dna_noc / src /cli /run_aggressive_tuning.py
manhngvu's picture
🎉 Update: Optimized models with F1=0.7135 + Complete research report + Analysis (2026-05-08)
4636192 verified
Raw
History Blame Contribute Delete
11.3 kB
"""Run aggressive hyperparameter tuned pipeline for F1 > 0.5."""
from __future__ import annotations
import argparse
import json
import sys
import time
from pathlib import Path
from typing import Dict, List, Tuple
import numpy as np
import pandas as pd
from catboost import CatBoostClassifier
from lightgbm import LGBMClassifier
from sklearn.metrics import (
accuracy_score,
average_precision_score,
confusion_matrix,
f1_score,
precision_score,
recall_score,
roc_auc_score,
)
from sklearn.preprocessing import StandardScaler
from xgboost import XGBClassifier
try:
from imblearn.over_sampling import SMOTE
SMOTE_AVAILABLE = True
except ImportError:
SMOTE_AVAILABLE = False
ROOT = Path(__file__).resolve().parents[2]
if str(ROOT) not in sys.path:
sys.path.insert(0, str(ROOT))
EPS = 1e-9
def _agg_numeric(frame: pd.DataFrame, key: str, value_cols, prefix: str) -> pd.DataFrame:
grouped = frame.groupby(key, sort=False)[list(value_cols)]
agg = grouped.agg(["mean", "std", "min", "max", "median"])
agg.columns = [f"{prefix}_{col}_{stat}" for col, stat in agg.columns]
agg = agg.reset_index()
for col in agg.columns:
if col != key:
agg[col] = agg[col].fillna(0.0)
return agg
def _build_features(benchmark_dir) -> pd.DataFrame:
marker_path = benchmark_dir / "marker_table.csv"
sample_key = "sample_file"
marker_cols = [
sample_key, "marker", "dye", "peak_count_total", "peak_count_non_ol",
"max_height", "sum_height", "has_ol",
]
marker = pd.read_csv(marker_path, usecols=marker_cols, low_memory=False)
marker["peak_nonol_ratio"] = marker["peak_count_non_ol"] / (marker["peak_count_total"] + EPS)
marker["height_density"] = marker["sum_height"] / (marker["peak_count_total"] + EPS)
gm = marker.groupby(sample_key, sort=False)
mfeat = pd.DataFrame({
sample_key: gm.size().index,
"marker_rows": gm.size().values,
"marker_unique_count": gm["marker"].nunique().values,
"marker_has_ol_rate": gm["has_ol"].mean().values,
"marker_peak_total_sum": gm["peak_count_total"].sum().values,
"marker_peak_nonol_sum": gm["peak_count_non_ol"].sum().values,
"marker_peak_nonol_ratio_mean": gm["peak_nonol_ratio"].mean().values,
"marker_height_density_mean": gm["height_density"].mean().values,
})
mnum = _agg_numeric(
marker, key=sample_key,
value_cols=["peak_count_total", "peak_count_non_ol", "peak_nonol_ratio", "max_height", "sum_height", "height_density"],
prefix="marker",
)
mfeat = mfeat.merge(mnum, on=sample_key, how="left")
dye_sum = marker.pivot_table(
index=sample_key, columns="dye", values="sum_height", aggfunc="sum", fill_value=0.0,
)
dye_sum.columns = [f"marker_sum_height_dye_{c}" for c in dye_sum.columns]
mfeat = mfeat.merge(dye_sum.reset_index(), on=sample_key, how="left")
return mfeat
def _tune_threshold(y_true: np.ndarray, prob: np.ndarray, steps: int) -> Tuple[float, float]:
thresholds = np.linspace(0.01, 0.99, steps)
best_f1 = -1.0
best_th = 0.5
for th in thresholds:
pred = (prob >= th).astype(int)
score = f1_score(y_true, pred, zero_division=0)
