This commit is contained in:
mhilya 2026-06-09 01:46:27 +07:00
parent b5d5a9b63a
commit 2f52a2e70c
10 changed files with 1572 additions and 223 deletions

View File

@ -8,7 +8,7 @@ code_to_name = {}
with open('wilayah_dapodik_final.csv', mode='r', encoding='utf-8') as f:
reader = csv.DictReader(f)
for row in reader:
# Bersihkan nama kabupaten/kota dan provinsi agar pencarian lebih akurat
kab = row['Kabupaten_Kota'].replace('Kab. ', '').replace('Kota Adm. ', '').replace('Kota ', '')
prov = row['Provinsi'].replace('Prov. ', '')
@ -20,7 +20,6 @@ with open('wilayah_dapodik_final.csv', mode='r', encoding='utf-8') as f:
if kab == 'Adm. Kep. Seribu':
kab = 'Kepulauan Seribu'
# Format query pencarian, contoh: "Malang, Jawa Timur, Indonesia"
if prov == 'Luar Negeri':
query = kab
else:
@ -35,22 +34,20 @@ coords = {}
print(f"Mulai mencari koordinat untuk {len(code_to_name)} wilayah...")
# Looping melalui semua data yang ada di CSV
for code, info in code_to_name.items():
query = info['query']
name = info['name']
max_retries = 3
for attempt in range(max_retries):
try:
time.sleep(1) # Jeda 1 detik agar tidak terkena limit API Nominatim
# Menambahkan parameter timeout eksplisit jika diperlukan, default biasanya 1 detik
time.sleep(1)
location = geolocator.geocode(query, timeout=5)
if location:
coords[code] = {"name": name, "coords": [location.latitude, location.longitude]}
print(f"Geocoded: {code} -> {query} -> {coords[code]['coords']}")
else:
print(f"NOT FOUND: {code} -> {query}")
break # Berhasil diproses (entah ketemu atau tidak ketemu), hentikan loop retry
break
except GeocoderTimedOut:
print(f"TIMEOUT (Percobaan {attempt + 1}/{max_retries}): {code} -> {query}")
if attempt == max_retries - 1:

806
data/prepare_corpus.ipynb Normal file

File diff suppressed because one or more lines are too long

View File

@ -63,8 +63,6 @@ def clean_text(text: str) -> str:
tokens = [w for w in text.split() if w not in COMPANY_STOPWORDS and len(w) >= 3]
return " ".join([stemmer.stem(w) for w in tokens])
# [FIX v4] Logika lebih ketat: hanya flag jika token teks sepenuhnya subset dari nama
# (max 3 token) — bukan sekadar rasio overlap 50% seperti v3 yang rawan false positive.
def is_likely_name(text: str, full_name: str) -> bool:
if not text or not full_name: return False
text_clean = re.sub(r'[^\w\s]', '', text.lower())
@ -110,7 +108,6 @@ def main():
mask = ~df[col_status].str.lower().str.strip().isin(blacklist)
df = df[mask].copy()
# Stemming berat terjadi HANYA di sini — satu kali saat corpus preparation
df["job_text_raw"] = df[col_jabatan].apply(clean_text)
if col_klasifikasi:
@ -153,7 +150,6 @@ def main():
print("DATA PELATIHAN (TRAINING SET):")
print(train_df["label"].value_counts().to_string())
# [KEEP v3] Distribusi per kelas di test set — berguna untuk deteksi class imbalance
print("\nDATA PENGUJIAN (TEST SET):")
for cls in TARGET_CLASSES:
c = test_df["label"].value_counts().get(cls, 0)

