Spaces:
Sleeping
Sleeping
File size: 17,822 Bytes
7f105c8 | 1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 130 131 132 133 134 135 136 137 138 139 140 141 142 143 144 145 146 147 148 149 150 151 152 153 154 155 156 157 158 159 160 161 162 163 164 165 166 167 168 169 170 171 172 173 174 175 176 177 178 179 180 181 182 183 184 185 186 187 188 189 190 191 192 193 194 195 196 197 198 199 200 201 202 203 204 205 206 207 208 209 210 211 212 213 214 215 216 217 218 219 220 221 222 223 224 225 226 227 228 229 230 231 232 233 234 235 236 237 238 239 240 241 242 243 244 245 246 247 248 249 250 251 252 253 254 255 256 257 258 259 260 261 262 263 264 265 266 267 268 269 270 271 272 273 274 275 276 277 278 279 280 281 282 283 284 285 286 287 288 289 290 291 292 293 294 295 296 297 298 299 300 301 302 303 304 305 306 307 308 309 310 311 312 313 314 315 316 317 318 319 320 321 322 323 324 325 326 327 328 329 330 331 332 333 334 335 336 337 338 339 340 341 342 343 344 345 346 347 348 349 350 351 352 353 354 355 356 357 358 359 360 361 362 363 364 365 366 367 368 369 370 371 372 373 374 375 376 377 378 379 380 381 382 383 384 385 386 387 388 389 390 391 392 393 394 395 396 397 398 399 400 401 402 403 404 405 406 407 408 409 410 | """
evaluate_compare_ensemble_20260505.py
======================================
담당: 경이 (kyeongyi)
작성일: 2026-05-05
목적:
Simple + KcELECTRA 소프트 투표 앙상블 평가.
KcELECTRA 단독(v3, Macro F1 0.8545)이 목표 +5%에 0.71%p 미달하여
devlog 우선순위 3에 따라 앙상블 시도.
[비교 모델]
1. Simple : TF-IDF + Logistic Regression (baseline 0.8116)
2. KcELECTRA: v3 파인튜닝 단독 (0.8545)
3. Ensemble : KcELECTRA×0.7 + Simple×0.3 (목표: 0.8616+)
[앙상블 방식]
소프트 투표(Soft Voting) — 두 모델의 확률을 가중 합산 후 argmax
combined[i] = weight_kc × kc_prob[i] + (1-weight_kc) × s_prob[i]
[출력 파일 - data/20260505/]
eval_results_ensemble_20260505.json
eval_comparison_summary_ensemble_20260505.csv
실행:
cd model/classification
python scripts/evaluate_compare_ensemble_20260505.py
python scripts/evaluate_compare_ensemble_20260505.py --weight_kc 0.6
python scripts/evaluate_compare_ensemble_20260505.py --search_weights
"""
import argparse
import json
import pickle
import sys
from pathlib import Path
import numpy as np
import pandas as pd
from sklearn.metrics import classification_report, confusion_matrix, f1_score
from sklearn.pipeline import Pipeline
_BASE = Path(__file__).parent.parent
sys.path.insert(0, str(_BASE / "src"))
SPLIT_CSV = _BASE / "data" / "split_v5_20260505.csv"
SIMPLE_PKL = _BASE / "checkpoints" / "simple_tfidf_logreg_v3_20260505.pkl"
KCELECTRA_CKPT = _BASE / "checkpoints" / "kcelectra-category-v3"
OUT_DIR = _BASE / "data" / "20260505"
# HF Hub fallback (로컬 체크포인트 없을 때)
_HF_REPO = "kysophia/kcelectra-category"
_HF_SUBFOLDER = "kcelectra-category-v3"
LABELS = ["일정", "준비물", "제출", "비용", "건강·안전", "기타"]
_LABEL_TO_COL = {lbl: i for i, lbl in enumerate(LABELS)}
# ──────────────────────────────────────────────────────────────────
# 데이터 로드
# ──────────────────────────────────────────────────────────────────
def load_split(split: str = "test") -> tuple[list[str], list[str]]:
if not SPLIT_CSV.exists():
raise FileNotFoundError(f"{SPLIT_CSV} 없음")
df = pd.read_csv(SPLIT_CSV, encoding="utf-8-sig")
df = df[df["split"] == split]
df = df[df["category"].isin(LABELS)]
return df["text"].tolist(), df["category"].tolist()
# ──────────────────────────────────────────────────────────────────
# Simple 확률 행렬 (N, 6)
# ──────────────────────────────────────────────────────────────────
def _load_simple() -> Pipeline:
if not SIMPLE_PKL.exists():
raise FileNotFoundError(
f"{SIMPLE_PKL.name} 없음.\n"
" 먼저 실행: python scripts/evaluate_compare_v3_20260505.py"
)
with open(SIMPLE_PKL, "rb") as f:
return pickle.load(f)
def get_simple_proba(texts: list[str]) -> np.ndarray:
"""Simple predict_proba → LABELS 순서로 열 정렬한 (N, 6) 행렬."""
