Building a Pokemon Image Generator with MLOps best practices!
The training is done with PyTorch and HuggingFace on Lightning.ai Studios. The data for it is stored on a GCP Cloud Bucket and downloaded to the GPU device for training.
The frontend is a streamlit UI that renders the generated images.
It is continously deployed with GCP Cloud Build and is live on a GCP Cloud Run instance.
The image generation service uses FastAPI to serve the latest trained model and runs on a GCP Cloud Run instance.
Autmated Training is described best in this video: https://vimeo.com/948396185
The Frontend and Backend are both built and deployed automatically when a Git Tag is added to a certain commit with either a frontend/VERSION or backend/VERSION.