在tensorflow中使用字符串张量时，grpc服务器内存泄漏？

问题描述

Ubutun 18，python 3.6，gRPC 1.24.3

server.py：

import time
import grpc
from concurrent import futures
import tensorflow as tf
from google.protobuf import any_pb2
import numpy as np
import random
import compute_pb2,compute_pb2_grpc  


class ComputeServicer(compute_pb2_grpc.ComputeServicer):
    def SayHello(self,request,ctx):
        max_len = str(len(request.helloworld))
        return compute_pb2.HelloReply(result=max_len)

    def Compute(self,ctx):
        n = 100
        # This 2 lines makes memory leak
        for i in range(n):
            value = tf.constant([str(random.randint(10000000,99999999)) * 16] * 5000000,dtype=tf.string)
        '''
        for i in range(n):
            value = tf.constant([i] * 100000000,dtype=tf.int64)
        ''' 
        value = tf.constant([1] * 50000000,dtype=tf.int64)
        tensor_proto = tf.make_tensor_proto(value)
        any_pb = any_pb2.Any()
        any_pb.Pack(tensor_proto)
        value_proto = compute_pb2.Value(tensor=any_pb)
        return compute_pb2.ComputeResponse(value=value_proto)


def main():
    options = [
        ('grpc.max_send_message_length',-1),('grpc.max_receive_message_length',-1)
    ]
    server_kwargs = {}
    server_kwargs['options'] = options
    server = grpc.server(futures.ThreadPoolExecutor(max_workers=10),**server_kwargs)
    servicer = ComputeServicer()
    compute_pb2_grpc.add_ComputeServicer_to_server(servicer,server)
    server.add_insecure_port('127.0.0.1:19999')
    server.start()
    try:
        print("running...")
        time.sleep(1000)
    except KeyboardInterrupt:
        print("stopping...")
        server.stop(0)


if __name__ == '__main__':
    main()

client.py

import grpc
import compute_pb2
import compute_pb2_grpc
import tensorflow as tf
import time

_HOST = '127.0.0.1'
_PORT = '19999'


def main():
    num = 0
    while num < 100:
        with grpc.insecure_channel("{0}:{1}".format(_HOST,_PORT),options=[('grpc.max_send_message_length',-1)]
                                   ) as channel:
            client = compute_pb2_grpc.ComputeStub(channel=channel)
            response = client.Compute(compute_pb2.HelloRequest(helloworld="123456"))
            value_proto = response.value
            tensor_proto = tf.make_tensor_proto(values=0)
            if not value_proto.tensor.Unpack(tensor_proto):
                raise ValueError('Unable to unpack the received tensor value.')

            tensor_value = tf.make_ndarray(tensor_proto)
            print(num)
            num+=1



if __name__ == '__main__':
    main()

compute.proto

Syntax = "proto3"; 
import "google/protobuf/any.proto";


package compute;
service Compute {
    rpc SayHello (HelloRequest) returns (HelloReply) {}
    rpc Compute(HelloRequest) returns (ComputeResponse) {}
}

message HelloRequest {
    string helloworld = 1;
}

message HelloReply {
    string result = 1;
}

message ComputeResponse {
  Value value = 1;
}

message Value {
  oneof value {
    google.protobuf.Any tensor = 1;
  }
}

python -m grpc_tools.protoc -I ./ --python_out =。/ --grpc_python_out =。/ compute.proto

python server.py

python client.py

使用上述运行命令时，服务器端的内存使用量会随着时间增加。如果我删除服务器中的字符串张量或将其替换为int64（请参阅我的注释），则服务器的内存使用情况是稳定的。

谢谢。

解决方法

暂无找到可以解决该程序问题的有效方法，小编努力寻找整理中！

如果你已经找到好的解决方法，欢迎将解决方案带上本链接一起发送给小编。

小编邮箱:dio#foxmail.com (将#修改为@）