib")
if __name__ == "__main__":
main()