人工智能 | Hugging Face 的应用

大纲

Hugging-Face 介绍

• Hugging-Face 大语言模型 LLM 管理
• Transformers 机器学习框架
• 文本生成推理 (TGI)

Hugging Face

Hugging-Face -- 大语言模型界的 Github

Hugging Face 专门开发用于构建机器学习应用的工具。该公司的代表产品是其为自然语言处理应用构建的 transformers 库，以及允许用户共享机器学习模型和数据集的平台

大模型平台 hugging face

国内对标 -- 百度千帆

百度智能云千帆大模型平台（以下简称千帆或千帆大模型平台）是面向企业开发者的一站式大模型开发及服务运行平台。千帆不仅提供了包括文心一言底层模型和第三方开源大模型，还提供了各种 AI 开发工具和整套开发环境，方便客户轻松使用和开发大模型应用。支持数据管理、自动化模型 SFT 以及推理服务云端部署的一站式大模型定制服务，助力各行业的生成式 AI 应用需求落地。

百度千帆

国内对标 -- 魔搭社区

ModelScope 社区成立于 2022 年 6 月，是一个模型开源社区及创新平台，由阿里巴巴通义实验室（Institute for Intelligent Computing），联合 CCF 开源发展委员会，共同作为项目发起方。社区联合国内 AI 领域合作伙伴与高校机构， 致力于通过开放的社区合作，构建深度学习相关的模型开源社区，并开放相关模型创新技术，推动基于“模型即服务”（Model-as-a-Service）理念的模型应用生态的繁荣发展。

Hugging Face 是行业风向标

• 开发者仓库 Hub
• 机器学习模型 Hub
• 数据集 Hub
• 机器学习算法库与封装工具
• 解决方案
• 文档

大语言模型管理

通过 Git 下载大模型

git lfs install
#模型会下载到本地目录，并不在hugging face的标准管理目录里，可能会导致后续Transformers库重新下载
git clone https://huggingface.co/google/owlv2-base-patch16-ensemble

客户端工具 huggingface_hub

python -m pip install huggingface_hub
# 需要开启代理
huggingface-cli login

高速下载

python -m pip install huggingface_hub[hf_transfer]

#缺点：目前还不支持代理，导致下载大文件可能有时候更慢。
export HF_HUB_ENABLE_HF_TRANSFER=1

huggingface-cli download gpt2 config.json

huggingface-cli 用法

(base) hogwarts: llama seveniruby$ huggingface-cli -h
usage: huggingface-cli <command> [<args>]

positional arguments:
  {env,login,whoami,logout,repo,upload,download,lfs-enable-largefiles,lfs-multipart-upload,scan-cache,delete-cache}
                        huggingface-cli command helpers
    env                 Print information about the environment.
    login               Log in using a token from huggingface.co/settings/tokens
    whoami              Find out which huggingface.co account you are logged in as.
    logout              Log out
    repo                {create, ls-files} Commands to interact with your huggingface.co repos.
    upload              Upload a file or a folder to a repo on the Hub
    download            Download files from the Hub
    lfs-enable-largefiles
                        Configure your repository to enable upload of files > 5GB.
    lfs-multipart-upload
                        Command will get called by git-lfs, do not call it directly.
    scan-cache          Scan cache directory.
    delete-cache        Delete revisions from the cache directory.

options:
  -h, --help            show this help message and exit

使用示例

huggingface-cli login
huggingface-cli download gpt2
huggingface-cli download gpt2 config.json
huggingface-cli download gpt2 --revision main

使用 Hub client library 管理

from huggingface_hub import hf_hub_download
hf_hub_download(repo_id="google/pegasus-xsum", filename="config.json")

from huggingface_hub import hf_hub_download
hf_hub_download(
    repo_id="google/pegasus-xsum",
    filename="config.json",
    revision="4d33b01d79672f27f001f6abade33f22d993b151"
)

hugging face 的本地缓存文件结构

#文件快照
./snapshots
./snapshots/187d69aae616814bc1386bdfe38e256518d17637
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/added_tokens.json
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/tokenizer_config.json
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/special_tokens_map.json
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/config.json
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/README.md
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/merges.txt
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/.gitattributes
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/pytorch_model.bin
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/vocab.json
./snapshots/187d69aae616814bc1386bdfe38e256518d17637/preprocessor_config.json
#真正的文件缓存
./blobs
./blobs/76e821f1b6f0a9709293c3b6b51ed90980b3166b
./blobs/69feda8b53b1c9e2a85ae756bf58c120c3c1b4b4a4d97d4876578c1809a63d76
./blobs/39808ac074d69141c0b2f065d16f56dfa5194cac
./blobs/1c4c373d81adc0c63bb3368a89f4a92574395c2c
./blobs/b24bfd579e8682df31c7794a3e1c3ef6a9a22a7a
./blobs/5ff4ac542d5e800dd3aadbdcda354868b5f24ddd
./blobs/a6344aac8c09253b3b630fb776ae94478aa0275b
./blobs/dfc1049705d082052a81e7d18bde6d11fc137a0d
./blobs/ee790e862e5a47398da080893272d3c2b76cf0e5
./blobs/469be27c5c010538f845f518c4f5e8574c78f7c8