if score > best_f1:
best_f1 = float(score)
best_th = float(th)
return best_th, best_f1
def _metrics(y_true: np.ndarray, prob: np.ndarray, threshold: float) -> Dict[str, float]:
pred = (prob >= threshold).astype(int)
tn, fp, fn, tp = confusion_matrix(y_true, pred, labels=[0, 1]).ravel()
specificity = tn / (tn + fp + EPS)
precision = precision_score(y_true, pred, zero_division=0)
recall = recall_score(y_true, pred, zero_division=0)
f2 = (5 * precision * recall) / (4 * precision + recall + EPS)
return {
"roc_auc": float(roc_auc_score(y_true, prob)),
"pr_auc": float(average_precision_score(y_true, prob)),
"f1": float(f1_score(y_true, pred, zero_division=0)),
"f2": float(f2),
"precision": float(precision),
"recall": float(recall),
"specificity": float(specificity),
"accuracy": float(accuracy_score(y_true, pred)),
"tp": int(tp), "fp": int(fp), "tn": int(tn), "fn": int(fn),
}
def run(args: argparse.Namespace) -> Tuple[pd.DataFrame, pd.DataFrame]:
args.out_dir.mkdir(parents=True, exist_ok=True)
print("=" * 80)
print("🚀 AGGRESSIVE HYPERPARAMETER TUNING PIPELINE")
print("=" * 80)
per_split_rows: List[Dict[str, object]] = []
total_start = time.time()
for benchmark_name in args.benchmarks:
print(f"\n📊 Processing benchmark: {benchmark_name}")
benchmark_dir = args.benchmark_root / benchmark_name
labels = pd.read_csv(benchmark_dir / "sample_labels_all_splits.csv", low_memory=False)
features = _build_features(benchmark_dir)
for split_idx, split_id in enumerate(sorted(labels["split_id"].unique())):
print(f" Split {split_idx + 1}/{len(labels['split_id'].unique())}: {split_id}", end=" ")
split_start = time.time()
split_df = labels[labels["split_id"] == split_id].copy()
data = split_df.merge(features, on="sample_file", how="inner")
drop_cols = {"benchmark_id", "split_id", "partition", "study_id", "panel", "sample_file", "sample_family_id", "true_contributors", "known_contributors_true", "unknown_contributors_true", "num_known_in_sample", "num_unknown_in_sample", "unknown_present", "total_contributors"}
feature_cols = [c for c in data.columns if c not in drop_cols]
x_all = data[feature_cols].astype(float).values
y_all = data["unknown_present"].astype(int).values
partition = data["partition"].values
train_idx, dev_idx, test_idx = partition == "train", partition == "dev", partition == "test"
x_train, y_train = x_all[train_idx], y_all[train_idx]
x_dev, y_dev = x_all[dev_idx], y_all[dev_idx]
x_test, y_test = x_all[test_idx], y_all[test_idx]
if SMOTE_AVAILABLE:
smote = SMOTE(random_state=42, k_neighbors=3) # Reduced k_neighbors
x_train, y_train = smote.fit_resample(x_train, y_train)
scaler = StandardScaler()
x_train = scaler.fit_transform(x_train)
x_dev = scaler.transform(x_dev)
x_test = scaler.transform(x_test)
pos = float((y_train == 1).sum())
neg = float((y_train == 0).sum())
spw = max(1.0, neg / max(pos, 1.0))
# AGGRESSIVE HYPERPARAMETERS
lgbm = LGBMClassifier(
n_estimators=2000, # More trees
learning_rate=0.005, # Slower learning
num_leaves=50, # Fewer leaves (less overfitting)
subsample=0.7, # Lower subsample
colsample_bytree=0.7,
min_child_samples=40, # Stricter constraints
lambda_l1=5.0, # More L1 regularization
lambda_l2=5.0,
