Building a Keras + deep learning REST API（三部曲之一）-白红宇

一、基本环境

$ pip install flask gevent requests pillow 

其中 flask不需要解释

gevent 是用于自动切换进程的；

pillow 是用来进行python下的图像处理的；

requests 是用来进行python下request处理的。

二、核心代码解释

# import the necessary packages
from keras.applications 
import ResNet50 
from keras.preprocessing.image 
import img_to_array 
from keras.applications 
import imagenet_utils 
from PIL 
import Image 
import numpy 
as np 
import flask 
import io 

引入所需的头文件。其中注意keras的几个类库是很有通用性的；

# initialize our Flask application and the Keras model
        app = flask.Flask( 
__name__) 
        model = 
None

类库的初始化

def
load_model(): 
# load the pre-trained Keras model (here we are using a model
# pre-trained on ImageNet and provided by Keras, but you can
# substitute in your own networks just as easily)
global model 
            model = ResNet50( 
weights= 
"imagenet") 

引入model模型，如果想引入自己的模型(CBIR)的话，就在这里引入。

def
prepare_image( 
image, 
target): 
# if the image mode is not RGB, convert it
if image.mode != 
"RGB": 
                image = image.convert( 
"RGB") 
# resize the input image and preprocess it
            image = image.resize(target) 
            image = img_to_array(image) 
            image = np.expand_dims(image, 
axis= 
0) 
            image = imagenet_utils.preprocess_input(image) 
# return the processed image
return image 

image的预处理，这里使用的是keras+PIL，和opencv之间的比较，需要有时间来做。

@app.route( 
"/predict", 
methods=[ 
"POST"]) 
def
predict(): 
# initialize the data dictionary that will be returned from the
# view
            data = { 
"success": 
False} 
# ensure an image was properly uploaded to our endpoint
if flask.request.method == 
"POST": 
if flask.request.files.get( 
"image"): 
# read the image in PIL format
                    image = flask.request.files[ 
"image"].read() 
                    image = Image.open(io.BytesIO(image)) 
# preprocess the image and prepare it for classification
                    image = prepare_image(image, 
target=( 
224, 
224)) 
# classify the input image and then initialize the list
# of predictions to return to the client
                    preds = model.predict(image) 
                    results = imagenet_utils.decode_predictions(preds) 
                    data[ 
"predictions"] = [] 
# loop over the results and add them to the list of
# returned predictions
for (imagenetID, label, prob) 
in results[ 
0]: 
                        r = { 
"label": label, 
"probability": 
float(prob)} 
                        data[ 
"predictions"].append(r) 
# indicate that the request was a success
                    data[ 
"success"] = 
True
# return the data dictionary as a JSON response
return flask.jsonify(data) 

虽然是核心部分，但是其实非常容易被复用。就是读取数据，然后进行处理的过程。

# if this is the main thread of execution first load the model and
# then start the server
if
__name__ == 
"__main__": 
print(( 
"* Loading Keras model and Flask starting server..."
"please wait until server has fully started")) 
            load_model() 
            app.run() 

比不可少的main过程。缺少不可运行。

三、运行效果

使用VPS能够更快地得到效果，至少你不需要下载resnet*.h5，一个链路不是太好的大物件。

flask的运行效果，使用curl进行处理的效果

img_f8ecebdbd07866f1b2d301b17059e52c.jpe

从结果上来看，curch排在了第2，而将这张图片识别为钟楼或者修道院、城堡，宫殿

，似乎也没有什么不妥。

四、小结反思

真的仅仅是通过了几行代码，就实现了flask部署的核心问题。不过光是跑这个简单的过程，机器就已经发出巨大的热量了；另一方面，整个的结构是什么，也需要进一步去研究清楚才对。

目前方向：图像拼接融合、图像识别联系方式：jsxyhelu@foxmail.com