753
data/train_model.ipynb Normal file

File diff suppressed because one or more lines are too long

View File

@ -1,181 +1,3 @@
# import pandas as pd
# import numpy as np
# import re
# import joblib
# import json
# import warnings
# from pathlib import Path
# from sklearn.feature_extraction.text import TfidfVectorizer
# from sklearn.linear_model import LogisticRegression
# from sklearn.model_selection import StratifiedKFold
# from sklearn.metrics import classification_report, confusion_matrix, ConfusionMatrixDisplay
# from sklearn.pipeline import Pipeline
# import matplotlib
# matplotlib.use('Agg')
# import matplotlib.pyplot as plt
# warnings.filterwarnings('ignore')
# BASE_DIR = Path(__file__).parent
# TRAIN_FILE = BASE_DIR / "processed" / "training_corpus.csv"
# TEST_FILE = BASE_DIR / "processed" / "test_set.csv"
# ML_DIR = BASE_DIR.parent / "fastapi" / "ml_assets"
# ML_DIR.mkdir(parents=True, exist_ok=True)
# TARGET_CLASSES = ["Programmer", "Data Analyst", "Wirausaha Informatika", "Non-IT"]
# CONFIDENCE_THRESHOLD = 0.50
# # [v4] Tidak import Sastrawi — corpus sudah di-stem oleh prepare_corpus.py
# # Stemming dua kali pada data yang sama dapat mendegradasi teks.
# STOPWORDS = {
# "yang", "di", "ke", "dari", "dan", "atau", "dengan", "untuk", "pada", "dalam",
# "adalah", "ini", "itu", "tidak", "juga", "sudah", "akan", "bisa", "ada", "oleh",
# "karena", "secara", "serta", "sebagai", "bagi", "telah", "maka", "namun", "sehingga",
# "jika", "agar", "ketika", "saat", "sebelum", "sesudah", "hingga", "sampai", "antara",
# "sekitar", "hanya", "saja", "belum", "masih", "lagi", "pun", "justru", "walaupun",
# "meskipun", "bahkan", "cukup", "sangat", "paling", "lebih", "kurang", "lain",
# "macam", "cara", "hal", "tentang", "mengenai", "terhadap", "kepada", "menuju",
# "kecuali", "selain", "tanpa", "demi", "guna", "khususnya", "umumnya", "kebanyakan",
# "sebagian", "beberapa", "semua", "setiap", "tiap", "satu", "dua", "tiga", "empat",
# "lima", "enam", "tujuh", "delapan", "sembilan", "sepuluh", "ratus", "ribu", "juta"
# }
# def preprocess_text(text):
# """Pembersihan dasar — tanpa stemming ulang karena korpus sudah di-stem."""
# if pd.isna(text) or not isinstance(text, str): return ""
# text = text.lower().strip()
# if text in {"nan", "none", "null", "-", "0", "", "tidak diisi"}: return ""
# text = text.replace('-', ' ').replace('/', ' ').replace('_', ' ').replace(',', ' ')
# text = re.sub(r'[^\w\s]', '', text)
# text = re.sub(r'\d+', '', text)
# words = [w for w in text.split() if w not in STOPWORDS and len(w) > 2]
# return " ".join(words)
# def main():
# if not TRAIN_FILE.exists():
# print("File training_corpus.csv belum ada. Jalankan prepare_corpus.py terlebih dahulu."); return
# df_train = pd.read_csv(TRAIN_FILE, sep=";", dtype=str).dropna(subset=["job_text_raw", "label"])
# df_test = pd.read_csv(TEST_FILE, sep=";", dtype=str).dropna(subset=["job_text_raw", "label"]) if TEST_FILE.exists() else None
# print("Sedang memproses teks (cleaning dasar tanpa stemming redundan)...")
# X_train = df_train["job_text_raw"].apply(preprocess_text)
# y_train = df_train["label"]
# mask = X_train.str.len() > 0
# X_train, y_train = X_train[mask], y_train[mask]
# X_test, y_test = pd.Series(dtype=str), pd.Series(dtype=str)
# if df_test is not None:
# X_test = df_test["job_text_raw"].apply(preprocess_text)
# y_test = df_test["label"]
# mask_t = X_test.str.len() > 0
# X_test, y_test = X_test[mask_t], y_test[mask_t]
# print(f"Jumlah data latih: {len(X_train)} baris | Data uji: {len(X_test)} baris\n")
# # [KEEP v3] K-Fold cross validation — estimasi variance performa model
# print("Melakukan pengujian K-Fold (5 putaran) pada data latih...")
# skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
# fold_metrics = []
# for i, (tr_idx, te_idx) in enumerate(skf.split(X_train, y_train)):
# model = Pipeline([
# ('tfidf', TfidfVectorizer(max_features=3000, ngram_range=(1,2), sublinear_tf=True, min_df=1)),
# ('clf', LogisticRegression(max_iter=1000, class_weight='balanced', solver='lbfgs'))
# ])
# model.fit(X_train.iloc[tr_idx], y_train.iloc[tr_idx])
# y_pred = model.predict(X_train.iloc[te_idx])
# rep = classification_report(y_train.iloc[te_idx], y_pred, output_dict=True, zero_division=0)
# fold_metrics.append(rep)
# print(f" Akurasi putaran ke-{i+1}: {rep['accuracy']:.4f}")
# avg_acc = np.mean([f['accuracy'] for f in fold_metrics])
# print(f"\n RATA-RATA PENGUJIAN K-FOLD:")
# print(f" Akurasi Keseluruhan : {avg_acc:.4f} ± {np.std([f['accuracy'] for f in fold_metrics]):.4f}")
# for cls in TARGET_CLASSES:
# p = np.mean([f.get(cls, {}).get('precision', 0) for f in fold_metrics])
# r = np.mean([f.get(cls, {}).get('recall', 0) for f in fold_metrics])
# f1 = np.mean([f.get(cls, {}).get('f1-score', 0) for f in fold_metrics])
# print(f" {cls:25} | P: {p:.3f} | R: {r:.3f} | F1: {f1:.3f}")
# print("\n Membuat model final dari seluruh data latih yang tersedia...")
# final_model = Pipeline([
# ('tfidf', TfidfVectorizer(max_features=3000, ngram_range=(1,2), sublinear_tf=True, min_df=1)),
# ('clf', LogisticRegression(max_iter=1000, class_weight='balanced', solver='lbfgs'))
# ])
# final_model.fit(X_train, y_train)
# metrics_test = None
# if len(X_test) > 0:
# y_pred = final_model.predict(X_test)
# rep_test = classification_report(y_test, y_pred, output_dict=True, zero_division=0)
# metrics_test = rep_test
# print("\n HASIL PENGUJIAN PADA DATA TEST:")
# print(classification_report(y_test, y_pred, zero_division=0))
# for cls in TARGET_CLASSES:
# sup = rep_test[cls]['support'] if cls in rep_test else 0
# if sup < 5: print(f"Catatan: Kelas '{cls}' cuma punya {sup} data uji — skor F1-nya mungkin kurang akurat.")
# # [KEEP v3] Confusion matrix
# cm = confusion_matrix(y_test, y_pred, labels=TARGET_CLASSES)
# disp = ConfusionMatrixDisplay(confusion_matrix=cm, display_labels=TARGET_CLASSES)
# disp.plot(cmap='Blues', values_format='d', xticks_rotation=45, colorbar=False)
# plt.title('Confusion Matrix - Hold-Out Test'); plt.tight_layout()
# plt.savefig(ML_DIR / 'confusion_matrix_test.png', dpi=300); plt.close()
# print("Grafik confusion matrix berhasil disimpan ke confusion_matrix_test.png")
# # [BUGFIX] Feature importance per kelas
# # clf.coef_ diurutkan mengikuti clf.classes_ (alfabetikal sklearn),
# # BUKAN urutan TARGET_CLASSES yang hardcoded. Pakai clf.classes_ sebagai
# # iterator agar index i selalu selaras dengan baris koefisien yang benar.
# vec, clf = final_model.named_steps['tfidf'], final_model.named_steps['clf']
# feats, coeffs = vec.get_feature_names_out(), clf.coef_
# print("\n5 KATA PALING BERPENGARUH UNTUK TIAP KELAS:")
# for i, cls in enumerate(clf.classes_): # <-- clf.classes_, bukan TARGET_CLASSES
# top = [(feats[j], coeffs[i][j]) for j in coeffs[i].argsort()[-5:][::-1] if coeffs[i][j] > 0]
# print(f" [{cls}] " + ", ".join([f"{w}({c:.2f})" for w, c in top]))
# # [KEEP v3] Confidence distribution
# proba = final_model.predict_proba(X_test)
# max_p = np.max(proba, axis=1)
# plt.hist(max_p, bins=15, edgecolor='black', alpha=0.7, color='teal')
# plt.axvline(x=CONFIDENCE_THRESHOLD, color='red', linestyle='--', linewidth=2, label=f'Threshold {CONFIDENCE_THRESHOLD}')
# plt.xlabel('Confidence Score'); plt.ylabel('Count'); plt.title('Distribusi Confidence Score')
# plt.legend(); plt.grid(axis='y', alpha=0.3); plt.tight_layout()
# plt.savefig(ML_DIR / 'confidence_distribution.png', dpi=300); plt.close()
# below = np.sum(max_p < CONFIDENCE_THRESHOLD)
# print(f"\nPengecekan Keyakinan Model: {below} dari {len(max_p)} prediksi ({below/len(max_p)*100:.1f}%) di bawah threshold {CONFIDENCE_THRESHOLD}.")
