Open erikerlandson opened 3 years ago
basic concept would be the seldon custom model mode: https://github.com/SeldonIO/seldon-core#deploy-your-custom-model-using-language-wrappers
In this scenario something like a pdf object would be pickled and passed over using jsonData
format or binData
format and the predict
function unpacks it and runs the detection model, and returns the resulting output as return json instead of outputting a csv file.
see also this batched-pipeline example workflow: https://medium.com/analytics-vidhya/automated-data-pipeline-using-ceph-notifications-and-kserving-5e1e9b996661
workloads like the corporate data stream ("nlp/pdf") can probably be modified so that they are deployable via Seldon s2i and model management. To explore this it will help to have ODH seldon deployed on the data commons cluster.
cc @MichaelClifford @ChristianMeyndt @HumairAK @redmikhail