Managing Machine Learning Experiments with MLflow
💻 Abstract
Successfully building and deploying a machine learning model is difficult. Enabling other data scientists to reproduce your pipeline, compare the results of different versions, and rollback models are much harder. This talk will introduce MLflow, an open-source project that helps developers reproduce and share ML experiments, manage models, and control the challenges associated with making models "production-ready."
🔊🔊 Speaker Bio
Jules S. Damji is a Developer Advocate at Databricks and an MLflow contributor. He is a hands-on developer with over 20 years of experience and has worked as a software engineer at leading companies such as Sun Microsystems, Netscape, @Home, LoudCloud/Opsware, VeriSign, ProQuest, and Hortonworks, building large-scale distributed systems. He holds a BSc and MSc in computer science and an MA in political advocacy and communication from Oregon State University, Cal State, and Johns Hopkins University.
Brooke Wenig is a Machine Learning Practice Lead at Databricks.
If you enjoyed this talk, visit us at https://mlopsworld.com/ and come participate in our next gathering!
Would you like to receive email summaries of these talks? Join our newsletter FREE here: http://bit.ly/MLOps_Summaries
Timestamps:
0:00 Intro
0:10 Introduction of the speakers
1:02 Agenda
2:58 Introducing MLflow
4:24 Key concepts in MLflow tracking
4:54 Tracking experiments
5:23 MLflow tracking
7:31 MLflow models
8:16 Example of MLflow model
8:50 Model flavors example
9:48 The model management problem
10:12 MLflow model registry
12:14 Demo
❓ Q&A ❓
32:34 How does MLflow differ from Azure ML workspace/AWS Sage maker?
34:28 Can an existing Git repo be transformed into an ml flow project?
36:10 How do MLflow projects relate to onnx?
36:49 Do you have any recommendations for hosting the tracking server and the registry server on the same or different machines?
38:23 How can you embed images into the UI?
39:32 Does MLflow provide standardized rust or gRPC API for model inference, no matter the underlying framework, pytorch, TensorFlow, etc?
40:23 Does MLflow provide an easy way to dockerize a model?
40:44 Do ml flow tracking server support authentication and role permissions?
43:14 Are these functions are identical and accessed through Azure Data bricks?
45:01 Can you connect different workflows in something like a DAG?
45:12 Does the model registry allow for custom versioning scheme?
46:23 Do you have any recommendations on how ml flow can be integrated into existing CI/CD?
48:19 Can we also search for model versions by filters?
48:57 How does MLflow compare with Kubeflow?
49:47 how does MLflow deal when the data set changes?
51:36 Do you provide a data platform for students?
51:59 How to migrate from data lake and Hadoop to Delta lake?
55:31 How we're sharing the slides?
55:58 Closing remarks