The Fabricius Classifier API is required by the Fabricius Worbench https://github.com/googleartsculture/workbench and uses Google Cloud Auto ML to suggest Gardiner Codes for glyphs in an image.
Following the update to Vertex AI the confidence needs to be above 50% to return a prediction. This has led to a few errors, when the model doesn't return anything.
The service needs updating so that:
Service doesn't 500 if there are no predictions and returns an sensible message to the client
Lower confidence score as there is a lot of divergence in potential quality of image that is being sent
@justingrayston -- one of the routes we're taking to explore is to see if we can improve the training corpus to be more specific to time periods and sources. Is there a way we can apply for CloudML credits for academic research in this regard? Also, are there deployment instructions for the classifier?
Currently the Cloud Build configs built the environment every run which is wasteful as the code underneath hasn't actually changed necessarily (for example promoting to higher environments).
Why build anyway?
The pip install from requirements could take up to 8 minutes to run as the grpc dependency takes a while to load (especially on a Cloud Build machine). It is much quicker to build the docker image than it to run it on the skinny build image as more dependencies are already present.
This is obviously subject to change and future testing is needed.