from transformers import pipeline
 
classifier = pipeline("sentiment-analysis")
result = classifier("This movie is fantastic!")
# [{'label': 'POSITIVE', 'score': 0.9998}]

classifier = pipeline("sentiment-analysis", model="cardiffnlp/twitter-roberta-base-sentiment-latest")

from transformers import AutoTokenizer, AutoModelForSequenceClassification
import torch
 
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
model = AutoModelForSequenceClassification.from_pretrained("bert-base-uncased", num_labels=2)
 
text = "This is a great product!"
# tokenize：把文本转成 token id
inputs = tokenizer(text, return_tensors="pt", truncation=True, max_length=512)
 
# 模型前向传播
with torch.no_grad():
    outputs = model(**inputs)
 
# outputs.logits 形状：[batch_size, num_labels]
logits = outputs.logits
probabilities = torch.softmax(logits, dim=-1)
predicted_class = torch.argmax(probabilities, dim=-1).item()

inputs = tokenizer("Hello world", return_tensors="pt")
# {
#   'input_ids': tensor([[101, 7592, 2088, 102]]),       # token id
#   'token_type_ids': tensor([[0, 0, 0, 0]]),            # 句子段落 id（BERT 用）
#   'attention_mask': tensor([[1, 1, 1, 1]])             # 1 表示真实 token，0 表示 padding
# }

# config.json 里有 "model_type": "bert"
# AutoTokenizer 自动选择 BertTokenizer
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")
 
# AutoModel 自动选择 BertModel
model = AutoModel.from_pretrained("bert-base-uncased")

from huggingface_hub import list_models
 
# 搜索中文文本分类模型，按下载量排序，取前 5 个
models = list(list_models(
    task="text-classification",
    language="zh",
    sort="downloads",
    limit=5,
))
for m in models:
    print(m.id, m.downloads)

~/.cache/huggingface/hub/

~/.cache/huggingface/hub/
  models--bert-base-uncased/
    snapshots/
      <commit-hash>/       # 模型文件（config.json、pytorch_model.bin 等）
    refs/
      main                 # 指向最新 commit hash 的指针

export HF_HOME=/data/hf_cache

import os
os.environ["TRANSFORMERS_OFFLINE"] = "1"
 
# 此时 from_pretrained 只从本地缓存加载，找不到则报错
tokenizer = AutoTokenizer.from_pretrained("bert-base-uncased")

from transformers import pipeline
 
# 情感分析：判断文本是正面还是负面
# 典型应用：用户评价分析、舆情监控
classifier = pipeline(
    "sentiment-analysis",
    model="distilbert-base-uncased-finetuned-sst-2-english"
)
result = classifier("The service was incredibly disappointing.")
# [{'label': 'NEGATIVE', 'score': 0.9994}]

from transformers import pipeline
 
# NER：识别文本中的人名、地名、组织名等实体
# 典型应用：信息抽取、知识图谱构建
ner = pipeline(
    "ner",
    model="dbmdz/bert-large-cased-finetuned-conll03-english",
    aggregation_strategy="simple"  # 合并同一实体的多个 token
)
result = ner("Hugging Face is a company based in New York City.")
# [
#   {'entity_group': 'ORG', 'word': 'Hugging Face', 'score': 0.99},
#   {'entity_group': 'LOC', 'word': 'New York City', 'score': 0.99}
# ]

from transformers import pipeline
 
# 抽取式问答：从给定段落中找到问题的答案
# 典型应用：文档问答、客服知识库检索
qa = pipeline(
    "question-answering",
    model="deepset/roberta-base-squad2"
)
result = qa(
    question="Where is Hugging Face based?",
    context="Hugging Face is an AI company headquartered in New York City, with offices in Paris."
)
# {'answer': 'New York City', 'score': 0.98, 'start': 45, 'end': 58}

from transformers import pipeline
 
# 文本摘要：将长文本压缩为简短摘要
# 典型应用：新闻摘要、会议纪要、文档压缩
summarizer = pipeline(
    "summarization",
    model="facebook/bart-large-cnn"
)
article = """
The Transformer architecture, introduced in the paper "Attention Is All You Need" (2017),
revolutionized natural language processing. Unlike previous recurrent models, Transformers
use self-attention mechanisms to process all tokens in parallel, making training significantly
faster. This architecture became the foundation for BERT, GPT, and virtually all modern
large language models.
"""
result = summarizer(article, max_length=60, min_length=20, do_sample=False)
# [{'summary_text': 'The Transformer architecture ...'}]

from transformers import pipeline
 
# 文本生成：给定 prompt，续写文本
# 典型应用：内容创作辅助、代码补全、对话
generator = pipeline(
    "text-generation",
    model="gpt2"
)
result = generator(
    "The future of artificial intelligence",
    max_new_tokens=50,
    num_return_sequences=1,
    do_sample=True,   # 采样解码，结果有随机性
    temperature=0.8,  # 控制随机性：越低越保守，越高越多样
)
# [{'generated_text': 'The future of artificial intelligence ...'}]

from transformers import pipeline
 
# 特征提取：获取文本的向量表示
# 典型应用：语义搜索、RAG、文本聚类、相似度计算
feature_extractor = pipeline(
    "feature-extraction",
    model="sentence-transformers/all-MiniLM-L6-v2",
    tokenize_kwargs={"truncation": True}
)
result = feature_extractor("Hello, how are you?")
# result 是一个嵌套列表，形状 [1, seq_len, hidden_size]
# 取平均池化得到句子向量（单句无 padding，直接取均值）
import numpy as np
sentence_embedding = np.mean(result[0], axis=0)  # 形状 [hidden_size]
print(f"句子向量维度：{sentence_embedding.shape}")  # (384,)