工业界的算法问题本质上都是工程问题。深度学习算法也当然不能例外。日常除了训练好一个模型之外，还需要将模型部署到生产环境中去，这就涉及到大量工程技术和一定的开发能力。

最近，UC Berkeley 2021全栈深度学习课程开源了其关于深度学习模型部署与监控（Deployment & Monitoring）的课程，课程分为两大部分，模型部署与模型监控。

模型部署部分分别从深度学习模型部署类型、REST API部署、边缘设备部署等三个方面来阐述。模型监控主要从数据流、监控方法、监控工具等方面来阐述。

完整课程目录如下：

I - Model Deployment

1 - Types of Deployment

Batch Prediction

Model-In-Service

Model-As-Service

2 - Building A Model Service

REST APIs

Dependency Management

ONNX

Docker

Performance Optimization

GPU or no GPU?

Concurrency

Model distillation

Model quantization

Caching

Batching

Sharing The GPU

Model Serving Libraries

Horizontal Scaling Container Orchestration

Deploying Code As Serverless Functions

Model Deployment

Managed Options

Take aways

3 - Edge Deployment

Tools For Edge Deployment

More Efficient Models

Mindset For Edge Deployment

Takeaways

II - Model Monitoring

1 - Why Model Degrades Post-Deployment?

2 - Data Drift

3 - What Should You Monitor?

4 - How Do You Measure Distribution Changes?

Select A Reference Window

Select A Measurement Window

Compare Windows Using A Distance Metric

5 - How Do You Tell If A Change Is Bad?

6 - Tools For Monitoring

7 - Evaluation Store

Conclusion

部分PPT内容如下：

对深度学习模型部署感兴趣或者是有需要用到的读者可以扫描下方二维码加笔者微信获取。

往期精彩：

时隔一年！深度学习语义分割理论与代码实践指南.pdf第二版来了！
 新书预告 | 《机器学习公式推导与代码实现》出版在即！