Deployment Framework - Static Machine Learning Model

- Separate Training from the Server:

• Static model is trained offline. That is, we train the model exactly once and then use that trained model for a while.

• The model training is done on the local machine and once the model is complete, it is saved and transferred to server to make live predictions.

• Packaging ML Model: the machine learning model is packaged as file in the following format (python):

  • Pickle file: converts python object to byte stream.

  • Predictive model markup language: xml based file format for saving prediction model.

  • Joblib File: A faster method for saving the model trained on a large dataset.

  • