# Databricks开源项目MLflow入门学习

By [liyao](https://paragraph.com/@liyao) · 2022-02-11

---

> An open source platform for the machine learning lifecycle [https://mlflow.org/](https://mlflow.org/)

### 上下文

*   为开发者提供类似谷歌TFX、Facebook FBLearner Flow等平台类似好处
    
*   可以支持任何工具和算法
    

### 项目架构

![](https://storage.googleapis.com/papyrus_images/809a62e5eb1ecb2a0a495573fa6df7deea51974a2afd607fb8d40171fe718267.png)

#### MLflow Tracking

*   记录和查询实验：代码、数据、配置和结果
    
*   [https://www.mlflow.org/docs/latest/tracking.html](https://www.mlflow.org/docs/latest/tracking.html)
    

#### MLflow Projects

*   可在任何平台上重复运行的打包格式
    
*   [https://www.mlflow.org/docs/latest/projects.html](https://www.mlflow.org/docs/latest/projects.html)
    

#### MLflow Models

*   将模型发送到各种部署工具的通用格式
    
*   [https://www.mlflow.org/docs/latest/models.html](https://www.mlflow.org/docs/latest/models.html)
    

#### Model Registry

*   中央存储库：存储、注释、发现和管理模型
    
*   [https://mlflow.org/docs/latest/model-registry.html](https://mlflow.org/docs/latest/model-registry.html)
    

### 项目使用

*   [Anaconda](https://docs.conda.io/projects/conda/en/latest/user-guide/install/)环境安装，可以便捷获取包且对包能够进行管理，同时对环境可以统一管理的发行版本
    

#### mlflow安装

#### UI启动

*   examples/mlflow\_tracking.py
    

    cd examples
    mlflow ui
    

#### 模型运行

    # 模型生成
    python3 sklearn_logistic_regression/train.py
    # API启动
    mlflow models serve -m runs:/<Model_id>/model --port 1234
    可能报错：mlflow.utils.process.ShellCommandException: Non-zero exitcode: 1
    mlflow models serve -m runs:/<Model_id>/model --port 1234 --no-conda
    

### 流处理Demo

*   [https://github.com/mlflow/mlflow/tree/master/examples/multistep\_workflow](https://github.com/mlflow/mlflow/tree/master/examples/multistep_workflow)
    

### 思考

*   官方待引入（MLflow Monitoring）
    
*   MLFlow现阶段还是类似Pipeline的辅助工具，定义了Pipeline工具和标准，缺少AI平台概念
    
*   如何与Spark、Docker做集成
    

* * *

That's all!

---

*Originally published on [liyao](https://paragraph.com/@liyao/databricks-mlflow)*