# # [KEEP v3] Daftar prediksi yang meleset
# mis = np.where(y_test.values != y_pred)[0]
# if len(mis) > 0:
# print("\n DAFTAR PREDIKSI YANG MELESET:")
# for idx in mis[:min(6, len(mis))]:
# txt = df_test.iloc[idx]['job_text_raw'][:80]
# print(f" • Seharusnya: {y_test.iloc[idx]:20} | Ditebak: {y_pred[idx]:20} | Teks: '{txt}...'")
# else:
# print("\n Semua prediksi pada test set benar.")
# model_path = ML_DIR / "ml_pipeline_internal.pkl"
# joblib.dump(final_model, model_path)
# metrics = {
# "methodology": "Internal MIF only",
# "k_fold": {
# "accuracy_mean": float(avg_acc),
# "accuracy_std": float(np.std([f['accuracy'] for f in fold_metrics])),
# "folds": fold_metrics
# },
# "threshold_config": CONFIDENCE_THRESHOLD
# }
# if metrics_test: metrics["hold_out_test"] = metrics_test
# with open(ML_DIR / "metrics_internal_only.json", 'w') as f: json.dump(metrics, f, indent=2)
# print(f"\n Model berhasil disimpan di: {model_path}")
# print(f"Laporan metrik disimpan di: {ML_DIR / 'metrics_internal_only.json'}")
# if __name__ == "__main__":
# main()
import pandas as pd
import numpy as np
import re
@ -192,9 +14,6 @@ import matplotlib
matplotlib.use('Agg')
import matplotlib.pyplot as plt
# PERBAIKAN 1: Hapus warnings.filterwarnings('ignore') secara global.
# Biarkan warning muncul. Jika ada ConvergenceWarning, Anda akan tahu bahwa max_iter perlu dinaikkan lagi.
BASE_DIR = Path(__file__).parent
TRAIN_FILE = BASE_DIR / "processed" / "training_corpus.csv"
TEST_FILE = BASE_DIR / "processed" / "test_set.csv"
@ -223,10 +42,6 @@ def preprocess_text(text):
if text in {"nan", "none", "null", "-", "0", "", "tidak diisi"}: return ""
text = text.replace('-', ' ').replace('/', ' ').replace('_', ' ').replace(',', ' ')
text = re.sub(r'[^\w\s]', '', text)
# PERBAIKAN 2: Penambahan komentar eksplisit.
# Anda harus sadar bahwa menghapus angka menghilangkan sinyal seperti "Python 3" atau "5 tahun".
# Jika ini keputusan sadar, biarkan. Jika tidak, hapus baris di bawah ini.
text = re.sub(r'\d+', '', text)
words = [w for w in text.split() if w not in STOPWORDS and len(w) > 2]
@ -239,14 +54,14 @@ def get_pipeline():
max_features=3000,
ngram_range=(1, 2),
sublinear_tf=True,
min_df=2, # PERBAIKAN 3: Diubah dari 1 ke 2 untuk mencegah overfitting pada noise/typo.
norm='l2' # PERBAIKAN 4: Eksplisit menyatakan normalisasi L2 (default, tapi ditulis agar transparan).
min_df=2,
norm='l2'
)),
('clf', LogisticRegression(
max_iter=2000, # PERBAIKAN 5: Dinaikkan dari 1000 untuk mencegah ConvergenceWarning pada data sparse.
max_iter=2000,
class_weight='balanced',
solver='lbfgs',
random_state=42 # PERBAIKAN 6: Ditambahkan untuk reproducibility (hasil yang konsisten).
random_state=42
))
])
@ -279,7 +94,6 @@ def main():
fold_metrics = []
for i, (tr_idx, te_idx) in enumerate(skf.split(X_train, y_train)):
# PERBAIKAN 7: Menggunakan fungsi get_pipeline() untuk menghindari duplikasi kode dan inkonsistensi
model = get_pipeline()
model.fit(X_train.iloc[tr_idx], y_train.iloc[tr_idx])
y_pred = model.predict(X_train.iloc[te_idx])
@ -297,7 +111,6 @@ def main():
print(f" {cls:25} | P: {p:.3f} | R: {r:.3f} | F1: {f1:.3f}")
print("\n Membuat model final dari seluruh data latih yang tersedia...")
# PERBAIKAN 8: Menggunakan fungsi get_pipeline()
final_model = get_pipeline()
final_model.fit(X_train, y_train)
@ -325,7 +138,6 @@ def main():
plt.close()
print("Grafik confusion matrix berhasil disimpan ke confusion_matrix_test.png")
# PERBAIKAN 9: Feature importance yang menampilkan koefisien POSITIF dan NEGATIF
vec = final_model.named_steps['tfidf']
clf = final_model.named_steps['clf']
feats = vec.get_feature_names_out()

