relaion2b-natural / classifier /compare_classifiers.py
andropar's picture
Add trained naturalness classifier with weights, diagnostics, and evaluation
ac53f17 verified
"""Compare classifier architectures on the ViT-L/14 features.
Conservative resource usage — single-threaded, sequential, no large ensembles.
"""
import pickle
import json
import warnings
import numpy as np
import matplotlib
matplotlib.use("Agg")
import matplotlib.pyplot as plt
from pathlib import Path
from sklearn.model_selection import StratifiedKFold, train_test_split
from sklearn.linear_model import LogisticRegression
from sklearn.neural_network import MLPClassifier
from sklearn.ensemble import HistGradientBoostingClassifier
from sklearn.preprocessing import StandardScaler
from sklearn.pipeline import Pipeline
from sklearn.metrics import (
roc_auc_score, average_precision_score, roc_curve,
precision_recall_curve, f1_score, accuracy_score,
)
warnings.filterwarnings("ignore")
OUT_DIR = Path(__file__).parent
DATA_PATH = Path(
"/home/jroth/photograph_detector/scripts/outputs/"
"extract_openai_vitl14_features/clip_vitl14_features_labeled.pkl"
)
N_FOLDS = 5
RANDOM_STATE = 42
def get_classifiers():
"""Conservative set of classifiers — no n_jobs, modest sizes."""
return {
"LogReg (C=0.1)": Pipeline([
("scaler", StandardScaler()),
("clf", LogisticRegression(C=0.1, max_iter=1000, random_state=RANDOM_STATE)),
]),
"LogReg (C=1)": Pipeline([
("scaler", StandardScaler()),
("clf", LogisticRegression(C=1.0, max_iter=1000, random_state=RANDOM_STATE)),
]),
"LogReg (C=10)": Pipeline([
("scaler", StandardScaler()),
("clf", LogisticRegression(C=10.0, max_iter=1000, random_state=RANDOM_STATE)),
]),
"MLP (256)": Pipeline([
("scaler", StandardScaler()),
("clf", MLPClassifier(
hidden_layer_sizes=(256,), max_iter=300,
early_stopping=True, validation_fraction=0.1,
random_state=RANDOM_STATE,
)),
]),
"MLP (512, 256)": Pipeline([
("scaler", StandardScaler()),
("clf", MLPClassifier(
hidden_layer_sizes=(512, 256), max_iter=300,
early_stopping=True, validation_fraction=0.1,
random_state=RANDOM_STATE,
)),
]),
"MLP (512, 256, 128)": Pipeline([
("scaler", StandardScaler()),
("clf", MLPClassifier(
hidden_layer_sizes=(512, 256, 128), max_iter=300,
early_stopping=True, validation_fraction=0.1,
random_state=RANDOM_STATE,
)),
]),
"MLP (256, alpha=1e-3)": Pipeline([
("scaler", StandardScaler()),
("clf", MLPClassifier(
hidden_layer_sizes=(256,), max_iter=300,
early_stopping=True, validation_fraction=0.1,
alpha=1e-3, random_state=RANDOM_STATE,
)),
]),
"HistGBM (100, d5)": HistGradientBoostingClassifier(
max_iter=100, max_depth=5, learning_rate=0.1,
early_stopping=True, validation_fraction=0.1,
random_state=RANDOM_STATE,
),
"HistGBM (300, d6)": HistGradientBoostingClassifier(
max_iter=300, max_depth=6, learning_rate=0.05,
early_stopping=True, validation_fraction=0.1,
random_state=RANDOM_STATE,
),
}
def manual_cv(clf_factory, X, y, n_folds=N_FOLDS):
"""Run CV manually one fold at a time to keep memory low."""
