Transformers学习笔记3. HuggingFace管道函数Pipeline_transformers pipeline 批量推理

一、简介

Hugging face提供了管道函数——Pipeline，可以使用极少的代码快速开启一个NLP任务。

Pipeline 具备了数据预处理、模型处理、模型输出后处理等步骤，可以直接输入原始数据，然后给出预测结果，十分方便。

给定一个任务之后，pipeline会自动调用一个预训练好的模型，然后根据你给的输入执行下面三个步骤：

1. 预处理输入文本，让它可被模型读取
2. 模型处理
3. 模型输出的后处理，让预测结果可读

虽然Pipeline使用很简单，但对于专业人士缺乏灵活性。

当前在下面网址查到当前有效的Pipeline：
https://huggingface.co/docs/transformers/main_classes/pipelines

本文介绍其中一些管道模型的使用。

二、一些管道模型示例

1. 情感分析

from transformers import pipeline

classifier = pipeline("sentiment-analysis")
classifier("I am happy.")
1
2
3
4

输出：

[{'label': 'POSITIVE', 'score': 0.9998760223388672}]
1

也可以传列表作为参数。

2. 零样本文本分类

from transformers import pipeline

classifier = pipeline("zero-shot-classification")
classifier(
    ["This is a course about the Transformers library",
        "New policy mix to propel turnaround in China's economy"],
    candidate_labels=["education", "politics", "business"],
)
1
2
3
4
5
6
7
8

[{'sequence': 'This is a course about the Transformers library', 'labels': ['education', 'business', 'politics'], 'scores': [0.8445969820022583, 0.11197575181722641, 0.0434272475540638]}, 
{'sequence': "New policy mix to propel turnaround in China's economy", 'labels': ['business', 'politics', 'education'], 'scores': [0.6015452146530151, 0.348330557346344, 0.05012420192360878]}]

1
2
3
4

3. 实体名称识别

from transformers import pipeline

ner = pipeline("ner", grouped_entities=True)
print(ner("My name is Sylvain and I work at Hugging Face in Brooklyn."))
1
2
3
4

输出：

[{'entity_group': 'PER', 'score': 0.9981694, 'word': 'Sylvain', 'start': 11, 'end': 18}, 
{'entity_group': 'ORG', 'score': 0.9796019, 'word': 'Hugging Face', 'start': 33, 'end': 45}, 
{'entity_group': 'LOC', 'score': 0.9932106, 'word': 'Brooklyn', 'start': 49, 'end': 57}]

1
2
3
4

4. 摘要

from transformers import pipeline

# use bart in pytorch
summarizer = pipeline("summarization")
summarizer("Sam Shleifer writes the best docstring examples in the whole world.", min_length=5, max_length=8)


1
2
3
4
5
6
7

输出：

# max_length=8
[{'summary_text': ' Sam Shleifer writes'}]
# max_length=12
[{'summary_text': ' Sam Shleifer writes the best docstring'}]
1
2
3
4

5. 文本生成

from transformers import pipeline

generator = pipeline('text-generation', model='liam168/chat-DialoGPT-small-zh')
print(generator('今天早上早点到公司,', max_length=100))
1
2
3
4

6. GPT2英文文本生成

from transformers import pipeline

generator = pipeline("text-generation", model="distilgpt2")
print(generator(
    "In this course, we will teach you how to",
    max_length=30,
    num_return_sequences=2,
))
1
2
3
4
5
6
7
8

结果：

[{'generated_text': 'In this course, we will teach you how to write a powerful and useful resource for your students.\n\n\n\nHow can you understand your own'}, 
{'generated_text': 'In this course, we will teach you how to program a \u202an–n\u202f\u202f\u202f and learn how to use it.'}]
1
2

7. 遮挡字还原

from transformers import pipeline

unmasker = pipeline('fill-mask')

print(unmasker('What the <mask>?', top_k=3))
1
2
3
4
5

结果：

[{'score': 0.378376841545105, 'token': 17835, 'token_str': ' heck', 'sequence': 'What the heck?'}, 
{'score': 0.32931089401245117, 'token': 7105, 'token_str': ' hell', 'sequence': 'What the hell?'}, 
{'score': 0.1464540809392929, 'token': 26536, 'token_str': ' fuck', 'sequence': 'What the fuck?'}]

1
2
3
4

8. 问答

from transformers import pipeline

question_answerer = pipeline("question-answering")
print(question_answerer(
    question="Where do I work?",
    context="My name is Sylvain and I work at Hugging Face in Brooklyn",
))
1
2
3
4
5
6
7

输出：

{'score': 0.6949763894081116, 'start': 33, 'end': 45, 'answer': 'Hugging Face'}

1
2

9. 翻译

一个中文到广东话翻译器：

from transformers import pipeline

translator = pipeline("translation", model="botisan-ai/mt5-translate-zh-yue")
print(translator("今天吃早饭没有？"))
1
2
3
4

输出：

[{'translation_text': '今日食早飯未?'}]
1