20行代码：Serverless架构下用Python轻松搞定图像分类

• 2020 年 3 月 2 日
• 筆記

图像分类是人工智能领域的一个热门话题，同样在生产环境中也会经常会遇到类似的需求，那么怎么快速搭建一个图像分类，或者图像内容是别的API呢？

首先，给大家推荐一个图像相关的库：ImageAI

通过官方给的代码，我们可以看到一个简单的Demo：

from imageai.Prediction import ImagePrediction  import os  execution_path = os.getcwd()    prediction = ImagePrediction()  prediction.setModelTypeAsResNet()  prediction.setModelPath(os.path.join(execution_path, "resnet50_weights_tf_dim_ordering_tf_kernels.h5"))  prediction.loadModel()    predictions, probabilities = prediction.predictImage(os.path.join(execution_path, "1.jpg"), result_count=5 )  for eachPrediction, eachProbability in zip(predictions, probabilities):      print(eachPrediction + " : " + eachProbability)

通过这个Demo我们可以考虑将这个模块部署到云函数：

首先，我们在本地创建一个Python的项目：

mkdir imageDemo

然后新建文件：vim index.py

from imageai.Prediction import ImagePrediction  import os, base64, random    execution_path = os.getcwd()    prediction = ImagePrediction()  prediction.setModelTypeAsSqueezeNet()  prediction.setModelPath(os.path.join(execution_path, "squeezenet_weights_tf_dim_ordering_tf_kernels.h5"))  prediction.loadModel()      def main_handler(event, context):      imgData = base64.b64decode(event["body"])      fileName = '/tmp/' + "".join(random.sample('zyxwvutsrqponmlkjihgfedcba', 5))      with open(fileName, 'wb') as f:          f.write(imgData)      resultData = {}      predictions, probabilities = prediction.predictImage(fileName, result_count=5)      for eachPrediction, eachProbability in zip(predictions, probabilities):          resultData[eachPrediction] =  eachProbability      return resultData

创建完成之后，我们需要下载一下我们所依赖的模型：

- SqueezeNet（文件大小：4.82 MB，预测时间最短，精准度适中）  - ResNet50 by Microsoft Research （文件大小：98 MB，预测时间较快，精准度高）  - InceptionV3 by Google Brain team （文件大小：91.6 MB，预测时间慢，精度更高）  - DenseNet121 by Facebook AI Research （文件大小：31.6 MB，预测时间较慢，精度最高）

我们先用第一个SqueezeNet来做测试：

在官方文档复制模型文件地址：

使用wget直接安装：

wget https://github.com/OlafenwaMoses/ImageAI/releases/download/1.0/squeezenet_weights_tf_dim_ordering_tf_kernels.h5

接下来，我们就需要进行安装依赖了，这里面貌似安装的内容蛮多的：

而且这些依赖有一些需要编译的，这就需要我们在centos + python2.7/3.6的版本下打包才可以，这样就显得非常复杂，尤其是mac/windows用户，伤不起。

所以这时候，直接用我之前的打包网址：

直接下载解压，然后放到自己的项目中：

最后，一步了，我们创建serverless.yaml

imageDemo:    component: "@serverless/tencent-scf"    inputs:      name: imageDemo      codeUri: ./      handler: index.main_handler      runtime: Python3.6      region: ap-guangzhou      description: 图像识别/分类Demo      memorySize: 256      timeout: 10      events:        - apigw:            name: imageDemo_apigw_service            parameters:              protocols:                - http              serviceName: serverless              description: 图像识别/分类DemoAPI              environment: release              endpoints:                - path: /image                  method: ANY

完成之后，执行我们的sls --debug部署，部署过程中会有扫码的登陆，登陆之后等待即可，完成之后，我们可以复制生成的URL：

通过Python语言进行测试，url就是我们刚才复制的+/image：

import urllib.request  import base64    with open("1.jpg", 'rb') as f:      base64_data = base64.b64encode(f.read())      s = base64_data.decode()    url = 'http://service-9p7hbgvg-1256773370.gz.apigw.tencentcs.com/release/image'    print(urllib.request.urlopen(urllib.request.Request(      url = url,      data=s.encode("utf-8")  )).read().decode("utf-8"))

通过网络搜索一张图片，例如我找了这个：

得到运行结果：

{"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}

将代码修改一下，进行一下简单的耗时测试：

import urllib.request  import base64, time    for i in range(0,10):      start_time = time.time()      with open("1.jpg", 'rb') as f:          base64_data = base64.b64encode(f.read())          s = base64_data.decode()        url = 'http://service-hh53d8yz-1256773370.bj.apigw.tencentcs.com/release/test'        print(urllib.request.urlopen(urllib.request.Request(          url = url,          data=s.encode("utf-8")      )).read().decode("utf-8"))      print("cost: ", time.time() - start_time)

输出结果：

{"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  2.1161561012268066  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  1.1259253025054932  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  1.3322770595550537  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  1.3562259674072266  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  1.0180821418762207  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  1.4290671348571777  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  1.5917718410491943  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  1.1727900505065918  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  2.962592840194702  {"cheetah": 83.12643766403198, "Irish_terrier": 2.315458096563816, "lion": 1.8476998433470726, "teddy": 1.6655176877975464, "baboon": 1.5562783926725388}  cost:  1.2248001098632812

这个数据，整体性能基本是在我可以接受的范围内。

至此，我们通过Serveerless架构搭建的Python版本的图像识别/分类小工具做好了。