1.什么是TFserving
当你训好你的模型,需要提供给外部使用的时候,你就需要把模型部署到线上,并提供合适的接口给外部调用。你可能会考虑一些问题:
目前流行的深度学习框架Tensorflow和Pytorch, Pytorch官方并没有提供合适的线上部署方案;Tensorflow则提供了TFserving方案来部署线上模型推理。另外,Model Server for Apache MXNet 为MXNet模型提供推理服务。
本文为TFServing的使用指南。如果你是pytorch或者MXNet模型,也可以通过ONNX转成TFserving的模型,部署在TFServing上。
那什么是TFserving?
TFserving是Google 2017推出的线上推理服务;采用C/S架构,客户端可通过gRPC和RESTfull API与模型服务进行通信。
TFServing的特点:
2.TFserving安装
强烈建议采用docker方式安装TFserving,安装依赖docker和nvidia-docker(TFserving的gpu需要)
- #安装yum-utils工具和device-mapper相关依赖包
- yum install -y yum-utils \
- device-mapper-persistent-data \
- lvm2
- #添加docker-ce stable版本的仓库
- yum-config-manager \
- --add-repo \
- https://download.docker.com/linux/centos/docker-ce.repo
- #更新yum缓存文件
- yum makecache fast
- #查看所有可安装的docker-ce版本
- yum list docker-ce --showduplicates | sort -r
- # 安装docker-ce
- yum install docker-ce-17.12.1.ce-1.el7.centos
- #允许开机启动docker-ce服务
- systemctl enable docker.service
- #启动Docker-ce服务
- systemctl start docker
- #运行测试容器hello-world
- docker run --rm hello-world
- # 安装nvidia-docker2
- yum install -y nvidia-docker2-2.0.3-1.docker17.12.1.ce
- # 重启docker服务
- service docker restart
- docker pull tensorflow/serving:latest-gpu
- # 可以选择其他版本如 docker pull tensorflow/serving:1.14.0-rc0-gpu
注意:docker版本和nvidia-docker要匹配
3.TFserving使用说明
3.1 模型转换
TFserving的模型需要转换成TFserving的格式, 不支持通常的checkpoint和pb格式。
TFserving的模型包含一个.pb文件和variables目录(可以为空),导出格式如下:.
- ├── 1
- │ ├── saved_model.pb
- │ └── variables
- ├── 2
- │ ├── saved_model.pb
- │ └── variables
不同的深度学习框架的转换路径:
- (1) pytorch(.pth)--> onnx(.onnx)--> tensorflow(.pb) --> TFserving
- (2) keras(.h5)--> tensorflow(.pb) --> TFserving
- (3) tensorflow(.pb) --> TFserving
这里详细介绍下pb转换成TFserving模型
- import tensorflow as tf
- def create_graph(pb_file):
- """Creates a graph from saved GraphDef file and returns a saver."""
- # Creates graph from saved graph_def.pb.
- with tf.gfile.FastGFile(pb_file, 'rb') as f:
- graph_def = tf.GraphDef()
- graph_def.ParseFromString(f.read())
- _ = tf.import_graph_def(graph_def, name='')
- def pb_to_tfserving(pb_file, export_path, pb_io_name=[], input_node_name='input', output_node_name='output', signature_name='default_tfserving'):
- # pb_io_name 为 pb模型输入和输出的节点名称,
- # input_node_name为转化后输入名
- # output_node_name为转化后输出名
- # signature_name 为签名
- create_graph(pb_file)
- # tensor_name_list = [tensor.name for tensor in tf.get_default_graph().as_graph_def().node]
- input_name = '%s:0' % pb_io_name[0]
- output_name = '%s:0' % pb_io_name[1]
- with tf.Session() as sess:
- in_tensor = sess.graph.get_tensor_by_name(input_name)
- out_tensor = sess.graph.get_tensor_by_name(output_name)
- builder = tf.saved_model.builder.SavedModelBuilder(export_path) ## export_path导出路径
- inputs = {input_node_name: tf.saved_model.utils.build_tensor_info(in_tensor)}
- outputs = {output_node_name: tf.saved_model.utils.build_tensor_info(out_tensor)}
- signature = tf.saved_model.signature_def_utils.build_signature_def(
- inputs, outputs, method_name=tf.saved_model.signature_constants.PREDICT_METHOD_NAME)
- builder.add_meta_graph_and_variables(
- sesssess=sess, tags=[tf.saved_model.tag_constants.SERVING],
- signature_def_map={signature_name: signature}, clear_devices=True) ## signature_name为签名,可自定义
- builder.save()
- pb_model_path = 'test.pb'
- pb_to_tfserving(pb_model_path, './1', pb_io_name=['input_1_1','output_1'],signature_name='your_model')
3.2 TFserving配置和启动
模型导出后,同一个模型可以导出不同的版本(版本后数字),可以TFserving配置中指定模型和指定版本。TFserving的模型是通过模型名称和签名来唯一定位。TFserving 可以配置多个模型,充分利用GPU资源。
- # models.config
- model_config_list {
- config {
- name: 'your_model'
- base_path: '/models/your_model/'
- model_platform: 'tensorflow'
- # model_version_policy {
- # specific {
- # versions: 42
- # versions: 43
- # }
- # }
- # version_labels {
- # key: 'stable'
- # value: 43
- # }
- # version_labels {
- # key: 'canary'
- # value: 43
- # }
- }
- config {
- name: "mnist",
- base_path: "/models/mnist",
- model_platform: "tensorflow",
- model_version_policy: {
- specific: {
- versions: 1,
- versions: 2
- }
- }
- }
- # 可以通过model_version_policy 进行版本的控制
- # 建议把模型和配置文件放在docker外的本地路径,如/home/tfserving/models, 通过-v 挂载到docker内部
- # --model_config_file: 指定模型配置文件
- # -e NVIDIA_VISIBLE_DEVICES=0: 指定GPU
- # -p 指定端口映射 8500为gRpc 8501为restful api端口
- # -t 为docker镜像
- nvidia-docker run -it --privileged -d -e NVIDIA_VISIBLE_DEVICES=0 -v /home/tfserving/models:/models -p 8500:8500 -p 8501:8501 \
- -t tensorflow/serving:latest-gpu \
- --model_config_file=/models/models.config
- # /home/tfserving/models 结构
- ├── models.config
- └── your_model
- ├── 1
- │ ├── saved_model.pb
- │ └── variables
- └── 2
- ├── saved_model.pb
- └── variables
- # test
- curl http://192.168.0.3:8501/v1/models/your_model
- {
- "model_version_status": [
- {
- "version": "2",
- "state": "AVAILABLE",
- "status": {
- "error_code": "OK",
- "error_message": ""
- }
- }
- ]
- }
- # 其他启动方式
- # 如果多个模型在不同的目录,可以通过-mount 单独加载
- nvidia-docker run -it --privileged -d -e NVIDIA_VISIBLE_DEVICES=0 \
- --mount type=bind,source=/home/tfserving/models/your_model,target=/models/your_model \
- --mount type=bind,source=/home/tfserving/models/your_model/models.config,target=/models/models.config \
- -p 8510:8500 -p 8501:8501 \
- -t tensorflow/serving:latest-gpu \
- --model_config_file=/models/models.config
3.3 TFserving服务调用
客户端可以通过gRpc和http方式调用TFserving服务模型,支持多种客户端语言,这里提供python的调用方式; 调用都是通过模型名称和签名来唯一对应一个模型
- #
- # -*-coding:utf-8 -*-
- import tensorflow as tf
- from tensorflow_serving.apis import predict_pb2
- from tensorflow_serving.apis import prediction_service_pb2_grpc
- import grpc
- import time
- import numpy as np
- import cv2
- class YourModel(object):
- def __init__(self, socket):
- """
- Args:
- socket: host and port of the tfserving, like 192.168.0.3:8500
- """
- self.socket = socket
- start = time.time()
- self.request, selfself.stub = self.__get_request()
- end = time.time()
- print('initialize cost time: ' + str(end - start) + ' s')
- def __get_request(self):
- channel = grpc.insecure_channel(self.socket, options=[('grpc.max_send_message_length', 1024 * 1024 * 1024),
- ('grpc.max_receive_message_length', 1024 * 1024 * 1024)]) # 可设置大小
- stub = prediction_service_pb2_grpc.PredictionServiceStub(channel)
- request = predict_pb2.PredictRequest()
- request.model_spec.name = "your_model" # model name
- request.model_spec.signature_name = "your_model" # model signature name
- return request, stub
- def run(self, image):
- """
- Args:
- image: the input image(rgb format)
- Returns: embedding is output of model
- """
- img = image[..., ::-1]
- self.request.inputs['input'].CopyFrom(tf.contrib.util.make_tensor_proto(img)) # images is input of model
- result = self.stub.Predict(self.request, 30.0)
- return tf.make_ndarray(result.outputs['output'])
- def run_file(self, image_file):
- """
- Args:
- image_file: the input image file
- Returns:
- """
- image = cv2.imread(image_file)
- image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
- return self.run(image)
- if __name__ == '__main__':
- model = YourModel('192.168.0.3:8500')
- test_file = './test.jpg'
- result = model.run_file(test_file)
- print(result)
- # [8.014745e-05 9.999199e-01]
- import cv2
- import requests
- class SelfEncoder(json.JSONEncoder):
- def default(self, obj):
- if isinstance(obj, np.ndarray):
- return obj.tolist()
- elif isinstance(obj, np.floating):
- return float(obj)
- elif isinstance(obj, bytes):
- return str(obj, encoding='utf-8');
- return json.JSONEncoder.default(self, obj)
- image_file = '/home/tfserving/test.jpg'
- image = cv2.imread(image_file)
- image = cv2.cvtColor(image, cv2.COLOR_BGR2RGB)
- img = image[..., ::-1]
- input_data = {
- "signature_name": "your_model",
- "instances": img
- }
- data = json.dumps(input_data, cls=SelfEncoder, indent=None)
- result = requests.post("http://192.168.0.3:8501/v1/models/your_model:predict", datadata=data)
- eval(result .content)
- # {'predictions': [8.01474525e-05, 0.999919891]}
5.总结
本文介绍了TFserving部署线上推理服务,从模型的转换,部署启动和调用推理,欢迎交流,希望对你有帮助。我们来回答下开篇提出的问题
网站名称:用TFserving部署深度学习模型
文章起源:http://www.shufengxianlan.com/qtweb/news8/378408.html
网站建设、网络推广公司-创新互联,是专注品牌与效果的网站制作,网络营销seo公司;服务项目有等
声明:本网站发布的内容(图片、视频和文字)以用户投稿、用户转载内容为主,如果涉及侵权请尽快告知,我们将会在第一时间删除。文章观点不代表本网站立场,如需处理请联系客服。电话:028-86922220;邮箱:631063699@qq.com。内容未经允许不得转载,或转载时需注明来源: 创新互联