Keras模型生产环境的部署[源码+教程]

发表: 2018-12-07 浏览: 2563

机器学习 Python

作者：徐静硕士研究生、算法工程师
兴趣方向：统计机器学习，深度学习，模型的线上化部署、网络爬虫，前端可视化。
个人博客：https://dataxujing.github.io/
项目地址：
https://github.com/DataXujing/tensorflow-serving-Wechat

1. 需要的系统环境

（1）一台生产服务器，系统为CentOs6.5（ubuntu只需在此基础上做简单修改即可完成部署）（2）默认为Python2.X（CentOs6.5默认的Python安装版本，特别提醒，不要卸载该版本，因为很多命令行函数基于Python2.X）（3）需要安装Python3.6环境

本文将演示从零开始把Keras训练的MNIST数据集项目部署到生产环境。

需要的模块：

tensorflow
Flask
gevent
gunicorn
keras
numpy
h5py
pillow
uwsgi
supervisor

2.核心代码分享

代码目录如下：

其中Keras模型部分如下(对应文件结构中的train.py):


'''
Gets to 98.78% test accuracy after 12 epochs
https://dataxujing.github.io/用CNN实现MNIST数据集/
'''

#python 2/3 compatibility
from __future__ import print_function
import keras
from keras.datasets import mnist
from keras.models import Sequential
from keras.layers import Dense, Dropout, Flatten
from keras.layers import Conv2D, MaxPooling2D

from keras import backend as K
# 模型结构画出来
from keras.utils.vis_utils import plot_model

batch_size = 128
num_classes = 10
epochs = 12

# 输入图片的大小
# 28x28 pixel images. 
img_rows, img_cols = 28, 28

# 数据划分
(x_train, y_train), (x_test, y_test) = mnist.load_data()

#3D数据, "channels_last" assumes (conv_dim1, conv_dim2, conv_dim3, channels) 
#while "channels_first" assumes (channels, conv_dim1, conv_dim2, conv_dim3).
if K.image_data_format() == 'channels_first':
   x_train = x_train.reshape(x_train.shape[0], 1, img_rows, img_cols)
   x_test = x_test.reshape(x_test.shape[0], 1, img_rows, img_cols)
   input_shape = (1, img_rows, img_cols)
else:
   x_train = x_train.reshape(x_train.shape[0], img_rows, img_cols, 1)
   x_test = x_test.reshape(x_test.shape[0], img_rows, img_cols, 1)
   input_shape = (img_rows, img_cols, 1)

# 数据变换
x_train = x_train.astype('float32')
x_test = x_test.astype('float32')
x_train /= 255
x_test /= 255
print('x_train shape:', x_train.shape)
print(x_train.shape[0], 'train samples')
print(x_test.shape[0], 'test samples')

# one-hot我们的label
y_train = keras.utils.to_categorical(y_train, num_classes)
y_test = keras.utils.to_categorical(y_test, num_classes)

# 构建模型
model = Sequential()
# cov2D + Relu
model.add(Conv2D(32, kernel_size=(3, 3),
                activation='relu',
                input_shape=input_shape))
model.add(Conv2D(64, (3, 3), activation='relu'))
# Pooling
model.add(MaxPooling2D(pool_size=(2, 2)))
# Drop out
model.add(Dropout(0.25))
#flatten 
model.add(Flatten())
#fully connected to get all relevant data
model.add(Dense(128, activation='relu'))
# Drop out 
model.add(Dropout(0.5))
# softmax
model.add(Dense(num_classes, activation='softmax'))
#Adaptive learning rate (adaDelta)  
model.compile(loss=keras.losses.categorical_crossentropy,
             optimizer=keras.optimizers.Adadelta(),
             metrics=['accuracy'])

#train 
model.fit(x_train, y_train,
         batch_size=batch_size,
         epochs=epochs,
         verbose=1,
         validation_data=(x_test, y_test))

# metric
score = model.evaluate(x_test, y_test, verbose=0)
print('Test loss:', score[0])
print('Test accuracy:', score[1])


#Save the model
# serialize model to JSON
model_json = model.to_json()
with open("model.json", "w") as json_file:
   json_file.write(model_json)
# serialize weights to HDF5
model.save_weights("model.h5")
print("Saved model to disk")