View File

@ -138,7 +138,6 @@ def classify_ml(job_text: str, pipeline) -> dict:
logger.error(f"ML_INFER_ERROR | job_text='{job_text[:50]}' | err={e}")
raise HTTPException(status_code=500, detail=f"ML inference failed: {str(e)}")
# [KEEP v3] Ekstraksi teks dari format Kemendikbud
def extract_kemendik_text(row: pd.Series, df_columns: list) -> str:
f5b_col = find_column_by_code(df_columns, "f5b")
f5c_col = find_column_by_code(df_columns, "f5c")
@ -175,7 +174,6 @@ def health_check():
@app.post("/api/v1/retrain", dependencies=[Depends(verify_token)])
async def trigger_retrain(file: UploadFile = File(None)):
# [v4] FileLock mencegah race condition dari concurrent requests
lock = FileLock(LOCK_PATH, timeout=2)
try:
with lock:
@ -221,7 +219,6 @@ def retrain_status():
if not STATUS_PATH.exists():
return JSONResponse(content={"stage": "idle", "message": "Belum ada proses re-training."})
try:
# [v4] Lock saat baca untuk menghindari partial write dari worker
lock = FileLock(LOCK_PATH, timeout=2)
with lock:
data = json.loads(STATUS_PATH.read_text(encoding="utf-8"))
@ -241,7 +238,6 @@ def retrain_status():
return JSONResponse(content={"stage": "unknown", "message": "Status tidak terbaca."})
# [KEEP v3] Endpoint reload manual — berguna jika pipeline perlu di-reload tanpa restart server
@app.post("/api/v1/retrain/reload", dependencies=[Depends(verify_token)])
def reload_model():
try:

