TensorFlow学习记录：用TF_Serving将人脸识别模型Inception ResNet V2部署到服务器上

部署前先到这里下载已训练好的Inception ResNet V2模型。
下载后解压得到如下图所示文件

将上图所示pb文件用saved_model模块导出为带有签名的模型文件

import tensorflow as tf
import numpy as np
from tensorflow.python.saved_model import tag_constants

model_path = "/home/boss/Study/face_recognition_flask/20180402-114759/model.pb"
image_test = np.random.randint(0,10,(1,160,160,3))    # 随机生成一个[1,160,160,3]的数组，代表一张图片，用于测试

with tf.gfile.FastGFile(model_path,'rb')as f:
	graph_def = tf.GraphDef()
	graph_def.ParseFromString(f.read())
	tf.import_graph_def(graph_def,name='')

	my_graph = tf.get_default_graph()
	 # 从读取的pb文件中找到相关张量，构造要保存到saved_model模型中的张量
	images_placeholder = my_graph.get_tensor_by_name('input:0')          # 传入Inception ResNet V2的placeholder
	embeddinigs = my_graph.get_tensor_by_name('embeddings:0')          # Inception ResNet V2的输出，为人脸的特征数组，形状为[1,512]
	phase_train_placeholder = my_graph.get_tensor_by_name('phase_train:0')# 传入Inception ResNet V2告诉神经网络是否是在训练阶段，值域为[True,False]

	with tf.Session()as sess:
		builder = tf.saved_model.builder.SavedModelBuilder('tfservingmodel')
		inputs = {'input_1':tf.saved_model.utils.build_tensor_info(images_placeholder),'input_2':tf.saved_model.utils.build_tensor_info(phase_train_placeholder)}
		outputs = {'output':tf.saved_model.utils.build_tensor_info(embeddinigs)}

		signature = tf.saved_model.signature_def_utils.build_signature_def(inputs,outputs,'signame')      # 构造签名对象
		builder.add_meta_graph_and_variables(sess,[tag_constants.SERVING],{'my_signature':signature})
		builder.save()

	tf.reset_default_graph()     # 清除默认图形堆栈并重置全局默认图形。
	with tf.Session()as sess:
		meta_graph_def = tf.saved_model.loader.load(sess,[tag_constants.SERVING],'tfservingmodel')
		signature = meta_graph_def.signature_def                # 从meta_graph_def取出SignatureDef对象

		images_placeholder = signature['my_signature'].inputs['input_1'].name        # 从SignatureDef对象中找出具体的输入输出张量
		phase_train_placeholder = signature['my_signature'].inputs['input_2'].name 
		embeddinigs = signature['my_signature'].outputs['output'].name

		feed_dict = {images_placeholder:image_test,phase_train_placeholder:False}
		emb = sess.run(embeddinigs,feed_dict=feed_dict)                     # 使用一开始构造的数据测试，Inception ResNet V2模型会把人脸转换成一个[1,512]的特征数组
		print(emb)

运行结果截图（部分）

生成支持远程调用的模型，和上面的代码几乎一样，不一样的地方是在生成签名对象时加入用于预测任务的签名信息

import tensorflow as tf
from tensorflow.python.saved_model import tag_constants

model_path = "/home/boss/Study/face_recognition_flask/20180402-114759/model.pb"

with tf.gfile.FastGFile(model_path,'rb')as f:
	graph_def = tf.GraphDef()
	graph_def.ParseFromString(f.read())
	tf.import_graph_def(graph_def,name='')

	my_graph = tf.get_default_graph()
	images_placeholder = my_graph.get_tensor_by_name('input:0')
	embeddinigs = my_graph.get_tensor_by_name('embeddings:0')
	phase_train_placeholder = my_graph.get_tensor_by_name('phase_train:0')

	with tf.Session()as sess:
		builder = tf.saved_model.builder.SavedModelBuilder('tfservingmodelv1')
		inputs = {'input_1':tf.saved_model.utils.build_tensor_info(images_placeholder),'input_2':tf.saved_model.utils.build_tensor_info(phase_train_placeholder)}
		outputs = {'output':tf.saved_model.utils.build_tensor_info(embeddinigs)}

		signature = tf.saved_model.signature_def_utils.build_signature_def(inputs=inputs,outputs=outputs,method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)
		builder.add_meta_graph_and_variables(sess,[tag_constants.SERVING],{'my_signature':signature})
		builder.save()

运行结果截图

在Linux中开启TF_Serving服务前，先构建模型版本号

cd tfservingmodelv1/
mkdir 123456
mv saved_model.pb 123456/
mv variables 123456/

开启服务

tensorflow_model_server --port=9000 --model_base_path=/home/boss/Study/face_recognition_flask/20180402-114759/tfservingmodelv1/ --model_name=md --rest_api_port=8500

如果出现与下面类似的图片，代表服务启动成功

编写一个脚本发送HTTP POST请求测试

import requests
import json
import numpy as np

url = "http://localhost:8500/v1/models/md/versions/123456:predict"
image_test = np.random.randint(0,10,(1,160,160,3))
def default(obj):
	if isinstance(obj,(np.ndarray,)):
		return obj.tolist()
s = json.dumps({"inputs":{'input_1':default(image_test),'input_2':False},"signature_name":"my_signature"})    # numpy不能转换成json格式，所以要先把numpy转换成list，再转json
r = requests.post(url,data=s)
print(r.text)

运行结果截图（部分）

参考书籍：《深度学习之TensorFlow工程化项目实战》 李金洪 编著