@deep-diver and I have been constantly working on MLOps projects, and we want to share one of our latest works with you guys, “Dual Deployments on Vertex AI”. We leverage several components from the ML tooling provided by Google such as TensorFlow, TFX, Vertex AI, Cloud Build, and so on.
Dual Deployments is a common machine learning design pattern.
As described in the blog post, it can be applied to two scenarios of online/offline predictions and layered predictions.
In this blog post, we introduce two different approaches on how to write a machine learning pipeline to realize the dual deployment pattern.
TFX + custom model based approach
: DenseNet for cloud and MobileNet v3 for mobile deployments.
KFP + GCP’s AutoML based approach
: AutoML for cloud and AutoML Edge for mobile deployments.
In both cases, you can find out how custom components can be written including
- At the time of writing the blog post, TFX didn’t support uploading/hosting trained models on Vertex AI, so we wrote ones ourselves.
- For mobile deployment, we wrote a custom component to publish the TFLite model to Firebase ML.
If you find this brief description interesting, please find more information:
- GCP Blog Post: Dual deployments on Vertex AI | Google Cloud Blog
- GitHub Repo: https://github.com/sayakpaul/Dual-Deployments-on-Vertex-AI