471_机器翻译原理¶

# 使用 HuggingFace 中的 Helsinki-NLP 翻译模型（Transformer架构）

from transformers import MarianMTModel, MarianTokenizer

# ---- 1. 加载中→英翻译模型 ----
# Helsinki-NLP 提供200+语言对的预训练翻译模型
model_name = 'Helsinki-NLP/opus-mt-zh-en'  # 中文→英文
tokenizer = MarianTokenizer.from_pretrained(model_name)
model = MarianMTModel.from_pretrained(model_name)

# ---- 2. 批量翻译 ----
texts = [
    "今天天气很好",
    "机器翻译是自然语言处理的重要任务",
    "深度学习改变了机器翻译的格局"
]

# 分词编码
inputs = tokenizer(
    texts,
    return_tensors='pt',
    padding=True,          # 批量需要padding到相同长度
    truncation=True,
    max_length=512
)

# ---- 3. 生成翻译（beam search）----
translated = model.generate(
    **inputs,
    num_beams=5,          # beam宽度5，质量和速度平衡
    max_length=256,
    early_stopping=True
)

# ---- 4. 解码输出 ----
results = tokenizer.batch_decode(translated, skip_special_tokens=True)
for src, tgt in zip(texts, results):
    print(f"原文: {src}")
    print(f"译文: {tgt}")
    print()

# ---- 5. Transformer 注意力可视化（概念演示）----
import torch
import matplotlib.pyplot as plt

def visualize_attention(model, tokenizer, src_text):
    """可视化编码器自注意力权重"""
    inputs = tokenizer(src_text, return_tensors='pt')
    with torch.no_grad():
        # 输出注意力权重
        outputs = model.model.encoder(
            **inputs,
            output_attentions=True  # 返回注意力权重
        )
    # outputs.attentions: tuple of (batch, heads, seq, seq)
    attn = outputs.attentions[-1][0, 0].numpy()  # 最后一层第1个head
    tokens = tokenizer.convert_ids_to_tokens(inputs['input_ids'][0])

    plt.figure(figsize=(8, 6))
    plt.imshow(attn, cmap='Blues')
    plt.xticks(range(len(tokens)), tokens, rotation=45)
    plt.yticks(range(len(tokens)), tokens)
    plt.title('Encoder Self-Attention')
    plt.tight_layout()
    plt.savefig('attention.png')
    return attn

# ---- 6. BLEU评估 ----
# pip install sacrebleu
from sacrebleu.metrics import BLEU

bleu = BLEU()
hypothesis = ["The weather is very good today"]
references = [["Today the weather is very nice"]]
score = bleu.corpus_score(hypothesis, references)
print(f"BLEU: {score.score:.2f}")