tensorflow serving

1.安装tensorflow serving

   1.1确保当前环境已经安装并可运行tensorflow

   从github上下载源码

git clone --recurse-submodules https://github.com/tensorflow/serving

   

   进入到serving目录下的tensorflow运行./configure,并安装步骤完成（需将 2问题解决的的步骤全操作完后执行安装步骤）

1.2.编译example代码

bazel build tensorflow_serving/example/...

  

 1.3.运行mnist例子导出model到/tmp/mnist_export目录下，目录下会根据export_version创建一个目录名为 /tmp/mnist_export/00000001

rm -rf /tmp/mnist_export/(第一次执行不存在，不必操作)

 

bazel-bin/tensorflow_serving/example/mnist_export --training_iteration=10000 --export_version=1 /tmp/mnist_export

 

Training model...

('Extracting', '/tmp/train-images-idx3-ubyte.gz')

('Extracting', '/tmp/train-labels-idx1-ubyte.gz')

('Extracting', '/tmp/t10k-images-idx3-ubyte.gz')

('Extracting', '/tmp/t10k-labels-idx1-ubyte.gz')

training accuracy 0.9219

Done training!

Exporting trained model to /tmp/mnist_export

Done exporting!

  1.4 执行inference开启服务，端口9000，目录指向之前导出的目录

bazel-bin/tensorflow_serving/example/mnist_inference --port=9000 /tmp/mnist_export/00000001

I tensorflow_serving/session_bundle/session_bundle.cc:130] Attempting to load a SessionBundle from: /tmp/mnist_export/00000001

I tensorflow_serving/session_bundle/session_bundle.cc:107] Running restore op for SessionBundle

I tensorflow_serving/session_bundle/session_bundle.cc:178] Done loading SessionBundle

I tensorflow_serving/example/mnist_inference.cc:163] Running...

1.6 执行测试client

bazel-bin/tensorflow_serving/example/mnist_client --num_tests=1000 --server=localhost:9000

 

('Extracting', '/tmp/train-images-idx3-ubyte.gz')

('Extracting', '/tmp/train-labels-idx1-ubyte.gz')

('Extracting', '/tmp/t10k-images-idx3-ubyte.gz')

('Extracting', '/tmp/t10k-labels-idx1-ubyte.gz')

........................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................................

Inference error rate: 9.2%

 

2.问题解决

no such package '@boringssl_git//': Error cloning repository:https://boringssl.googlesource.com/boringssl: cannot open git-upload-pack and referenced by '//external:libssl'.

 由于GFW把google的很多地址给墙了，所以无法下载相关的内容，修改seving目录下的tensorflow/tensorflow/workspace.bzl 文件相关的repository

git_repository(

name = "boringssl_git",

#commit = "436432d849b83ab90f18773e4ae1c7a8f148f48d",

commit = "db0729054d5964feab9e60089ba2d06a181e78b1",

init_submodules = True,

#remote = "https://github.com/mdsteele/boringssl-bazel.git",

remote = "https://github.com/doubler/boringssl-bazel.git",

)

https://github.com/tensorflow/serving/issues/6

 

 运行mnist_client时报错

Traceback (most recent call last):

File "/root/tensorflow-serving/bazel-bin/tensorflow_serving/example/mnist_client.runfiles/__main__/tensorflow_serving/example/mnist_client.py", line 34, in <module>

from grpc.beta import implementations

ImportError: No module named grpc.beta

 使用pip安装grpcio模块

pip install grpcio

https://github.com/grpc/grpc/tree/master/src/python/grpcio

 

export过程报错缺少 manifest_pb2.py 的解决方法：

首先编译serving下的example目录得到

bazel-bin/tensorflow_serving/example/mnist_export.runfiles/org_tensorflow/tensorflow/contrib/session_bundle/manifest_pb2.py

随后copy到python主目录下的lib目录 例如：/usr/lib/python2.7/site-packages/

如果还报相同错误

修改 /usr/lib/python2.7/site-packages/tensorflow/contrib/session_bundle/目录下的exporter.py文件

删除 from tensorflow.contrib.session_bundle import manifest_pb2

增加 import manifest_pb2

解决思路：主要就是PYTHONPATH中缺少manifest_pb2.py，所以需要设置，设置路径可以找到manifest_pb2.py即可

manifest_pb2.py为tensorflow/contrib/session_bundle/manifest.proto 的protobuf生成文件，若有条件可手动生成

 

3.参考地址

  tensorflow serving github : https://github.com/tensorflow/serving

https://github.com/tensorflow/serving/blob/master/tensorflow_serving/g3doc/setup.md

  bazel http://www.bazel.io/ (需FQ)

 

4.Serving Framework  

 

4.1.Train 

  训练模型的过程

4.2.exporter

  负责将训练好的模型导出

4.3.Sever

  负责存储操作，例如将对象存储到磁盘

4.4.Server

  提供grpc server,组织request调用Module，将结果response client

4.5.ModuleManager

  负责加载训练好的模型

4.6.Scheduler

  负责请求的调度，例如BatchScheduler(buffer 某一批数据才发给Service)

4.7.client

 负责发送Request请求接收Response

 

5.如何编写Serving

5.1 export model

     模型训练完成后，需要export model

     1) 需要确定 signature : (classification_signature,regression_signature,generic_signature)

         classification_signature: input , classes , scores

         regression_signature: input , output

         generic_signature: map<string,tensor_name>

 

         signature规定了输入和输出的tensor_name, 这个tensor_name应该对应到graph里的tensor

         例如 mnist 为classification模型 训练输入了 x 训练出 y 则在export的时候使用：

signature = exporter.classification_signature(input_tensor=x, scores_tensor=y)

    

5.2 确定输出的路径

      导出model需要一个可存储的路径，这个路径会在inference程序读取时使用

 

5.3 编写inference

      inference主要流程：

      1）获得    SessionBundle

std::unique_ptr<SessionBundleFactory> bundle_factory;

 

TF_QCHECK_OK(

 

SessionBundleFactory::Create(session_bundle_config, &bundle_factory));

 

std::unique_ptr<SessionBundle> bundle(new SessionBundle);

 

TF_QCHECK_OK(bundle_factory->CreateSessionBundle(bundle_path, &bundle));

      2)  提供输入与输出的tensor

Tensor input(tensorflow::DT_FLOAT, {1, kImageDataSize});

 

    std::copy_n(request->image_data().begin(), kImageDataSize,

 

    input.flat<float>().data());

 

std::vector<Tensor> outputs;

 

      3)  通过signatrue传入input,output到session_bundle并执行

const tensorflow::Status status = bundle_->session->Run(

 

        {{signature_.input().tensor_name(), input}},

 

        {signature_.scores().tensor_name()}, {}, &outputs);