pipe = _load_simple()
raw = pipe.predict_proba(texts) # (N, K), pipe.classes_ 순서
classes = list(pipe.classes_)
out = np.zeros((len(texts), len(LABELS)))
for j, lbl in enumerate(LABELS):
if lbl in classes:
out[:, j] = raw[:, classes.index(lbl)]
return out
# ──────────────────────────────────────────────────────────────────
# KcELECTRA 확률 행렬 (N, 6)
# ──────────────────────────────────────────────────────────────────
def _load_kcelectra_model():
"""(device, tokenizer, model, id2label) 반환. 로컬 우선, HF Hub fallback."""
try:
import torch
from transformers import AutoModelForSequenceClassification, AutoTokenizer
except ImportError:
raise ImportError("pip install torch transformers")
device = "cuda" if torch.cuda.is_available() else "cpu"
local_ready = (
KCELECTRA_CKPT.exists()
and (KCELECTRA_CKPT / "config.json").exists()
and any(
(KCELECTRA_CKPT / f).exists()
for f in ("pytorch_model.bin", "model.safetensors")
)
)
if local_ready:
tokenizer = AutoTokenizer.from_pretrained(str(KCELECTRA_CKPT))
model = AutoModelForSequenceClassification.from_pretrained(
str(KCELECTRA_CKPT), num_labels=len(LABELS), ignore_mismatched_sizes=True
)
src = str(KCELECTRA_CKPT)
else:
print(f"[kcelectra] 로컬 없음 → HF Hub 다운로드: {_HF_REPO}/{_HF_SUBFOLDER}")
tokenizer = AutoTokenizer.from_pretrained(_HF_REPO, subfolder=_HF_SUBFOLDER)
model = AutoModelForSequenceClassification.from_pretrained(
_HF_REPO, subfolder=_HF_SUBFOLDER, num_labels=len(LABELS)
)
src = f"{_HF_REPO}/{_HF_SUBFOLDER}"
model.to(device).eval()
print(f"[kcelectra] 모델 로드: {src} → device={device}")
labels_file = KCELECTRA_CKPT / "label2id.json"
if local_ready and labels_file.exists():
with open(labels_file, encoding="utf-8") as f:
label2id: dict[str, int] = json.load(f)
id2label = {v: k for k, v in label2id.items()}
else:
id2label = {i: lbl for i, lbl in enumerate(LABELS)}
return device, tokenizer, model, id2label
def get_kcelectra_proba(texts: list[str], batch_size: int = 32) -> np.ndarray:
"""KcELECTRA softmax → LABELS 순서로 열 정렬한 (N, 6) 행렬. 배치 처리."""
import torch
device, tokenizer, model, id2label = _load_kcelectra_model()
out = np.zeros((len(texts), len(LABELS)))
with torch.no_grad():
for start in range(0, len(texts), batch_size):
batch = texts[start : start + batch_size]
enc = tokenizer(
batch,
return_tensors="pt",
truncation=True,
padding=True,
max_length=128,
).to(device)
probs = torch.softmax(model(**enc).logits, dim=-1).cpu().numpy()
for b_i, prob_row in enumerate(probs):
for k_i, p in enumerate(prob_row):
lbl = id2label.get(k_i, "기타")
col = _LABEL_TO_COL.get(lbl, _LABEL_TO_COL["기타"])
out[start + b_i, col] = p
return out
# ──────────────────────────────────────────────────────────────────
# 앙상블 평가
# ──────────────────────────────────────────────────────────────────
def evaluate_ensemble(
split: str,
weight_kc: float,
s_proba: np.ndarray,
kc_proba: np.ndarray,
true_labels: list[str],
) -> dict:
"""소프트 투표. s_proba / kc_proba를 받아 가중 합산 → 분류 리포트 반환."""
