演習：Transformerモデルを活用しよう

from transformers import pipeline

# 感情分析パイプライン
classifier = pipeline("sentiment-analysis")

# 英語レビュー
reviews_en = [
    "This product is amazing! Best purchase ever.",
    "Terrible quality. Broke after one week.",
    "Delivery was slow but the product is decent.",
    "Excellent customer service and fast shipping.",
    "Not worth the price. Very disappointed.",
]

print("=== 英語レビューの感情分析 ===")
for review in reviews_en:
    result = classifier(review)[0]
    print(f"  [{result['label']:>8s}] ({result['score']:.3f}) {review[:50]}")

# 日本語レビュー（多言語モデル）
classifier_multi = pipeline(
    "sentiment-analysis",
    model="nlptown/bert-base-multilingual-uncased-sentiment"
)

reviews_ja = [
    "この商品は品質が素晴らしいです。大満足！",
    "配送が遅すぎます。二度と利用しません。",
    "普通の商品です。可もなく不可もなく。",
    "コスパ最高！友達にも勧めたい。",
    "説明と違う商品が届きました。返品します。",
]

print("\n=== 日本語レビューの感情分析 ===")
for review in reviews_ja:
    result = classifier_multi(review)[0]
    print(f"  [{result['label']:>10s}] ({result['score']:.3f}) {review}")

from transformers import (
    DistilBertForSequenceClassification,
    DistilBertTokenizerFast,
    TrainingArguments,
    Trainer
)
from datasets import load_dataset
import numpy as np
from sklearn.metrics import accuracy_score, f1_score

# === データ準備 ===
dataset = load_dataset("imdb")

# サブセットを作成（学習効率のため）
train_dataset = dataset["train"].shuffle(seed=42).select(range(1000))
test_dataset = dataset["test"].shuffle(seed=42).select(range(500))

# === トークナイズ ===
tokenizer = DistilBertTokenizerFast.from_pretrained("distilbert-base-uncased")

def tokenize_function(examples):
    return tokenizer(
        examples["text"],
        padding="max_length",
        truncation=True,
        max_length=256
    )

train_tokenized = train_dataset.map(tokenize_function, batched=True)
test_tokenized = test_dataset.map(tokenize_function, batched=True)

# === モデル準備 ===
model = DistilBertForSequenceClassification.from_pretrained(
    "distilbert-base-uncased",
    num_labels=2
)

# === 評価関数 ===
def compute_metrics(eval_pred):
    logits, labels = eval_pred
    predictions = np.argmax(logits, axis=-1)
    return {
        "accuracy": accuracy_score(labels, predictions),
        "f1": f1_score(labels, predictions, average="binary"),
    }

# === 学習設定 ===
training_args = TrainingArguments(
    output_dir="./results",
    num_train_epochs=3,
    per_device_train_batch_size=16,
    per_device_eval_batch_size=32,
    learning_rate=2e-5,
    weight_decay=0.01,
    evaluation_strategy="epoch",
    save_strategy="epoch",
    load_best_model_at_end=True,
    logging_steps=50,
)

# === Trainer ===
trainer = Trainer(
    model=model,
    args=training_args,
    train_dataset=train_tokenized,
    eval_dataset=test_tokenized,
    compute_metrics=compute_metrics,
)

# === 学習 ===
trainer.train()

# === 評価 ===
results = trainer.evaluate()
print(f"\n=== 最終評価 ===")
print(f"  Accuracy: {results['eval_accuracy']:.4f}")
print(f"  F1 Score: {results['eval_f1']:.4f}")

import torch
import matplotlib.pyplot as plt
import seaborn as sns
import numpy as np

def visualize_attention(text, model, tokenizer):
    """Attention 重みを可視化"""
    inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=64)

    model.eval()
    with torch.no_grad():
        outputs = model(**inputs, output_attentions=True)

    # 予測結果
    logits = outputs.logits
    predicted = torch.argmax(logits, dim=-1).item()
    probs = torch.softmax(logits, dim=-1)
    label = "Positive" if predicted == 1 else "Negative"
    confidence = probs[0][predicted].item()

    # Attention 重み（最終層の平均）
    attention = outputs.attentions[-1]  # 最終層
    attention_avg = attention.mean(dim=1).squeeze()  # ヘッドの平均

    # トークンに変換
    tokens = tokenizer.convert_ids_to_tokens(inputs["input_ids"][0])

    # [CLS] トークンの注目度を表示
    cls_attention = attention_avg[0].numpy()

    # 可視化
    fig, ax = plt.subplots(figsize=(12, 3))
    display_tokens = tokens[:20]  # 最初の20トークンのみ表示
    display_attn = cls_attention[:20]

    bars = ax.bar(range(len(display_tokens)), display_attn)
    ax.set_xticks(range(len(display_tokens)))
    ax.set_xticklabels(display_tokens, rotation=45, ha='right')
    ax.set_title(f"Prediction: {label} ({confidence:.2%})")
    ax.set_ylabel("Attention Weight")

    # 上位3トークンを赤色にハイライト
    top_indices = np.argsort(display_attn)[-3:]
    for idx in top_indices:
        bars[idx].set_color('red')

    plt.tight_layout()
    plt.show()

# テスト
reviews = [
    "This product is absolutely fantastic and high quality!",
    "Terrible experience. The product broke immediately.",
    "Shipping was slow but the item itself is pretty good.",
]

for review in reviews:
    visualize_attention(review, model, tokenizer)
    print()