class_weight="balanced",
random_state=42,
verbose=-1,
n_jobs=-1,
)
xgb = XGBClassifier(
n_estimators=2000,
learning_rate=0.005,
max_depth=5, # Shallower trees
min_child_weight=10, # Stricter constraints
subsample=0.7,
colsample_bytree=0.7,
reg_alpha=5.0, # More regularization
reg_lambda=5.0,
scale_pos_weight=spw,
eval_metric="logloss",
random_state=42,
n_jobs=-1,
)
cb = CatBoostClassifier(
iterations=2000,
learning_rate=0.005,
depth=5,
l2_leaf_reg=5.0,
scale_pos_weight=spw,
random_state=42,
verbose=0,
task_type="CPU",
)
lgbm.fit(x_train, y_train)
xgb.fit(x_train, y_train)
cb.fit(x_train, y_train)
p_dev_l = lgbm.predict_proba(x_dev)[:, 1]
p_dev_x = xgb.predict_proba(x_dev)[:, 1]
p_dev_c = cb.predict_proba(x_dev)[:, 1]
p_test_l = lgbm.predict_proba(x_test)[:, 1]
p_test_x = xgb.predict_proba(x_test)[:, 1]
p_test_c = cb.predict_proba(x_test)[:, 1]
th_l, dev_f1_l = _tune_threshold(y_dev, p_dev_l, args.threshold_steps)
th_x, dev_f1_x = _tune_threshold(y_dev, p_dev_x, args.threshold_steps)
th_c, dev_f1_c = _tune_threshold(y_dev, p_dev_c, args.threshold_steps)
# Best ensemble
best_weight = [1/3, 1/3, 1/3]
best_th = 0.5
best_f1 = -1.0
for w_l in np.linspace(0, 1, 6):
for w_x in np.linspace(0, 1-w_l, 6):
w_c = 1.0 - w_l - w_x
p_dev = (w_l * p_dev_l) + (w_x * p_dev_x) + (w_c * p_dev_c)
th, dev_f1 = _tune_threshold(y_dev, p_dev, args.threshold_steps)
if dev_f1 > best_f1:
best_f1, best_th = dev_f1, th
best_weight = [w_l, w_x, w_c]
p_test_ens = (best_weight[0] * p_test_l) + (best_weight[1] * p_test_x) + (best_weight[2] * p_test_c)
for name, prob, th in [("lgbm", p_test_l, th_l), ("xgb", p_test_x, th_x), ("cb", p_test_c, th_c), ("ensemble", p_test_ens, best_th)]:
m = _metrics(y_test, prob, th)
per_split_rows.append({"benchmark": benchmark_name, "split_id": split_id, "model": name, **m})
print(f"✅ {time.time() - split_start:.1f}s")
per_split = pd.DataFrame(per_split_rows).sort_values(["benchmark", "model", "split_id"])
summary = per_split.groupby(["benchmark", "model"], as_index=False).agg(
f1_mean=("f1", "mean"), f1_std=("f1", "std"),
precision_mean=("precision", "mean"), recall_mean=("recall", "mean"),
)
per_split.to_csv(args.out_dir / "aggressive_tuning_per_split.csv", index=False)
summary.to_csv(args.out_dir / "aggressive_tuning_summary.csv", index=False)
print(f"\n{'='*80}\nTotal: {(time.time()-total_start)/60:.1f} min\n{'='*80}\n")
return per_split, summary
def main() -> None:
parser = argparse.ArgumentParser()
parser.add_argument("--benchmark-root", type=Path, default=Path("data/processed"))
parser.add_argument("--benchmarks", nargs="+", default=["rd14-fullref-50_multisplit_v2", "rd12-fullref-61_multisplit_v2"])
parser.add_argument("--out-dir", type=Path, default=Path("outputs/benchmarks/aggressive_tuning"))
parser.add_argument("--threshold-steps", type=int, default=199)
args = parser.parse_args()
_, summary = run(args)
print("\n📊 RESULTS:")
print(summary[["benchmark", "model", "f1_mean", "precision_mean", "recall_mean"]].to_string(index=False))
best_f1 = summary["f1_mean"].max()
if best_f1 > 0.5:
print(f"\n✅ SUCCESS! Best F1 = {best_f1:.4f}")
else:
print(f"\n⚠️ Target not met. Best F1 = {best_f1:.4f}")
if __name__ == "__main__":
main()