from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
import torch.nn.functional as F

# 1) Load Model & Tokenizer
MODEL_NAME = "LimYeri/HowRU-KoELECTRA-Emotion-Classifier"

tokenizer = AutoTokenizer.from_pretrained(MODEL_NAME)
model = AutoModelForSequenceClassification.from_pretrained(MODEL_NAME)

# GPU 사용 가능 시 자동 전환
device = torch.device("cuda" if torch.cuda.is_available() else "cpu")
model.to(device)
model.eval()

# 감정 라벨 매핑 (id2label)
id2label = model.config.id2label


# 2) Inference Function
def predict_emotion(text: str):
    """
    Returns:
        - top1_pred: 예측된 감정 라벨
        - probs_sorted: 감정별 확률(내림차순)
        - top2_pred: 상위 두 개의 감정
    """

    # 토크나이징
    inputs = tokenizer(
        text,
        return_tensors="pt",
        truncation=True,
        padding=True,
        max_length=512
    ).to(device)

    # 추론
    with torch.no_grad():
        logits = model(**inputs).logits
        probs = F.softmax(logits, dim=-1)[0]

    # 정렬된 확률
    probs_sorted = sorted(
        [(id2label[i], float(probs[i])) for i in range(len(probs))],
        key=lambda x: x[1],
        reverse=True
    )

    top1_pred = probs_sorted[0]
    top2_pred = probs_sorted[:2]

    return {
        "text": text,
        "top1_emotion": top1_pred,
        "top2_emotions": top2_pred,
        "all_probabilities": probs_sorted,
    }


# 3) Example
result = predict_emotion("오늘 정말 기분이 좋고 행복한 하루였어!")
print(result)

pipeline

python

from transformers import pipeline

MODEL_NAME = "LimYeri/HowRU-KoELECTRA-Emotion-Classifier"

classifier = pipeline(
    "text-classification",
    model=MODEL_NAME,
    tokenizer=MODEL_NAME,
    top_k=None   # 전체 감정 확률 반환
)

# 예측
text = "오늘 정말 기분이 좋고 행복한 하루였어!"
result = classifier(text)

result = result[0]

print("입력 문장:", text)
print("\nTop-1 감정:", result[0]['label'], f"({result[0]['score']:.4f})")
print("\n전체 감정 분포:")
for r in result:
    print(f"  {r['label']}: {r['score']:.4f}")

Training Details

Training Data

Total(8:2로 분할): 50,000행
Train: 40,000행
Validation: 10,000행

Training Procedure

Base Model: monologg/koelectra-base-v3-discriminator
Objective: Single-label classification
Max Length: 512

Training Hyperparameters

num_train_epochs: 3
learning_rate: 3e-5
weight_decay: 0.02
warmup_ratio: 0.15
per_device_train_batch_size: 32
per_device_eval_batch_size: 64
max_grad_norm: 1.0

Performance

Metric	Score
Eval Accuracy	0.95
Eval F1 Macro	0.95
Eval Loss	0.16

Model Architecture

1) ELECTRA Encoder (Base-size)

Hidden size: 768
Layers: 12 Transformer blocks
Attention heads: 12
MLP intermediate size: 3072
Activation: GELU
Dropout: 0.1

2) Classification Head

감정 8개 클래스를 예측하기 위한 추가 분류 헤드:

Dense Layer: 768 → 768
Activation: GELU
Dropout: 0.1
Output Projection: 768 → 8

Citation

bibtex

@misc{HowRUEmotion2025,
  title={HowRU KoELECTRA Emotion Classifier},
  author={Lim, Yeri},
  year={2025},
  publisher={Hugging Face},
  howpublished={\url{https://huggingface.co/LimYeri/HowRU-KoELECTRA-Emotion-Classifier}}
}