View File

@ -11,7 +11,6 @@ Alur:
6. Evaluasi candidate pada test set yang sama delta terhadap old model
7. Promote jika lebih baik; rollback jika tidak
8. Setiap write status dilindungi FileLock (atomic)
9. [v5] Semua event log disimpan ke file JSONL terpisah untuk audit BAB 4
"""
import sys
import json
@ -42,7 +41,7 @@ warnings.filterwarnings("ignore")
BASE_DIR = Path(__file__).parent.parent
ML_DIR = BASE_DIR / "ml_assets"
DATA_DIR = BASE_DIR.parent / "data" / "processed"
LOG_DIR = ML_DIR / "logs" # [v5] Direktori log terpisah
LOG_DIR = ML_DIR / "logs"
LOG_DIR.mkdir(parents=True, exist_ok=True)
PIPELINE_PATH = ML_DIR / "ml_pipeline_internal.pkl"
@ -54,7 +53,6 @@ LOCK_PATH = ML_DIR / "retrain_status.lock"
CORPUS_PATH = DATA_DIR / "training_corpus.csv"
TEST_PATH = DATA_DIR / "test_set.csv"
# [v5] Nama file log berbasis timestamp — satu file per sesi retraining
SESSION_TS = datetime.now().strftime("%Y%m%d_%H%M%S")
LOG_FILE_PATH = LOG_DIR / f"retrain_log_{SESSION_TS}.jsonl"
@ -81,7 +79,7 @@ STOPWORDS = {
stemmer = StemmerFactory().create_stemmer()
# ──────────────────────────────────────────────────────────────
# [v5] CUSTOM JSON FORMATTER — setiap log entry jadi JSON valid
# CUSTOM JSON FORMATTER
# ──────────────────────────────────────────────────────────────
class JsonFormatter(logging.Formatter):
"""Format log record sebagai JSON satu baris (JSONL)."""
@ -92,7 +90,6 @@ class JsonFormatter(logging.Formatter):
"logger": record.name,
"message": record.getMessage(),
}
# [v5] Jika ada extra data (stage, metrics, dll), gabungkan ke root
if hasattr(record, "stage"):
log_entry["stage"] = record.stage
if hasattr(record, "event_type"):
@ -102,7 +99,7 @@ class JsonFormatter(logging.Formatter):
return json.dumps(log_entry, ensure_ascii=False)
# ──────────────────────────────────────────────────────────────
# [v5] DUAL LOGGER SETUP — console + file JSONL
# DUAL LOGGER SETUP
# ──────────────────────────────────────────────────────────────
logger = logging.getLogger("retrain_worker")
logger.setLevel(logging.INFO)
@ -123,7 +120,7 @@ logger.addHandler(file_handler)
logger.info(f"[init] Log file tersimpan di: {LOG_FILE_PATH}")
# ──────────────────────────────────────────────────────────────
# [v5] STRUCTURED LOG HELPER — untuk event penting dengan metadata
# STRUCTURED LOG HELPER
# ──────────────────────────────────────────────────────────────
def log_event(event_type: str, message: str, stage: str = "info", metadata: dict = None):
"""
@ -148,7 +145,7 @@ def compute_override_weight(n_corpus: int, n_override: int) -> float:
return round(w_log, 4)
# ──────────────────────────────────────────────────────────────
# STATUS WRITER — atomic dengan FileLock (v4)
# STATUS WRITER
# ──────────────────────────────────────────────────────────────
def write_status(stage: str, message: str, extra: dict = None):
lock = FileLock(LOCK_PATH, timeout=10)
@ -156,7 +153,6 @@ def write_status(stage: str, message: str, extra: dict = None):
payload = {"stage": stage, "message": message, "timestamp": datetime.now().isoformat()}
if extra: payload.update(extra)
