Model training as a CI/CD system

TensorFlow Show and Tell
, , , ,
1

Hi folks,Continuous integration and deployment (CI/CD) is a common topic of discussion when it comes to DevOps. No wonder why it has also become so for MLOps. With MLOps though, we have another piece of continuity - continuous re-training and evaluation.

In our latest two-part article from the GCP blog, @deep-diver and I dive deep into incorporating CI/CD for ML with TensorFlow, TFX, and Vertex AI along with other services from GCP. We take the scenario where we need to incorporate code changes (be it for better training techniques or better model architectures) for an ML system and perform CI/CD in a meaningful manner. Below are the links:

Happy to answer q’s.

7 Likes
3

Hey, this is what I was looking for!
Thanks for sharing!!

1 Like
5

Thanks for sharing, this is very helpful!

I have a question, in the partial-pipeline-deployment.yaml in the ‘Create Pipeline’ step, in the tfx cli, why use the create command and not the update? If the pipeline is already created, it will raise that the pipeline already exists and will not update, thus the run create will be the same, or no?

1 Like
7

@deep-diver to reply to this one.

1 Like
9

@Sayak_Paul

CI/CD tool(cloud build or github action here) always starts from clean state with fresh container. That means there is no existing pipeline, but you can do update if you are using a dedicated stateful server.

2 Likes
11

Understood! Thank you.