TZRPC

让深度学习调用： 简单！！！ 高效！！！

深度学习模型在落地时需要提供高效快速交互接口，业务逻辑和深度模型解码通常运行在不同类型的机器上。 Http 并不适合大量数据的交互，而RPC (Remote Procedure Call) 远程过程调用, 而RPC在TCP层实现。提高了开发效率，算法工程师可以不必花费更多精力放在具体的接口实现上，而是专注于算法优化上。

tzrpc 框架基于google的 grpc 实现，需要Python 3.7及以上， 支持流式传输！！！

目前支持以下基础类型：

TZRPC 类型	python 类型	是否支持
String	str	✅
Integer	int	✅
Float	float	✅
Double	float	✅
Boolean	bool	✅
Bytes	bytes	✅
Numpy	numpy	✅
Tensor	torch.Tensor	✅
使用pickle模块反序列化	任意python对象	✅

快速使用

安装

pip install tzrpc

或拉取最新代码，安装

git clone https://github.com/lovemefan/TZRPC.git
cd TZRPC
pip install -e .

服务端 Server.py

from tzrpc import TZRPC_Server
import numbers
server = TZRPC_Server(__name__)


@server.register
def say_hello(text):
    return "hello world " + text

@server.register
def send_numpy_obj(data):
    return data * 2 + 1

@server.register
def send_torch_tensor_obj(data):
    return data @ data.T

@server.register
def send_bytes(data: bytes):
    return data + data

@server.register
def send_number(data: numbers.Number):
    return data * 2


@server.register
def send_bool(_bool: bool):
    return not _bool


@server.register
def send_python_obj(data):
    return data

@server.register(stream=True)
def gumbel(num):
    if num % 3 == 0:
        yield f"number is {num}, you win"

if __name__ == '__main__':
    server.run("localhost", 8000)

客户端 Client.py

from tzrpc import TZPRC_Client
import numpy as np
import torch
import numbers
SERVER_ADDRESS = "localhost:8000"
client = TZPRC_Client(SERVER_ADDRESS)


@client.register
def say_hello(text):
    return text

@client.register
def send_numpy_obj():
    data = np.array([[1, 2, 3], [4, 5, 6]])
    return data

@client.register
def send_torch_tensor_obj():
    data = torch.tensor([[1, 2, 3], [4, 5, 6]])
    return data

@client.register
def send_bytes():
    return b"just for test"

@client.register
def send_number(data: numbers.Number):
    return data


@client.register
def send_bool(_bool: bool):
    return _bool

@client.register
def send_python_obj(data):
    return data

@client.register(stream=True)
def gumbel(num):
    for i in range(num):
        yield i

if __name__ == '__main__':
    print(say_hello("lovemefan"))
    print(send_numpy_obj())
    print(send_torch_tensor_obj())
    print(send_bytes())
    print(send_number(2))
    print(send_number(1/3))
    print(send_bool(True))
    print(send_bool(False))
    
    class testOb:
        def __init__(self, name, age):
            self.name = name
            self.age = age

    python_obj = testOb("test_name", 20)
    print(send_python_obj(python_obj).__dict__)
    
    # 流式demo
    for i in gumbel(10):
        print(i)

客户端输出

hello world lovemefan

[[ 3  5  7]
 [ 9 11 13]]
 
tensor([[14, 32],
        [32, 77]])
        
b'just for testjust for test'

4

0.6666666666666666

False

True

{'name': 'test_name', 'age': 20}

number is 0, you win
number is 3, you win
number is 6, you win
number is 9, you win