inference_examples.py

"""
최소한의 BERT 추론 예제 모음
============================
논문의 각 태스크를 가장 작은 코드로 보여주는 함수들입니다. 환경 점검이나
복사-붙여넣기용으로 유용합니다.

모든 예제 실행:
    python inference_examples.py
"""

from __future__ import annotations

import torch
from transformers import (
    AutoTokenizer,
    AutoModelForMaskedLM,
    AutoModelForNextSentencePrediction,
    pipeline,
)


# ---------------------------------------------------------------------------
# 1. Masked Language Model (논문 §3.1, Task #1)
# ---------------------------------------------------------------------------
def example_mlm() -> None:
    print("=" * 60)
    print("1. Masked Language Model (마스크 토큰 예측)")
    print("=" * 60)

    tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
    model = AutoModelForMaskedLM.from_pretrained("bert-base-uncased")
    model.eval()

    text = "The capital of France is [MASK]."
    inputs = tokenizer(text, return_tensors="pt")
    mask_idx = (inputs.input_ids == tokenizer.mask_token_id).nonzero(as_tuple=True)[1]

    with torch.no_grad():
        logits = model(**inputs).logits

    # 마스크 위치에 대한 상위 5개 예측
    top_5 = torch.topk(logits[0, mask_idx], 5, dim=-1)
    print(f"입력: {text}")
    for token_id, score in zip(top_5.indices[0], top_5.values[0]):
        token = tokenizer.decode([token_id])
        print(f"  {token!r:<15} 로짓={score.item():.3f}")


# ---------------------------------------------------------------------------
# 2. Next Sentence Prediction (논문 §3.1, Task #2)
# ---------------------------------------------------------------------------
def example_nsp() -> None:
    print("\n" + "=" * 60)
    print("2. Next Sentence Prediction (다음 문장 예측)")
    print("=" * 60)

    tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
    model = AutoModelForNextSentencePrediction.from_pretrained("bert-base-uncased")
    model.eval()

    pairs = [
        ("The man went to the store.", "He bought a gallon of milk."),
        ("The man went to the store.", "Penguins are flightless birds."),
    ]
    for sent_a, sent_b in pairs:
        inputs = tokenizer(sent_a, sent_b, return_tensors="pt")
        with torch.no_grad():
            logits = model(**inputs).logits
        probs = torch.softmax(logits, dim=-1)[0].tolist()
        # HuggingFace BERT NSP: 라벨 0 = IsNext, 라벨 1 = NotNext
        verdict = "IsNext" if probs[0] > probs[1] else "NotNext"
        print(f"  A: {sent_a}")
        print(f"  B: {sent_b}")
        print(f"  -> {verdict}  (P(IsNext)={probs[0]:.3f}, P(NotNext)={probs[1]:.3f})\n")


# ---------------------------------------------------------------------------
# 3. 문장 쌍 분류 (MNLI; 논문 §4.1, Figure 4a)
# ---------------------------------------------------------------------------
def example_mnli() -> None:
    print("=" * 60)
    print("3. MNLI (문장 쌍 분류)")
    print("=" * 60)

    nlp = pipeline(
        "text-classification",
        model="textattack/bert-base-uncased-MNLI",
    )
    premise = "A soccer game with multiple males playing."
    hypothesis = "Some men are playing a sport."
    result = nlp(f"{premise} [SEP] {hypothesis}")[0]
    print(f"  전제:   {premise}")
    print(f"  가설:   {hypothesis}")
    print(f"  -> {result['label']}  ({result['score']:.3f})")


# ---------------------------------------------------------------------------
# 4. 단일 문장 분류 (SST-2; 논문 §4.1, Figure 4b)
# ---------------------------------------------------------------------------
def example_sst2() -> None:
    print("\n" + "=" * 60)
    print("4. SST-2 (단일 문장 감성 분류)")
    print("=" * 60)

    nlp = pipeline(
        "text-classification",
        model="textattack/bert-base-uncased-SST-2",
    )
    sentences = [
        "This movie was absolutely fantastic.",
        "What a complete waste of time.",
    ]
    for s in sentences:
        r = nlp(s)[0]
        print(f"  {s}\n    -> {r['label']}  ({r['score']:.3f})")


# ---------------------------------------------------------------------------
# 5. 질의응답 (SQuAD; 논문 §4.2, Figure 4c)
# ---------------------------------------------------------------------------
def example_squad() -> None:
    print("\n" + "=" * 60)
    print("5. SQuAD v1.1 (추출형 질의응답)")
    print("=" * 60)

    nlp = pipeline(
        "question-answering",
        model="bert-large-uncased-whole-word-masking-finetuned-squad",
    )
    context = (
        "BERT was introduced by Google AI Language in October 2018. "
        "It is pre-trained on the BooksCorpus (800M words) and English Wikipedia "
        "(2,500M words). BERT-Large has 340 million parameters."
    )
    question = "How many parameters does BERT-Large have?"
    result = nlp(question=question, context=context)
    print(f"  질문: {question}")
    print(f"  답변: {result['answer']!r}  (점수={result['score']:.3f})")


# ---------------------------------------------------------------------------
# 6. 개체명 인식 (CoNLL-2003; 논문 §5.3, Figure 4d)
# ---------------------------------------------------------------------------
def example_ner() -> None:
    print("\n" + "=" * 60)
    print("6. CoNLL-2003 NER (개체명 인식)")
    print("=" * 60)

    nlp = pipeline(
        "token-classification",
        model="dslim/bert-base-NER",
        aggregation_strategy="simple",
    )
    text = "Jacob Devlin works at Google in Mountain View, California."
    for ent in nlp(text):
        print(f"  {ent['word']:<20} {ent['entity_group']:<6} (점수={ent['score']:.3f})")


if __name__ == "__main__":
    example_mlm()
    example_nsp()
    example_mnli()
    example_sst2()
    example_squad()
    example_ner()