#分支，指定分支有助于解决重复下载问题
./refs
./refs/main

hugging face transformers 机器学习框架

Transformers 库

Transformers 是一个预训练的最先进模型库，用于自然语言处理 (NLP)、计算机视觉以及音频和语音处理任务。该库不仅包含 Transformer 模型，还包含非 Transformer 模型，例如用于计算机视觉任务的现代卷积网络。Transformers 提供 API 和工具，可轻松下载和训练最先进的预训练模型，它支持 PyTorch、TensorFlow 和 JAX 之间的框架互操作性。

Transformers 能力

• 自然语言处理：文本分类、命名实体识别、问答、语言建模、摘要、翻译、多项选择和文本生成。
• 计算机视觉：图像分类、对象检测和分割。
• 音频：自动语音识别和音频分类。
• 多模态：表格问答、光学字符识别、扫描文档信息提取、视频分类和视觉问答。

Task	Description	Modality	Pipeline identifier
Text classification	assign a label to a given sequence of text	NLP	pipeline(task=“sentiment-analysis”)
Text generation	generate text given a prompt	NLP	pipeline(task=“text-generation”)
Summarization	generate a summary of a sequence of text or document	NLP	pipeline(task=“summarization”)

Task	Description	Modality	Pipeline identifier
Image classification	assign a label to an image	Computer vision	pipeline(task=“image-classification”)
Image segmentation	assign a label to each individual pixel of an image (supports semantic, panoptic, and instance segmentation)	Computer vision	pipeline(task=“image-segmentation”)
Object detection	predict the bounding boxes and classes of objects in an image	Computer vision	pipeline(task=“object-detection”)

Task	Description	Modality	Pipeline identifier
Audio classification	assign a label to some audio data	Audio	pipeline(task=“audio-classification”)
Automatic speech recognition	transcribe speech into text	Audio	pipeline(task=“automatic-speech-recognition”)

Task	Description	Modality	Pipeline identifier
Visual question answering	answer a question about the image, given an image and a question	Multimodal	pipeline(task=“vqa”)
Document question answering	answer a question about the document, given a document and a question	Multimodal	pipeline(task=“document-question-answering”)
Image captioning	generate a caption for a given image	Multimodal	pipeline(task=“image-to-text”)

快速开始

pip install transformers datasets
pip install torch
pip install tensorflow

#测试
python -c "from transformers import pipeline; print(pipeline('sentiment-analysis')('we love you'))"

pipeline api 使用
pipe = pipeline("text-classification")
pipe("This restaurant is awesome")

#输出
[{'label': 'POSITIVE', 'score': 0.9998743534088135}]

使用第三方库

pipe = pipeline(model="FacebookAI/roberta-large-mnli")
pipe("This restaurant is awesome")

#输出
[{'label': 'NEUTRAL', 'score': 0.7313136458396912}]

图像识别

def test_zero_shot_object_detector():
    with open('coco_sample.png', mode='rb') as f:
        image = Image.open(f)
        object_detector = pipeline('object-detection', revision='main')

        predictions = object_detector(image)
        print(json.dumps(predictions, indent=2, ensure_ascii=False))

#输出
[
  {
    "score": 0.9982201457023621,
    "label": "remote",
    "box": {
      "xmin": 40,
      "ymin": 70,
      "xmax": 175,
      "ymax": 117
    }
  },
  {
    "score": 0.9960021376609802,
    "label": "remote",
    "box": {
      "xmin": 333,
      "ymin": 72,
      "xmax": 368,
      "ymax": 187
    }
  },
  {
    "score": 0.9954745173454285,
    "label": "couch",
    "box": {
      "xmin": 0,
      "ymin": 1,
      "xmax": 639,
      "ymax": 473
    }
  },
  {
    "score": 0.9988006353378296,
    "label": "cat",
    "box": {
      "xmin": 13,
      "ymin": 52,
      "xmax": 314,
      "ymax": 470
    }
  },
  {
    "score": 0.9986783862113953,
    "label": "cat",
    "box": {
      "xmin": 345,
      "ymin": 23,
      "xmax": 640,
      "ymax": 368
    }
  }
]

文本生成推理 Text Generation Inference

Text Generation Inference

文本生成推理 (TGI) 是一个用于部署和服务大型语言模型 (LLM) 的工具包。TGI 为最流行的开源 LLM 提供高性能文本生成，包括 Llama、Falcon、StarCoder、BLOOM、GPT-NeoX 和 T5。

model=tiiuae/falcon-7b-instruct
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:1.4 --model-id $model

model=teknium/OpenHermes-2.5-Mistral-7B
volume=$PWD/data # share a volume with the Docker container to avoid downloading weights every run

docker run --gpus all --shm-size 1g -p 8080:80 -v $volume:/data ghcr.io/huggingface/text-generation-inference:1.4 --model-id $model

TGI 接口

curl 127.0.0.1:8080/generate \
    -X POST \
    -d '{"inputs":"What is Deep Learning?","parameters":{"max_new_tokens":20}}' \
    -H 'Content-Type: application/json'

编程使用 TGI

import requests

headers = {
    "Content-Type": "application/json",
}

data = {
    'inputs': 'What is Deep Learning?',
    'parameters': {
        'max_new_tokens': 20,
    },
}

response = requests.post('http://127.0.0.1:8080/generate', headers=headers, json=data)
print(response.json())
# {'generated_text': '\n\nDeep Learning is a subset of Machine Learning that is concern