STATUS_PATH.write_text(json.dumps(payload, indent=2, ensure_ascii=False), encoding="utf-8")
# [v5] Tulis juga ke log file dengan stage tracking
logger.info(f"[{stage}] {message}", extra={"stage": stage})
# ──────────────────────────────────────────────────────────────
@ -228,7 +224,6 @@ def do_rollback(reason: str, old_f1: float, new_f1: float):
else:
logger.warning("File .bak tidak ditemukan — pkl aktif dibiarkan.")
# [v5] Log event rollback dengan metadata lengkap
log_event(
event_type="rollback",
message=f"Model lama dipertahankan. {reason}",
@ -246,7 +241,6 @@ def do_rollback(reason: str, old_f1: float, new_f1: float):
# MAIN
# ──────────────────────────────────────────────────────────────
def main(extra_csv_path: str = None):
# [v5] Log awal sesi — penting untuk audit trail
log_event(
event_type="session_start",
message=f"Retraining session dimulai. Log file: {LOG_FILE_PATH.name}",
@ -314,7 +308,6 @@ def main(extra_csv_path: str = None):
y_all = df_all["label"]
w_all = df_all["_weight"]
# [v5] Log event data merge dengan metadata
log_event(
event_type="data_merged",
message=f"Data training siap: {len(df_all)} baris",
@ -342,7 +335,7 @@ def main(extra_csv_path: str = None):
write_status("training", f"K-Fold validation & training... ({len(X_train)} sampel)")
skf = StratifiedKFold(n_splits=5, shuffle=True, random_state=42)
fold_metrics = []
fold_accuracies = [] # [v5] untuk log ringkas
fold_accuracies = []
for i, (tr_idx, te_idx) in enumerate(skf.split(X_train, y_train)):
m = Pipeline([
('tfidf', TfidfVectorizer(max_features=3000, ngram_range=(1, 2), sublinear_tf=True, min_df=1)),
@ -358,7 +351,6 @@ def main(extra_csv_path: str = None):
fold_accuracies.append(round(rep['accuracy'], 4))
write_status("training", f"K-Fold selesai: fold {i+1}/5 | acc={rep['accuracy']:.4f}")
# [v5] Log ringkas hasil K-Fold
log_event(
event_type="kfold_completed",
message=f"K-Fold 5 lipatan selesai. Akurasi per fold: {fold_accuracies}",
@ -405,7 +397,6 @@ def main(extra_csv_path: str = None):
f"minimum dibutuhkan: +{MIN_IMPROVEMENT_THRESHOLD*100:.1f}%)"
)
# [v5] Log event evaluasi dengan metadata lengkap — ini yang akan dikutip di BAB 4
log_event(
event_type="evaluation_completed",
message=f"Evaluasi selesai. Delta F1: {delta:+.4f} ({'promote' if should_promote else 'rollback'})",
@ -460,7 +451,6 @@ def main(extra_csv_path: str = None):
}
METRICS_PATH.write_text(json.dumps(new_metrics_full, indent=2, ensure_ascii=False), encoding="utf-8")
# [v5] Log event promote final
log_event(
event_type="model_promoted",
message=f"Model baru berhasil dipromote. {reason}",
@ -492,7 +482,6 @@ def main(extra_csv_path: str = None):
except Exception as e:
logger.error(f"ERROR saat training: {type(e).__name__}: {e}", exc_info=True)
# [v5] Log event error dengan traceback info
log_event(
event_type="training_error",
message=f"Training gagal: {type(e).__name__}: {str(e)[:200]}",

Binary file not shown.

Before

Width:  |  Height:  |  Size: 62 KiB

After

Width:  |  Height:  |  Size: 92 KiB

Binary file not shown.

Before

Width:  |  Height:  |  Size: 114 KiB

After

Width:  |  Height:  |  Size: 137 KiB