Kubeflow pipelines for multifamily price prediction

Problem

A commercial mortgage underwriter is interested in the future prices and valuations of various commercial real estate assets. In this example, we build a machine learning pipeline that utilizes various attributes of historical real estate valuations in order to predict the future valuation of a given asset. This pipline is deployed via Kubeflow to Vertex AI in order to automate and scale, efficiently leveraging GCP's serverless infrastucture.

Code Sample

The python code for this pipeline resides in https://github.com/secrettoad/kubeflow/blob/main/multifamily_pricing/main.py

To run the pipeline, call the run_pricing_pipeline function within main.py

To deploy a pipeline for automatic runs on new uploads to the associated gcs bucket, run the command below (you will need to authenticate to gcp first and have the appropriate permissions, as well as have created the bucket to be listened to)

gcloud functions deploy python-finalize-function \
--runtime=python38 \
--region=us-west1 \
--source=multifamily_pricing \
--entry-point=run_pricing_pipeline \
--trigger-bucket="coysu-demo-datasets" \
--memory=1024MB

The Google Cloud Functions cli command provided at the beginning of this document creates a serverless, asynchronous function that listens for new files being added to the relevant bucket. The pipeline then uses dask to load all relevant files within each train and test directory.

Within the function deploy_component a model is pushed to the gcp model registry and then exposed via an http endpoint.

Kubeflow in conjunction with Vertex AI is a combination of tools that enables the deployment of pipelines in a commercial-grade, production-ready manner without demanding an inordinate amount of time or resources. Contact John at john@coysu.com for more information.