tensorflow serving 编写配置文件platform_config_file的方法

1、安装grpc

gRPC 的安装：

$ pip install grpcio

安装 ProtoBuf 相关的 python 依赖库：

$ pip install protobuf

安装 python grpc 的 protobuf 编译工具：

$ pip install grpcio-tools

2、在serving目录运行脚本，生成*_pb2.py文件

 # run at root of tensorflow_serving repo

 TARGET_DIR="$1"

 python -m grpc.tools.protoc \
     -I . -I ./tensorflow \
     --python_out "$TARGET_DIR" \
     tensorflow_serving/servables/tensorflow/saved_model_bundle_source_adapter.proto \
     tensorflow_serving/servables/tensorflow/session_bundle_config.proto \
     tensorflow_serving/config/platform_config.proto

 pushd $TARGET_DIR

 touch tensorflow_serving/__init__.py
 touch tensorflow_serving/config/__init__.py
 touch tensorflow_serving/servables/__init__.py
 touch tensorflow_serving/servables/tensorflow/__init__.py

 popd

sh gen-tf-serving-proto-py.sh /tmp

3、将生成的*_pb2.py文件cp出来

cp -r /tmp/tensorflow_serving .

4、在当前目录运行gen-platform-config.py

 # -*- coding: utf-8 -*-

 import tensorflow as tf

 from tensorflow_serving.config import platform_config_pb2
 from tensorflow_serving.servables.tensorflow import session_bundle_config_pb2
 from tensorflow_serving.servables.tensorflow import saved_model_bundle_source_adapter_pb2

 session_config = tf.ConfigProto()
 # config whatever you want
 session_config.gpu_options.allow_growth = True
 session_config.gpu_options.per_process_gpu_memory_fraction = 0.4

 legacy_config=session_bundle_config_pb2.SessionBundleConfig(session_config=session_config)
 adapter = saved_model_bundle_source_adapter_pb2.SavedModelBundleSourceAdapterConfig(legacy_config=legacy_config)

 config_map = platform_config_pb2.PlatformConfigMap()
 config_map.platform_configs['tensorflow'].source_adapter_config.Pack(adapter)

 print(config_map)

5、生成platform_config_file.cfg文件

 platform_configs {
   key: "tensorflow"
   value {
     source_adapter_config {
       [type.googleapis.com/tensorflow.serving.SavedModelBundleSourceAdapterConfig] {
         legacy_config {
           session_config {
             gpu_options {
               per_process_gpu_memory_fraction: 0.4
               allow_growth: true
             }
           }
         }
       }
     }
   }
 }

6、运行tf_serving时添加参数--platform_config_file=./conf/platform_config_file.cfg

7、若同时需要配置batching_parameters_file，则需要将batching参数写入到platform_config_file.cfg内

 platform_configs {
   key: "tensorflow"
   value {
     source_adapter_config {
       [type.googleapis.com/tensorflow.serving.SavedModelBundleSourceAdapterConfig] {
         legacy_config {
       batching_parameters {
         max_batch_size { value:  }
         batch_timeout_micros { value:  }
         max_enqueued_batches { value:  }
         num_batch_threads { value:  }
           }
           session_config {
             allow_soft_placement: true
             gpu_options {
               per_process_gpu_memory_fraction: 0.4
               allow_growth: true
             }
           }
         }
       }
     }
   }
 }

详细信息参照：https://github.com/tensorflow/serving/issues/342

我运行后生成的cfg文件为

 platform_configs {
   key: "tensorflow"
   value {
     source_adapter_config {
       type_url: "type.googleapis.com/tensorflow.serving.SavedModelBundleSourceAdapterConfig"
       value: "\302>\017\022\r2\013\t\232\231\231\231\231\231\331? \001"
     }
   }
 }

并不能生成清晰的text格式的配置文件，目前还未找到原因