scripts/validate_text_emotion_english.py

#!/usr/bin/env python3
"""영어 텍스트 감정 모델 (DistilRoBERTa) sanity check.

hand-crafted 영어 문장으로 모델 로딩, 출력 포맷, 기본 정확도를 검증한다.

Usage:
    python scripts/validate_text_emotion_english.py
"""

from __future__ import annotations

import logging
import sys
from pathlib import Path

PROJECT_ROOT = Path(__file__).parent.parent
sys.path.insert(0, str(PROJECT_ROOT))

logging.basicConfig(
    level=logging.INFO,
    format="%(asctime)s - %(levelname)s - %(message)s",
)
logger = logging.getLogger("validate_text_en")

PROJECT_LABELS = ["neutral", "joy", "sadness", "anger", "surprise", "fear", "disgust"]

# (text, expected_emotion) — 명확한 감정 표현 문장
TEST_CASES = [
    # joy
    ("I'm so happy today, everything is wonderful!", "joy"),
    ("This is the best day of my life!", "joy"),
    ("I'm thrilled about the good news!", "joy"),
    # anger
    ("This makes me absolutely furious!", "anger"),
    ("I can't believe how unfair this is, I'm so angry!", "anger"),
    ("Stop doing that, it's driving me crazy!", "anger"),
    # sadness
    ("I feel so sad and lonely right now.", "sadness"),
    ("I miss you so much, it hurts.", "sadness"),
    ("I can't stop crying, everything feels hopeless.", "sadness"),
    # fear
    ("I'm really scared, something is wrong.", "fear"),
    ("I'm terrified of what might happen next.", "fear"),
    ("Help me, I'm so afraid!", "fear"),
    # surprise
    ("Oh my god, I can't believe it!", "surprise"),
    ("Wow, I never expected that to happen!", "surprise"),
    ("What?! That's absolutely incredible!", "surprise"),
    # neutral
    ("The meeting is scheduled for 3pm.", "neutral"),
    ("I need to buy groceries on the way home.", "neutral"),
    ("The temperature today is around 20 degrees.", "neutral"),
    # disgust
    ("That's absolutely disgusting, I feel sick.", "disgust"),
    ("This food tastes terrible, it's revolting.", "disgust"),
    ("I can't stand the smell, it's nauseating.", "disgust"),
]


def validate():
    """DistilRoBERTa 모델 검증."""
    from src.stage2.text_emotion import predict as text_predict

    logger.info("=" * 60)
    logger.info("영어 텍스트 감정 모델 (DistilRoBERTa) 검증 시작")
    logger.info("=" * 60)

    passed = 0
    failed = 0
    results = []

    for text, expected in TEST_CASES:
        result = text_predict(text, language="en")

        # 출력 포맷 검증
        assert "emotion" in result, f"Missing 'emotion' key in result"
        assert "confidence" in result, f"Missing 'confidence' key in result"
        assert "scores" in result, f"Missing 'scores' key in result"

        # 모든 프로젝트 라벨이 scores에 있는지 확인
        for label in PROJECT_LABELS:
            assert label in result["scores"], f"Missing label '{label}' in scores"

        # scores 합계 ~1.0 확인
        score_sum = sum(result["scores"].values())
        assert abs(score_sum - 1.0) < 0.01, f"Scores sum={score_sum}, expected ~1.0"

        predicted = result["emotion"]
        confidence = result["confidence"]
        match = predicted == expected

        if match:
            passed += 1
            status = "PASS"
        else:
            failed += 1
            status = "FAIL"

        results.append({
            "text": text[:50],
            "expected": expected,
            "predicted": predicted,
            "confidence": confidence,
            "match": match,
        })

        logger.info(
            f"  [{status}] expected={expected:10s} predicted={predicted:10s} "
            f"conf={confidence:.3f} | \"{text[:45]}...\""
            if len(text) > 45 else
            f"  [{status}] expected={expected:10s} predicted={predicted:10s} "
            f"conf={confidence:.3f} | \"{text}\""
        )

    total = passed + failed
    accuracy = passed / total if total > 0 else 0

    logger.info(f"\n{'='*60}")
    logger.info(f"결과: {passed}/{total} PASS ({accuracy:.1%})")
    logger.info(f"  Passed: {passed}")
    logger.info(f"  Failed: {failed}")

    if accuracy >= 0.8:
        logger.info("판정: PASS — 모델이 정상적으로 동작합니다.")
    elif accuracy >= 0.6:
        logger.info("판정: WARN — 일부 감정에서 부정확합니다.")
    else:
        logger.info("판정: FAIL — 모델에 문제가 있을 수 있습니다.")

    logger.info(f"{'='*60}")

    return accuracy >= 0.6


if __name__ == "__main__":
    success = validate()
    sys.exit(0 if success else 1)