combined = weight_kc * kc_proba + (1.0 - weight_kc) * s_proba
pred_idx = combined.argmax(axis=1)
pred_labels = [LABELS[i] for i in pred_idx]
report = classification_report(
true_labels, pred_labels,
labels=LABELS, output_dict=True, zero_division=0,
)
cm = confusion_matrix(true_labels, pred_labels, labels=LABELS)
macro_f1 = f1_score(true_labels, pred_labels, labels=LABELS,
average="macro", zero_division=0)
return {
"model": "ensemble",
"version": f"kc{weight_kc:.1f}+s{1-weight_kc:.1f}",
"weight_kc": weight_kc,
"macro_f1": round(macro_f1, 4),
"macro_precision": round(report["macro avg"]["precision"], 4),
"macro_recall": round(report["macro avg"]["recall"], 4),
"per_class": {
lbl: {
"precision": round(report[lbl]["precision"], 4),
"recall": round(report[lbl]["recall"], 4),
"f1": round(report[lbl]["f1-score"], 4),
"support": report[lbl]["support"],
}
for lbl in LABELS
},
"confusion_matrix": cm.tolist(),
"labels": LABELS,
"split_used": split,
"data_version": "v5_20260505",
}
# ──────────────────────────────────────────────────────────────────
# val 세트 가중치 그리드 탐색
# ──────────────────────────────────────────────────────────────────
def search_best_weight() -> float:
"""val 세트에서 weight_kc 0.4~0.9 탐색 → 최적 weight 반환."""
print("\n[weight 탐색] val 세트 기준 KcELECTRA 가중치 최적화")
val_texts, val_labels = load_split("val")
print(f" val {len(val_texts)}건 확률 계산 중...")
val_s = get_simple_proba(val_texts)
val_kc = get_kcelectra_proba(val_texts)
print(f"\n {'weight_kc':>10s} {'val Macro F1':>14s}")
print(" " + "-" * 28)
best_w, best_f1 = 0.7, 0.0
for w in [0.4, 0.5, 0.6, 0.7, 0.8, 0.9]:
res = evaluate_ensemble("val", w, val_s, val_kc, val_labels)
mark = " ← best" if res["macro_f1"] > best_f1 else ""
print(f" {w:>10.1f} {res['macro_f1']:>14.4f}{mark}")
if res["macro_f1"] > best_f1:
best_f1, best_w = res["macro_f1"], w
print(f"\n 최적 weight_kc = {best_w} (val Macro F1 = {best_f1:.4f})")
return best_w
# ──────────────────────────────────────────────────────────────────
# 저장 + 비교 출력
# ──────────────────────────────────────────────────────────────────
def _summary_row(res: dict) -> dict:
row = {
"macro_f1": res.get("macro_f1", "-"),
"macro_precision": res.get("macro_precision", "-"),
"macro_recall": res.get("macro_recall", "-"),
}
for lbl in LABELS:
row[f"f1_{lbl}"] = res.get("per_class", {}).get(lbl, {}).get("f1", "-")
return row
def _build_result_from_proba(
proba: np.ndarray,
true_labels: list[str],
model_name: str,
split: str,
) -> dict:
"""확률 행렬 → 분류 리포트 dict (JSON 없을 때 fallback)."""
pred_labels = [LABELS[i] for i in proba.argmax(axis=1)]
report = classification_report(true_labels, pred_labels,
labels=LABELS, output_dict=True, zero_division=0)
macro_f1 = f1_score(true_labels, pred_labels, labels=LABELS,
average="macro", zero_division=0)
cm = confusion_matrix(true_labels, pred_labels, labels=LABELS)
return {
"model": model_name,
"version": "v3",
"macro_f1": round(macro_f1, 4),
"macro_precision": round(report["macro avg"]["precision"], 4),
"macro_recall": round(report["macro avg"]["recall"], 4),
"per_class": {
lbl: {
"precision": round(report[lbl]["precision"], 4),
"recall": round(report[lbl]["recall"], 4),
"f1": round(report[lbl]["f1-score"], 4),
"support": report[lbl]["support"],
}
for lbl in LABELS
},
"confusion_matrix": cm.tolist(),
"labels": LABELS,
"split_used": split,
"data_version": "v5_20260505",
}
def save_and_compare(simple_res: dict, kc_res: dict, ens_res: dict) -> None:
OUT_DIR.mkdir(parents=True, exist_ok=True)
ens_path = OUT_DIR / "eval_results_ensemble_20260505.json"
with open(ens_path, "w", encoding="utf-8") as f:
json.dump(ens_res, f, ensure_ascii=False, indent=2)
