实验1：Hugging Face Pipeline 一站跑通经典任务

import torch
from transformers import pipeline
import time

# ===== 1. 环境验证 =====
print(f"PyTorch: {torch.__version__}")
print(f"CUDA available: {torch.cuda.is_available()}")
if torch.cuda.is_available():
    print(f"GPU: {torch.cuda.get_device_name(0)}")
    device = 0  # pipeline 的 device 参数：0 = 第一块 GPU
else:
    device = -1  # -1 表示 CPU

# ===== 2. 最简情感分类 =====
sentiment = pipeline("sentiment-analysis", device=device)
result = sentiment("I've been waiting for this class my whole life.")
print(result)
# [{'label': 'POSITIVE', 'score': 0.99...}]

# ===== 英文情感（默认模型）=====
clf_en = pipeline("sentiment-analysis", device=device)

# ===== 中文情感（指定中文模型）=====
clf_zh = pipeline(
    "sentiment-analysis",
    model="uer/roberta-base-finetuned-dianping-chinese",
    device=device,
)

samples = {
    "英文-正面": "This movie is absolutely fantastic, I loved every minute.",
    "英文-负面": "Terrible plot, horrible acting, complete waste of time.",
    "中文-正面": "这家餐厅的菜品精致，服务也非常贴心，值得推荐。",
    "中文-负面": "菜上得慢，态度又差，再也不会来了。",
}

print("\n=== 情感分类 ===")
for name, text in samples.items():
    clf = clf_zh if name.startswith("中文") else clf_en
    t0 = time.time()
    out = clf(text)
    dt = (time.time() - t0) * 1000
    print(f"  [{name}] {out[0]['label']} ({out[0]['score']:.3f}) | {dt:.0f} ms")
    print(f"           原文：{text[:40]}...")

# 英文 NER（默认：dslim/bert-base-NER）
ner_en = pipeline("ner", aggregation_strategy="simple", device=device)

# 中文 NER（用 ckiplab 或 xlm-roberta 的多语言版）
ner_zh = pipeline(
    "ner",
    model="Davlan/bert-base-multilingual-cased-ner-hrl",
    aggregation_strategy="simple",
    device=device,
)

en_text = "Apple CEO Tim Cook visited Beijing on March 3, 2026, meeting Li Qiang at Zhongnanhai."
zh_text = "2026 年 3 月，苹果公司 CEO 库克访问北京大学软件与微电子学院。"

print("\n=== NER ===")
for name, txt, m in [("英文", en_text, ner_en), ("中文", zh_text, ner_zh)]:
    t0 = time.time()
    out = m(txt)
    dt = (time.time() - t0) * 1000
    print(f"  [{name}] {dt:.0f} ms")
    for ent in out:
        print(f"    - '{ent['word']}' → {ent['entity_group']} ({ent['score']:.2f})")

# 英 → 中
translator_en_zh = pipeline(
    "translation",
    model="Helsinki-NLP/opus-mt-en-zh",
    device=device,
)

# 中 → 英
translator_zh_en = pipeline(
    "translation",
    model="Helsinki-NLP/opus-mt-zh-en",
    device=device,
)

en = "Large language models have fundamentally changed how we approach natural language processing."
zh = "大语言模型正在改变自然语言处理的研究范式，也在改变工业界的部署方式。"

print("\n=== 机器翻译 ===")
t0 = time.time()
print(f"  [EN→ZH] {translator_en_zh(en)[0]['translation_text']}")
print(f"          耗时 {(time.time()-t0)*1000:.0f} ms")

t0 = time.time()
print(f"  [ZH→EN] {translator_zh_en(zh)[0]['translation_text']}")
print(f"          耗时 {(time.time()-t0)*1000:.0f} ms")

summarizer = pipeline(
    "summarization",
    model="facebook/bart-large-cnn",  # 英文摘要的经典模型
    device=device,
)

long_text = """
The field of natural language processing has undergone a remarkable transformation
in the past decade. Once dominated by rule-based systems and statistical methods
relying on hand-crafted features like TF-IDF and n-grams, NLP has been reshaped
by the rise of neural networks and, more recently, by the Transformer architecture.
Today, large language models trained on trillions of tokens can perform tasks that
were unthinkable a decade ago, from translating between hundreds of language pairs
to writing functional code. Yet classical techniques have not vanished; they continue
to serve as strong baselines, retrieval backbones in RAG systems, and lightweight
alternatives when compute is constrained.
"""

print("\n=== 摘要 ===")
t0 = time.time()
out = summarizer(long_text, max_length=60, min_length=20, do_sample=False)
print(f"  原文：{len(long_text)} 字符 | 摘要：{out[0]['summary_text']}")
print(f"  耗时 {(time.time()-t0)*1000:.0f} ms")

qa = pipeline(
    "question-answering",
    model="deepset/roberta-base-squad2",  # SQuAD 2.0 训练
    device=device,
)

context = """
The Peking University School of Software and Microelectronics was founded in 2002.
It is located in the Daxing campus of Peking University. The school offers master's
and doctoral programs in software engineering, integrated circuits, and artificial
intelligence. Professor Zhijun Gao teaches the "Artificial Intelligence Practice
(Language Intelligence)" course to graduate students.
"""

questions = [
    "When was the School of Software and Microelectronics founded?",
    "Where is it located?",
    "Who teaches the Language Intelligence course?",
    "Does the school offer undergraduate programs?",  # 回答应当为 "empty" 或低分
]

print("\n=== 抽取式 QA ===")
for q in questions:
    t0 = time.time()
    out = qa(question=q, context=context)
    dt = (time.time() - t0) * 1000
    print(f"  Q: {q}")
    print(f"    → '{out['answer']}' (score={out['score']:.2f}, {dt:.0f} ms)")

# ===== 显存占用检查 =====
if torch.cuda.is_available():
    mem_mb = torch.cuda.memory_allocated() / 1024**2
    max_mem_mb = torch.cuda.max_memory_allocated() / 1024**2
    print(f"\n当前显存占用：{mem_mb:.0f} MB，峰值：{max_mem_mb:.0f} MB")

# ===== 与"传统 pipeline"的概念对比 =====
comparison = """
维度对比：传统 NLP pipeline vs Hugging Face pipeline（Transformer）

| 任务     | 传统方法                         | Transformer pipeline       |
|----------|----------------------------------|----------------------------|
| 情感分类 | 分词 + TF-IDF + LogReg           | BERT + 分类头              |
| NER      | 特征工程 + BiLSTM + CRF          | BERT + token 分类头        |
| 翻译     | 短语对齐 / Seq2Seq + Attention    | Transformer Enc-Dec        |
| 摘要     | 抽取式（TextRank）                | BART / T5 生成             |
| QA       | 模板 + 规则 / BiDAF               | RoBERTa + span head        |

关键观察：
1. 所有任务都通过 pipeline("任务名") 统一调用 —— 接口完全一致
2. 模型、tokenizer、后处理全部被封装 —— 你不需要知道 BERT 和 BART 的架构差异
3. 推理速度：GPU 下每条样本 10-100 ms 级
4. 代价：下载的模型权重总计 ~2 GB，远比传统 pipeline 重
"""
print(comparison)