[TextMatch框架] server

 

TextMatch

TextMatch is a semantic matching model library for QA & text search … It’s easy to train models and to export representation vectors.

 

TextMatch/server模块包含 ：（基于MLFlow的server）

model_server.py

import os
import sys
import time
import platform
import argparse
import cloudpickle
import numpy as np
import mlflow.pyfunc
from pyspark.sql import DataFrame
from sklearn.externals import joblib
from textmatch.config.constant import Constant as const
from textmatch.models.text_embedding.model_factory_sklearn import ModelFactory


cur_abs_dir = os.path.dirname(os.path.abspath(__file__))
code_home = cur_abs_dir
sys.path.insert(0, code_home)
print('[model_server] python version: %s' % platform.python_version())
print('[model_server] code_home: %s' % code_home)

start = time.time()
exec_time = time.time() - int(time.time()) % 900
local_time = time.localtime(exec_time - 30 * 60)
exec_day = time.strftime('%Y-%m-%d', local_time)
exec_hour = time.strftime('%H', local_time)
exec_minute = time.strftime('%M', local_time)


class TextMatchWrapper(mlflow.pyfunc.PythonModel):
    def __init__(self, experiment_name, version_name):
        self.experiment_name = experiment_name
        self.version_name = version_name
        self.wordstest_dict = {"0":"我去玉龙雪山并且喜欢玉龙雪山玉龙雪山", "1":"我在玉龙雪山并且喜欢玉龙雪山", "2":"我在九寨沟", "3":"你好"}   #["我去玉龙雪山并且喜欢玉龙雪山玉龙雪山","我在玉龙雪山并且喜欢玉龙雪山","我在九寨沟"]
        self.mf = ModelFactory( match_models=['bow', 'tfidf', 'ngram_tfidf'] )

    def load_context(self, context):
        # wordstest_dict = context.artifacts["wordstest_dict"]
        self.mf.init(words_dict=self.wordstest_dict, update=True)

    def predict(self, context, model_input):
        #print('model_input>>>', model_input)
        #print('model_input[text]>>>', model_input["text"].values)
        return self.mf.predict(model_input["text"].values[0])



# 模型预测主流程
def model_server(experiment_name, version_name, args): #, model_path='./data/'):

    artifacts = {
        "train_model": os.path.join(const.BOW_DIC_PATH )
    }
    if args.local_store:
        mlflow.pyfunc.save_model(path=args.model_file,
                                python_model=TextMatchWrapper(experiment_name, version_name),
                                artifacts=artifacts)
    else:
        mlflow.pyfunc.log_model(artifact_path=args.model_file,
                                 python_model=TextMatchWrapper(experiment_name, version_name),
                                 artifacts=artifacts)



def parse_argvs():
    parser = argparse.ArgumentParser(description='textmatch ---- 模型线上部署')
    parser.add_argument("--experiment_name", help="实验名称")
    parser.add_argument("--version_name", help="版本名称")
    parser.add_argument("--model_file", help="模型存储路径",default='model')
    parser.add_argument("--local_store", help="是否本地存储",action='store_true', default=True)
    args = parser.parse_args()
    print('[model_predictor] args: %s' % args)

    return parser, args

# python model_server.py --experiment_name "textmatch" --version_name "001" --model_file "textmodel"
# mlflow models serve -m /Users/qudian/Desktop/TextMatch/textmodel/ -h 0.0.0.0 -w 3 -p 5000 --no-conda
if __name__ == '__main__':
    parser, args = parse_argvs()
    # 输入参数解析
    experiment_name = args.experiment_name      # 实验名称
    input_version_name = args.version_name      # 输入的版本名称
    model_server(experiment_name=experiment_name, version_name=input_version_name, args=args)

    end = time.time()
    print('运行时长: {}s'.format(int(end - start)))

 

构建mlflow server模型

python model_server.py --experiment_name "textmatch" --version_name "001" --model_file "textmodel"

启动服务

mlflow models serve -m /Users/qudian/Desktop/TextMatch/textmodel/ -h 0.0.0.0 -w 3 -p 5000 --no-conda

 

post_data.py

import requests
import pandas as pd
import json
import time 


for i in range(100):
    # 构造需要进行推断的数据
    newJson = '{"text":"我在九寨沟,很喜欢"}'
    # 指定ip, 端口
    url = "http://127.0.0.1:5000/invocations"
    # 传递的参数需要从dataframe转化为json格式
    json_data = json.loads( newJson )
    model_input = pd.DataFrame([json_data])
    req_data = model_input.to_json(orient='split')
    headers = {'content-type': 'application/json; format=pandas-split'}
    # 使用POST方式调用REST api
    start_time = time.time()
    respond = requests.request("POST", url, data=req_data, headers=headers) 
    print ("time>>>>>>>", time.time() - start_time)
    print ( "respond>>", respond )
    # 获取返回值
    print (respond.json())