cv = StratifiedKFold(n_splits=n_folds, shuffle=True, random_state=RANDOM_STATE)
fold_metrics = []
for fold, (train_idx, val_idx) in enumerate(cv.split(X, y)):
clf = clf_factory()
clf.fit(X[train_idx], y[train_idx])
probs = clf.predict_proba(X[val_idx])[:, 1]
preds = (probs >= 0.5).astype(int)
fold_metrics.append({
"roc_auc": roc_auc_score(y[val_idx], probs),
"avg_precision": average_precision_score(y[val_idx], probs),
"accuracy": accuracy_score(y[val_idx], preds),
"f1": f1_score(y[val_idx], preds),
})
del clf # free memory between folds
return {
metric: {
"mean": float(np.mean([f[metric] for f in fold_metrics])),
"std": float(np.std([f[metric] for f in fold_metrics])),
}
for metric in ["roc_auc", "avg_precision", "accuracy", "f1"]
}
def main():
print("=" * 70)
print("CLASSIFIER COMPARISON ON VIT-L/14 FEATURES")
print("=" * 70)
with open(DATA_PATH, "rb") as f:
data = pickle.load(f)
X = data["features"]
y = data["labels"]
print(f"Dataset: {X.shape[0]} samples, {X.shape[1]} features")
print(f"Labels: {dict(zip(*np.unique(y, return_counts=True)))}")
X_train, X_test, y_train, y_test = train_test_split(
X, y, test_size=0.2, random_state=RANDOM_STATE, stratify=y,
)
print(f"Train: {len(y_train)}, Test: {len(y_test)}")
classifiers = get_classifiers()
# --- Phase 1: Cross-validation ---
print(f"\nPHASE 1: {N_FOLDS}-fold CV (sequential, single-threaded)")
print("-" * 70)
cv_results = {}
for name, clf_template in classifiers.items():
print(f" {name}...", end="", flush=True)
try:
# Create a factory that returns a fresh clone each fold
from sklearn.base import clone
factory = lambda t=clf_template: clone(t)
result = manual_cv(factory, X_train, y_train)
cv_results[name] = result
print(f" AUC={result['roc_auc']['mean']:.4f} +/- {result['roc_auc']['std']:.4f}")
except Exception as e:
print(f" FAILED: {e}")
cv_results[name] = {"error": str(e)}
# --- Phase 2: Holdout eval for curves ---
print(f"\nPHASE 2: Holdout evaluation")
print("-" * 70)
holdout_results = {}
for name, clf in classifiers.items():
if "error" in cv_results.get(name, {}):
continue
print(f" {name}...", end="", flush=True)
try:
from sklearn.base import clone
clf = clone(clf)
clf.fit(X_train, y_train)
probs = clf.predict_proba(X_test)[:, 1]
fpr, tpr, _ = roc_curve(y_test, probs)
prec, rec, _ = precision_recall_curve(y_test, probs)
holdout_results[name] = {
"roc_auc": float(roc_auc_score(y_test, probs)),
"avg_precision": float(average_precision_score(y_test, probs)),
"fpr": fpr, "tpr": tpr,
"precision": prec, "recall": rec,
}
print(f" AUC={holdout_results[name]['roc_auc']:.4f}")
del clf
except Exception as e:
print(f" FAILED: {e}")
# --- Phase 3: Plots ---
print(f"\nPHASE 3: Plots")
print("-" * 70)
valid = {k: v for k, v in cv_results.items() if "error" not in v}
names = sorted(valid.keys(), key=lambda k: valid[k]["roc_auc"]["mean"], reverse=True)
means = [valid[n]["roc_auc"]["mean"] for n in names]
stds = [valid[n]["roc_auc"]["std"] for n in names]
colors = ["#e74c3c" if "LogReg" in n else "#3498db" if "MLP" in n
else "#2ecc71" for n in names]
# Bar chart
fig, ax = plt.subplots(figsize=(10, 5))
ax.barh(range(len(names)), means, xerr=stds, color=colors, alpha=0.8,
edgecolor="white", linewidth=0.5, capsize=3)
ax.set_yticks(range(len(names)))
ax.set_yticklabels(names, fontsize=10)
ax.set_xlabel("ROC AUC (5-fold CV)", fontsize=12)
ax.set_title("Classifier Comparison on ViT-L/14 Features", fontsize=13)
ax.grid(axis="x", alpha=0.3)
ax.invert_yaxis()
for i, (m, s) in enumerate(zip(means, stds)):