print(f"\n[저장] {ens_path.name}")
label_e = f"Ensemble(kc{ens_res['weight_kc']:.1f}+s{1-ens_res['weight_kc']:.1f})"
rows = [
{"model": "Simple (TF-IDF+LR)", **_summary_row(simple_res)},
{"model": "KcELECTRA v3 (단독)", **_summary_row(kc_res)},
{"model": label_e, **_summary_row(ens_res)},
]
summary_path = OUT_DIR / "eval_comparison_summary_ensemble_20260505.csv"
pd.DataFrame(rows).to_csv(summary_path, index=False, encoding="utf-8-sig")
print(f"[저장] {summary_path.name}")
s_f1 = simple_res["macro_f1"]
kc_f1 = kc_res["macro_f1"]
e_f1 = ens_res["macro_f1"]
print("\n" + "=" * 60)
print(" 앙상블 성능 비교 (v5 데이터 - 4992행)")
print("=" * 60)
print(f" Simple Macro F1 : {s_f1:.4f} (baseline)")
print(f" KcELECTRA Macro F1 : {kc_f1:.4f} (Delta vs Simple: {kc_f1-s_f1:+.4f})")
print(f" Ensemble Macro F1 : {e_f1:.4f} (Delta vs Simple: {e_f1-s_f1:+.4f})")
print()
delta = e_f1 - s_f1
if delta >= 0.05:
print(" ★ 앙상블 5%+ 향상 달성! 채택 확정")
elif delta > 0:
print(f" ~ 앙상블 소폭 향상 ({delta:+.4f}) — --search_weights 또는 추가 튜닝 권장")
else:
print(" ✗ 앙상블이 Simple 미달 — weight_kc 조정 필요")
print("\n [카테고리별 F1 비교]")
print(f" {'카테고리':10s} {'Simple':>8s} {'KcELEC':>8s} {'Ensemble':>10s}")
print(" " + "-" * 42)
for lbl in LABELS:
s = simple_res["per_class"].get(lbl, {}).get("f1", 0.0)
k = kc_res["per_class"].get(lbl, {}).get("f1", 0.0)
e = ens_res["per_class"].get(lbl, {}).get("f1", 0.0)
best = " ★" if e > max(s, k) else (" " if e >= max(s, k) else " ")
print(f" {lbl:10s} {s:>8.4f} {k:>8.4f} {e:>10.4f}{best}")
print(f"\n[출력 폴더] {OUT_DIR}")
# ──────────────────────────────────────────────────────────────────
# CLI
# ──────────────────────────────────────────────────────────────────
def main() -> None:
parser = argparse.ArgumentParser(
description="KcELECTRA + Simple 소프트 투표 앙상블 평가"
)
parser.add_argument("--split", default="test",
choices=["train", "val", "test"])
parser.add_argument("--weight_kc", type=float, default=0.7,
help="KcELECTRA 가중치 (0.0~1.0). 기본값 0.7")
parser.add_argument("--search_weights", action="store_true",
help="val 세트에서 0.4~0.9 그리드 탐색 후 최적값으로 test 평가")
args = parser.parse_args()
print(f"앙상블 평가 시작 — split: {args.split}, weight_kc: {args.weight_kc}")
# ① val 세트 가중치 탐색 (--search_weights)
weight_kc = args.weight_kc
if args.search_weights:
weight_kc = search_best_weight()
print(f"\n[test 평가] 최적 weight_kc={weight_kc} 적용")
# ② test 확률 행렬
texts, true_labels = load_split(args.split)
print(f"\n[데이터] {args.split} 세트 {len(texts)}건")
print("\n[Simple] 확률 행렬 계산 중...")
s_proba = get_simple_proba(texts)
print("[KcELECTRA] 확률 행렬 계산 중... (CPU면 수 분 소요)")
kc_proba = get_kcelectra_proba(texts)
# ③ 앙상블 평가
ens_res = evaluate_ensemble(args.split, weight_kc, s_proba, kc_proba, true_labels)
# ④ 단독 결과 — 기존 JSON 재활용, 없으면 확률 행렬에서 직접 계산
simple_json = OUT_DIR / "eval_results_simple_v3_20260505.json"
kc_json = OUT_DIR / "eval_results_kcelectra_v3_20260505.json"
if simple_json.exists():
with open(simple_json, encoding="utf-8") as f:
simple_res = json.load(f)
print(f"[simple] 기존 JSON 재활용: {simple_json.name}")
else:
simple_res = _build_result_from_proba(s_proba, true_labels, "simple", args.split)
if kc_json.exists():
with open(kc_json, encoding="utf-8") as f:
kc_res = json.load(f)
print(f"[kcelectra] 기존 JSON 재활용: {kc_json.name}")
else:
kc_res = _build_result_from_proba(kc_proba, true_labels, "kcelectra", args.split)
# ⑤ 저장 + 비교 출력
print("\n[앙상블] 분류 리포트")
pred_labels = [LABELS[i] for i in (weight_kc * kc_proba + (1 - weight_kc) * s_proba).argmax(axis=1)]
print(classification_report(true_labels, pred_labels, labels=LABELS, zero_division=0))
save_and_compare(simple_res, kc_res, ens_res)
if __name__ == "__main__":
main()
|