# plot
plot_model(model, to_file='model_lx.png',show_shapes=True)

训练后的模型会保存在文件结构中的model文件夹，供Flask调用。

Flask serve部分(对应文件结构中的app.py部分):

from flask import Flask, render_template,request
#scientific computing library for saving, reading, and resizing images
from scipy.misc import imsave, imread, imresize
import numpy as np
import keras.models
import re

import sys 
import os
import base64

#tell our app where our saved model is
sys.path.append(os.path.abspath("./model"))
from load import * 

#initalize our flask app
app = Flask(__name__)

global model, graph
model, graph = init()

#decoding an image from base64 into raw representation
def convertImage(imgData1):
   imgstr = re.search(b'base64,(.*)',imgData1).group(1) # 匹配第一个括号
   with open('output.png','wb') as output:
       output.write(base64.b64decode(imgstr))
   

@app.route('/')
def index():
   return render_template("index.html")

@app.route('/predict/',methods=['GET','POST'])
def predict():
   imgData = request.get_data()
   convertImage(imgData)
   #read the image into memory
   x = imread('output.png',mode='L')
   #compute a bit-wise inversion so black becomes white and vice versa
   x = np.invert(x)
   #make it the right size
   x = imresize(x,(28,28))
   x = x.reshape(1,28,28,1)

   with graph.as_default():
       #perform the prediction
       out = model.predict(x)
       print(out)
       print(np.argmax(out,axis=1))
       #convert the response to a string
       response = np.array_str(np.argmax(out,axis=1))
       return response
   

if __name__ == "__main__":
   port = int(os.environ.get('PORT', 5060))
   app.run(host='0.0.0.0', port=port)

运行app.py可以测试训练的Keras模型是否正常运行。为了部署Flask应用，采用nginx+uwsgi+supervisor的方式把上述Flask应用部署到centOs服务器。

3.安装和配置uwsgi

uwsgi的配置内容如下(对应文件结构中的keras_uwsgi.ini文件):


[uwsgi]
#application's base folder
base = /home/soft/Keras_model

#python module to import
app = app
module = %(app)

home = %(base)/envKeras
#home =/root/anaconda3
pythonpath = %(base)

#socket file's location
socket = 127.0.0.1:8001


#permissions for the socket file
chmod-socket    = 777

#the variable that holds a flask application inside the module imported at line #6
callable = app

#location of log files
logto = /home/soft/Keras_model/log/%n.log


chdir = /home/soft/Keras_model/

# 处理器数
processes = 4
# 线程数
threads = 2

4.安装和配置nginx

首先服务器需要安装nginx,centOs运行一下命令即可安装：

yum install -y nginx

浏览器下测试是否安装成功:

安装成功后将下面配置项(对应文件结构中的keras_nginx.conf)复制到/etc/nginx/conf.d/中：


server {
   listen      5060;
   server_name XXX.XXX.XXX.XXX;
   charset     utf-8;


   location / {
       include uwsgi_params;
       uwsgi_pass unix: 127.0.0.1:8001;
   }
   location ~ /static/ {
       
       root /home/soft/Keras_model;
   }
}

5.安装和配置supervisor

supervisor可以在后台启动Flask应用，并且可以配置当服务器宕机时仍可以自启动Flask应用。其配置项如下（对应文件结构中的keras_supervisor.conf):

[program:app]

# 启动命令入口
command=uwsgi --ini /home/soft/Keras_model/keras_uwsgi.ini

# 命令程序所在目录
#directory=/home/soft/keras_model
#运行命令的用户名


autostart=true
autorestart=true
#日志地址
stdout_logfile=/home/soft/Keras_model/log/uwsgi_supervisor.log
stderr_logfile=/home/soft/Keras_model/log/uwsgi_supervisor_error.log

[inet_http_server]
port = 127.0.0.1:9001

[supervisord]
nodaemon=true
user=root

[supervisorctl]

6. 测试

最后运行最终的run.sh的Shell脚本就可以启动我们的基于Keras的Flask的应用，其中run.sh的脚本内容如下:

#!/bin/bash
echo "Keras+Flask!"
nginx  | uwsgi --ini /home/soft/Keras_model/keras_uwsgi.ini | 
supervisord -c /home/soft/Keras_model/keras_supervisor.conf

我们可以在浏览器中测试我们的MNIST,手写数字识别应用：