ax.text(m + s + 0.002, i, f"{m:.4f}", va="center", fontsize=9)
fig.tight_layout()
fig.savefig(OUT_DIR / "classifier_comparison_auc.png", dpi=200, bbox_inches="tight")
plt.close()
print(" Saved classifier_comparison_auc.png")
# All metrics
metrics = ["roc_auc", "avg_precision", "accuracy", "f1"]
metric_labels = ["ROC AUC", "Avg Precision", "Accuracy", "F1"]
fig, axes = plt.subplots(1, 4, figsize=(18, 5), sharey=True)
for ax, metric, label in zip(axes, metrics, metric_labels):
m_means = [valid[n][metric]["mean"] for n in names]
m_stds = [valid[n][metric]["std"] for n in names]
ax.barh(range(len(names)), m_means, xerr=m_stds, color=colors, alpha=0.8,
edgecolor="white", linewidth=0.5, capsize=3)
ax.set_xlabel(label, fontsize=11)
ax.grid(axis="x", alpha=0.3)
ax.invert_yaxis()
for i, (m, s) in enumerate(zip(m_means, m_stds)):
ax.text(m + s + 0.002, i, f"{m:.3f}", va="center", fontsize=8)
axes[0].set_yticks(range(len(names)))
axes[0].set_yticklabels(names, fontsize=10)
fig.suptitle("All Metrics (5-fold CV)", fontsize=13, y=1.01)
fig.tight_layout()
fig.savefig(OUT_DIR / "classifier_comparison_all_metrics.png", dpi=200, bbox_inches="tight")
plt.close()
print(" Saved classifier_comparison_all_metrics.png")
# ROC + PR curves
if holdout_results:
fig, (ax1, ax2) = plt.subplots(1, 2, figsize=(14, 6))
sorted_h = sorted(holdout_results.items(), key=lambda x: x[1]["roc_auc"], reverse=True)
cmap = plt.cm.tab10(np.linspace(0, 1, len(sorted_h)))
for i, (name, res) in enumerate(sorted_h):
ax1.plot(res["fpr"], res["tpr"], color=cmap[i], lw=1.5,
label=f'{name} ({res["roc_auc"]:.3f})')
ax2.plot(res["recall"], res["precision"], color=cmap[i], lw=1.5,
label=f'{name} ({res["avg_precision"]:.3f})')
ax1.plot([0, 1], [0, 1], "k--", lw=1, alpha=0.3)
ax1.set_xlabel("FPR"); ax1.set_ylabel("TPR")
ax1.set_title("ROC Curves (holdout)"); ax1.legend(fontsize=8, loc="lower right")
ax1.grid(alpha=0.3)
ax2.set_xlabel("Recall"); ax2.set_ylabel("Precision")
ax2.set_title("PR Curves (holdout)"); ax2.legend(fontsize=8, loc="lower left")
ax2.grid(alpha=0.3)
fig.tight_layout()
fig.savefig(OUT_DIR / "classifier_comparison_curves.png", dpi=200, bbox_inches="tight")
plt.close()
print(" Saved classifier_comparison_curves.png")
# --- Summary ---
print(f"\n{'=' * 70}")
print("SUMMARY (sorted by CV ROC AUC)")
print(f"{'=' * 70}")
print(f"{'Classifier':<25} {'CV AUC':>15} {'CV AP':>15} {'CV Acc':>15} {'Holdout AUC':>12}")
print("-" * 80)
for name in names:
cv = valid[name]
h_auc = holdout_results.get(name, {}).get("roc_auc", float("nan"))
print(f"{name:<25} "
f"{cv['roc_auc']['mean']:.4f}+/-{cv['roc_auc']['std']:.4f} "
f"{cv['avg_precision']['mean']:.4f}+/-{cv['avg_precision']['std']:.4f} "
f"{cv['accuracy']['mean']:.4f}+/-{cv['accuracy']['std']:.4f} "
f"{h_auc:>10.4f}")
# Save JSON
save_results = {}
for name in names:
cv = valid[name]
h = holdout_results.get(name, {})
save_results[name] = {
"cv": {m: {"mean": cv[m]["mean"], "std": cv[m]["std"]}
for m in ["roc_auc", "avg_precision", "accuracy", "f1"]},
"holdout": {"roc_auc": h.get("roc_auc"), "avg_precision": h.get("avg_precision")},
}
with open(OUT_DIR / "classifier_comparison_results.json", "w") as f:
json.dump(save_results, f, indent=2)
print(f"\nSaved classifier_comparison_results.json")
best = names[0]
baseline_auc = valid["LogReg (C=1)"]["roc_auc"]["mean"]
best_auc = valid[best]["roc_auc"]["mean"]
print(f"\nBaseline LogReg (C=1) AUC: {baseline_auc:.4f}")
print(f"Best: {best} (AUC: {best_auc:.4f}, delta: {best_auc - baseline_auc:+.4f})")
if __name__ == "__main__":
main()