首页 函数计算(旧版) 实践教程 基于函数计算和TensorFlow的Serverless AI推理

基于函数计算和TensorFlow的Serverless AI推理

更新时间: 2024-03-15 15:04:24

您可以通过Serverless Devs开发工具快速将AI推理类的应用部署到函数计算控制台,从而实现按需、自动扩展的AI模型推理,无需管理底层基础设施,节省成本并快速迭代。

背景信息

本文示例实现的能力是识别上传的照片内的动物是猫还是狗。关于代码详情,请参见示例工程dog and cat

前提条件

操作步骤

  1. 执行以下命令,克隆项目。

    git clone https://github.com/awesome-fc/cat-dog-classify.git
  2. 安装依赖。

    1. 执行以下命令进入项目目录。

      cd cat-dog-classify
    2. 执行以下命令安装依赖。

      sudo s build --use-docker

      输出示例:

      [2021-12-09 07:26:39] [INFO] [S-CLI] - Start ...
      [2021-12-09 07:26:40] [INFO] [FC-BUILD] - Build artifact start...
      [2021-12-09 07:26:40] [INFO] [FC-BUILD] - Use docker for building.
      [2021-12-09 07:26:40] [INFO] [FC-BUILD] - Build function using image: registry.<regionId>.aliyuncs.com/aliyunfc/runtime-python3.6:build-1.9.20
      [2021-12-09 07:26:40] [INFO] [FC-BUILD] - begin pulling image registry.<regionId>.aliyuncs.com/aliyunfc/runtime-python3.6:build-1.9.20, you can also use docker pull registry.cn-beijing.aliyuncs.com/aliyunfc/runtime-python3.6:build-1.9.20 to pull image by yourself.
      build-1.9.20: Pulling from aliyunfc/runtime-python3.6
      f49cf87b52c1: Already exists
      ......
      01ce50b4eb85: Already exists
      02b807385deb: Pull complete
      ......
      9b9fdb8de506: Pull complete
      Digest: sha256:a9a6dab2d6319df741ee135d9749a90b2bb834fd11ee265d1fb106053890****
      Status: Downloaded newer image for registry.<regionId>.aliyuncs.com/aliyunfc/runtime-python3.6:build-1.9.20
       builder begin to build
      [2021-12-09 07:27:57] [INFO] [FC-BUILD] - Build artifact successfully.
      
      Tips for next step
      ======================
      * Invoke Event Function: s local invoke
      * Invoke Http Function: s local start
      * Deploy Resources: s deploy
      End of method: build

    执行完安装依赖的命令后,Serverless Devs会自动安装相关依赖包,并将第三方库下载到.s/build/artifacts/cat-dog/classify/.s/python目录内。

  3. 上传依赖到NAS。

    当您在安装依赖时,函数计算引用的代码包在解压后可能会出现大于代码包限制的情况,为了减少代码包的体积,您可以将大体积的依赖和相对较大的模型参数文件存放在NAS中。

    1. 执行以下命令,初始化NAS。

      sudo s nas init

      输出示例:

      [2021-12-09 07:29:58] [INFO] [S-CLI] - Start ...
      [2021-12-09 07:29:59] [INFO] [FC-DEPLOY] - Using region: cn-shenzhen
      [2021-12-09 07:29:59] [INFO] [FC-DEPLOY] - Using access alias: default
      [2021-12-09 07:29:59] [INFO] [FC-DEPLOY] - Using accessKeyID: LTAI4G4cwJkK4Rza6xd9****
      [2021-12-09 07:29:59] [INFO] [FC-DEPLOY] - Using accessKeySecret: eCc0GxSpzfq1DVspnqqd6nmYNN****
      ......
      [2021-12-09 07:30:01] [INFO] [FC-DEPLOY] - Generated vpcConfig:
      securityGroupId: sg-wz90u1syk2h1f14b****
      vSwitchId: vsw-wz9qnuult4q5g4f7n****
      vpcId: vpc-wz9x9bzs0wtvjgt6n****
      
      ......
      [2021-12-09 07:30:15] [INFO] [FC-DEPLOY] - Checking Trigger httpTrigger exists
       Make service _FC_NAS_cat-dog success.
       Make function _FC_NAS_cat-dog/nas_dir_checker success.
       Make trigger _FC_NAS_cat-dog/nas_dir_checker/httpTrigger success.
      [2021-12-09 07:30:25] [INFO] [FC-DEPLOY] - Checking Service _FC_NAS_cat-dog exists
      [2021-12-09 07:30:25] [INFO] [FC-DEPLOY] - Checking Function nas_dir_checker exists
      [2021-12-09 07:30:26] [INFO] [FC-DEPLOY] - Checking Trigger httpTrigger exists
      
