#BERT家族详解：从双向编码器到主流变体及Hugging Face实战

def bert_vs_gpt_role():
    """
    BERT与GPT的应用定位对比
    """
    print("🔍 BERT（理解型选手）:")
    print("- Encoder-only架构")
    print("- 训练目标：MLM（掩码猜词）+ NSP（句子衔接）")
    print("- 代表token：[CLS]（分类锚点）、[SEP]（句子分隔）")
    print("- 擅长：分类、问答、NER、文本相似度")
    
    print("\n✍️ GPT（生成型选手）:")
    print("- Decoder-only架构")
    print("- 训练目标：标准LM（自回归预测下一词）")
    print("- 擅长：文本续写、对话、摘要")

bert_vs_gpt_role()

import random

def create_simple_mlm(text, mask_ratio=0.15, mask_token="[MASK]"):
    tokens = list(text)  # 简化中文单字分词
    masked_tokens = tokens.copy()
    num_mask = max(1, int(len(tokens) * mask_ratio))
    mask_indices = random.sample(range(len(tokens)), num_mask)
    
    for idx in mask_indices:
        rand = random.random()
        if rand < 0.8:
            masked_tokens[idx] = mask_token
        elif rand >= 0.9:  # 10%随机换，增强鲁棒性
            masked_tokens[idx] = random.choice("中文自然语言处理很有趣")
    
    return "".join(masked_tokens), mask_indices

# 测试
original = "自然语言处理是AI的核心分支"
masked, pos = create_simple_mlm(original)
print(f"原文：{original}")
print(f"掩码：{masked}")
print(f"训练目标：还原被遮的位置")

from transformers import BertTokenizer, BertForSequenceClassification
import torch

# 加载预训练中文模型
tokenizer = BertTokenizer.from_pretrained("bert-base-chinese")
model = BertForSequenceClassification.from_pretrained("bert-base-chinese", num_labels=2)

# 示例输入
texts = ["这个手机续航超棒！", "买了就后悔，屏幕太暗"]
inputs = tokenizer(texts, padding=True, truncation=True, max_length=128, return_tensors="pt")
inputs["labels"] = torch.tensor([1, 0])  # 1正面，0负面

# 前向传播
with torch.no_grad():
    outputs = model(**inputs)
    predictions = torch.argmax(outputs.logits, dim=1)

print(f"预测情感：{predictions.tolist()}")

from transformers import pipeline

# 1. 中文情感分析
sentiment_clf = pipeline(
    "sentiment-analysis",
    model="uer/roberta-base-finetuned-chinanews-chinese"
)
print(sentiment_clf("今天的火锅太好吃了！"))

# 2. 中文命名实体识别
ner = pipeline(
    "ner",
    model="ckiplab/bert-base-chinese-ner",
    aggregation_strategy="simple"  # 合并连续实体
)
print(ner("张三在腾讯深圳总部工作"))

from transformers import AutoTokenizer, AutoModel
from sklearn.metrics.pairwise import cosine_similarity
import torch

tokenizer = AutoTokenizer.from_pretrained("bert-base-chinese")
model = AutoModel.from_pretrained("bert-base-chinese")

def get_embedding(text):
    inputs = tokenizer(text, return_tensors="pt", padding=True, truncation=True)
    with torch.no_grad():
        outputs = model(**inputs)
    # 用<[BOS_never_used_51bce0c785ca2f68081bfa7d91973934]>的嵌入作为句子表示
    return outputs.last_hidden_state[:, 0, :].numpy()

# 测试
texts = ["AI如何改变医疗？", "人工智能对医疗行业的影响"]
emb1, emb2 = get_embedding(texts[0]), get_embedding(texts[1])
print(f"余弦相似度：{cosine_similarity(emb1, emb2)[0][0]:.4f}")