Hey everyone, I would like to share with you all a small project that I had worked at my spare time, I called it “BERT as a service”, the goal is to build an end-to-end TFX pipeline for sentiment analysis using BERT. This project is educational and is also aimed to provide a simple and easier reference for people that are looking to get familiar with MLOps using TFX and GCP (as was my case), but should not be that hard to tweak it to be more “production-ready”.
Here is an overview of what you will find:
An end-to-end ML pipeline, from data ingestion, data validation, transformation, training, and deployment.
Usage of specific components for infrastructure validation and hyperparameter tuning.
Orchestration using KubeFlow and the new Vertex AI
New version of TFX (1.0.0)
Colab version has a Tunner component that uses KerasTunner for HP search.
KubeFlow version has Infravalidator component to validate infra before blessing the model for deployment.
This was a very cool experience to get more familiar with MLOps concepts applied to the GCP ecosystem, if you have a similar project or any feedback I would love to know more.