      There is auto config in the service: _FC_NAS_cat-dog
      [2021-12-09 07:30:26] [INFO] [FC-DEPLOY] - Generated nasConfig:
      groupId: 10003
      mountPoints:
        - fcDir: /mnt/auto
          nasDir: /cat-dog
          serverAddr: 2bfb748****.cn-shenzhen.nas.aliyuncs.com
      userId: 10003
      
      cat-dog:
        userId:      10003
        groupId:     10003
        mountPoints:
          -
            serverAddr: 2bfb748****.cn-shenzhen.nas.aliyuncs.com
            nasDir:     /cat-dog
            fcDir:      /mnt/auto
    2. 执行以下命令,部署服务。

      sudo s deploy service
    3. 执行以下命令,上传依赖到NAS。

      sudo s nas upload -r .s/build/artifacts/cat-dog/classify/.s/python/ /mnt/auto/python

      输出示例:

      [2021-12-09 07:33:14] [INFO] [S-CLI] - Start ...
      Packing ...
      Package complete.
       Upload done
      
      Tips for next step
      ======================
      * Invoke remote function: s invoke
      End of method: nas
    4. 执行以下命令,上传模型到NAS。

      sudo s nas upload -r src/model/ /mnt/auto/model

      输出示例:

      [2021-12-09 07:52:26] [INFO] [S-CLI] - Start ...
      Packing ...
      Package complete.
       Upload done
      
      Tips for next step
      ======================
      * Invoke remote function: s invoke
      End of method: nas
    5. 执行以下命令,查看NAS目录。

      sudo s nas command ls /mnt/auto/

      输出示例:

      [2021-12-09 07:53:01] [INFO] [S-CLI] - Start ...
      model
      python
      
      
      Tips for next step
      ======================
      * Invoke remote function: s invoke
      End of method: nas
  4. 执行以下命令,部署项目。

    sudo s deploy

    输出示例:

    [2021-12-09 07:56:15] [INFO] [S-CLI] - Start ...
    [2021-12-09 07:56:16] [INFO] [FC-DEPLOY] - Using region: cn-shenzhen
    [2021-12-09 07:56:16] [INFO] [FC-DEPLOY] - Using access alias: default
    [2021-12-09 07:56:16] [INFO] [FC-DEPLOY] - Using accessKeyID: LTAI4G4cwJkK4Rza6xd9****
    [2021-12-09 07:56:16] [INFO] [FC-DEPLOY] - Using accessKeySecret: eCc0GxSpzfq1DVspnqqd6nmYNN****
     ......
    [2021-12-09 07:56:19] [INFO] [FC-DEPLOY] - Generated logConfig:
    enableInstanceMetrics: true
    enableRequestMetrics: true
    logBeginRule: ~
    logstore: fc-service-cat-dog-logstore
    project: 188077086902****-cn-shenzhen-logproject
    ......
    There is auto config in the service: cat-dog
    
    Tips for next step
    ======================
    * Display information of the deployed resource: s info
    * Display metrics: s metrics
    * Display logs: s logs
    * Invoke remote function: s invoke
    * Remove Service: s remove service
    * Remove Function: s remove function
    * Remove Trigger: s remove trigger
    * Remove CustomDomain: s remove domain
    
    
    
    cat-dog:
      region:   cn-shenzhen
      service:
        name: cat-dog
      function:
        name:       classify
        runtime:    python3
        handler:    predict.handler
        memorySize: 1024
        timeout:    120
      url:
        system_url:    https://188077086902****.cn-shenzhen.fc.aliyuncs.com/2016-08-15/proxy/cat-dog/classify/
        custom_domain:
          -
            domain: http://classify.cat-dog.188077086902****.cn-shenzhen.fc.devsapp.net
      triggers:
        -
          type: http
          name: httpTrigger

    成功部署该项目后,您可以在执行输出中查看到函数计算生成的临时域名,通过该域名可以访问刚部署的函数。生成的域名的格式为:http://classify.cat-dog.<account_id>.<region_id>.fc.devsapp.net

    使用浏览器访问该域名,上传图片后识别到的结果如下:cat-dogserverless devs

    说明 临时域名仅用作演示以及开发,具有时效性。如需用作生产,请绑定已经在阿里云备案的域名。详细信息,请参见配置自定义域名

使用预留消除冷启动毛刺

函数计算具有动态伸缩的特性,根据并发请求量,自动弹性扩容出执行环境。在这个典型的深度学习示例中,加载依赖和模型参数消耗的时间很长,在您设置的1 GB规格的函数中,并发访问的时间为10s左右,有时可能大于20s。

因此不可避免的会出现函数调用毛刺的情况,即冷启动时间大于10s,在这种情况下,您可以使用设置预留的方式来避免冷启动。您可以在项目目录内执行以下命令消除冷启动毛刺:

sudo s provision put --target 10 --qualifier LATEST

同时,当您需要了解服务器的最大承受能力,实现更好地运行和开发时,您可以使用Serverless Devs的压测命令对指定的函数进行压测。详细信息,请参见Serverless Devs操作命令

重要
  • 您可以通过以下命令获取预留实例详情:

    sudo s provision get --qualifier LATEST
  • 当您完成压测后,请执行以下命令取消预留:

    sudo s provision put --target 0 --qualifier LATEST

相关文档

阿里云首页 函数计算(旧版) 相关技术圈