背景

mlflow是Databrick开源的机器学习管理平台，它很好的解藕了算法训练和算法模型服务，使得算法工程师专注于模型的训练，而不需要过多的关注于服务的,
而且在我们公司已经有十多个服务稳定运行了两年多。

搭建

mlflow的搭建主要是mlflow tracking server的搭建，tracking server主要是用于模型的元数据以及模型的数据存储
我们这次以minio作为模型数据的存储后台，mysql作为模型元数据的存储，因为这种模式能满足线上的需求，不仅仅是用于测试

• minio的搭建
  参考我之前的文章MinIO的搭建使用，并且创建名为mlflow的bucket，便于后续操作

• mlflow的搭建

  • conda的安装
    参照install conda,根据自己的系统安装不同的conda环境
  • mlfow tracking server安装

  # 创建conda环境 并安装 python 3.6  
  conda create -n mlflow-1.11.0 python==3.6
  #激活conda环境
  conda activate mlflow-1.11.0
  # 安装mlfow tracking server python需要的依赖包
  pip install mlflow==1.11.0 
  pip install mysqlclient
  pip install boto3
  

  • mlflow tracking server的启动

  暴露出minio url以及需要的ID和KEY，因为mlflow tracking server在上传模型文件时需要        
  export AWS_ACCESS_KEY_ID=AKIAIOSFODNN7EXAMPLE      
  export AWS_SECRET_ACCESS_KEY=wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY      
  export MLFLOW_S3_ENDPOINT_URL=http://localhost:9001     
  mlflow server \
     --backend-store-uri mysql://root:AO,h07ObIeH-@localhost/mlflow_test \
     --host 0.0.0.0 -p 5002 \
     --default-artifact-root s3://mlflow
  

  访问localhost:5002, 就能看到如下界面:
  

  mlflow.png

使用

拷贝以下的wine.py文件

import os
import warnings
import sys

import pandas as pd
import numpy as np
from sklearn.metrics import mean_squared_error, mean_absolute_error, r2_score
from sklearn.model_selection import train_test_split
from sklearn.linear_model import ElasticNet
import mlflow.sklearn


def eval_metrics(actual, pred):
  rmse = np.sqrt(mean_squared_error(actual, pred))
  mae = mean_absolute_error(actual, pred)
  r2 = r2_score(actual, pred)
  return rmse, mae, r2


if __name__ == "__main__":
  warnings.filterwarnings("ignore")
  np.random.seed(40)

  # Read the wine-quality csv file (make sure you're running this from the root of MLflow!)
  wine_path = os.path.join(os.path.dirname(os.path.abspath(__file__)), "wine-quality.csv")
  data = pd.read_csv(wine_path)

  # Split the data into training and test sets. (0.75, 0.25) split.
  train, test = train_test_split(data)

  # The predicted column is "quality" which is a scalar from [3, 9]
  train_x = train.drop(["quality"], axis=1)
  test_x = test.drop(["quality"], axis=1)
  train_y = train[["quality"]]
  test_y = test[["quality"]]

  alpha = float(sys.argv[1]) if len(sys.argv) > 1 else 0.5
  l1_ratio = float(sys.argv[2]) if len(sys.argv) > 2 else 0.5
  mlflow.set_tracking_uri("http://localhost:5002")
  client = mlflow.tracking.MlflowClient()
  mlflow.set_experiment('http_metrics_test')
  with mlflow.start_run():
      lr = ElasticNet(alpha=alpha, l1_ratio=l1_ratio, random_state=42)
      lr.fit(train_x, train_y)

      predicted_qualities = lr.predict(test_x)

      (rmse, mae, r2) = eval_metrics(test_y, predicted_qualities)

      print("Elasticnet model (alpha=%f, l1_ratio=%f):" % (alpha, l1_ratio))
      print("  RMSE: %s" % rmse)
      print("  MAE: %s" % mae)
      print("  R2: %s" % r2)

      mlflow.log_param("alpha", alpha)
      mlflow.log_param("l1_ratio", l1_ratio)
      mlflow.log_metric("rmse", rmse)
      mlflow.log_metric("r2", r2)
      mlflow.log_metric("mae", mae)

      mlflow.sklearn.log_model(lr, "model")

注意：
1.mlflow.set_tracking_uri("http://localhost:5002") 设置为刚才启动的mlflow tracking server的地址
2.mlflow.set_experiment('http_metrics_test') 设置实验的名字
3.安装该程序所依赖的python包
4.如果不是在同一个conda环境中，还得执行

    export AWS_ACCESS_KEY_ID=AKIAIOSFODNN7EXAMPLE      
    export AWS_SECRET_ACCESS_KEY=wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY      
    export MLFLOW_S3_ENDPOINT_URL=http://localhost:9001

便于python客户端上传模型文件以及模型元数据
直接执行 python wine.py 如果成功，访问mlflow tracking server ui下有如下

mlflow-ui.png

点击 2020-10-30 10:34:38，如下：

mlflow-model1.png mlflow-model2.png

启动mlflow 算法服务

在同一个conda环境中执行命令

export MLFLOW_TRACKING_URI=http://localhost:5002 
mlflow models serve -m runs:/e69aed0b22fb45debd115dfc09dbc75a/model -p 1234 --no-conda

其中e69aed0b22fb45debd115dfc09dbc75a为mlflow tracking server ui中的run id

如遇到ModuleNotFoundError: No module named 'sklearn'
执行 pip install scikit-learn==0.19.1
遇到ModuleNotFoundError: No module named 'scipy'
执行pip install scipy

请求访问该model启动的服务:

curl -X POST -H "Content-Type:application/json; format=pandas-split" --data '{"columns":["alcohol", "chlorides", "citric acid", "density", "fixed acidity", "free sulfur dioxide", "pH", "residual sugar", "sulphates", "total sulfur dioxide", "volatile acidity"],"data":[[12.8, 0.029, 0.48, 0.98, 6.2, 29, 3.33, 1.2, 0.39, 75, 0.66]]}' http://127.0.0.1:1234/invocations

输出 [5.455573233630147] 则表明该模型服务成功部署

至此主要简单的mlflow使用就完成了，如果还有mlflow不支持的算法，可以参照自定义model