Deploying and visualizing models

Goals

• Easily re-train and re-deploy models
• Analyze models performance both over historical test set and new test data.

Deploying models

• Dockerize model (this should be handled automatically in flow). TimeModel class should include an infer function. Dockerfile should contain necessary packages.
• Model weights are automatically uploaded to GCS into a bucket called ts-model-prod
• Dockerfile should take as input path to the weight file. So that all that requires changing to updated model.
• Models predictions are continuously saved to a GCS bucket ts-predictions-prod/{model-id}

(Re)-Deployment Architecture

• New data gathers over a 50 day period.
• This data is split into two partitions. 18 of these days will be added to the test set and the remaining 32 days will be incorporated as new training data.
• Model preforms better on the previous historical dataset
• Both old model has test metric computed and new model is run on the newly acquired test set.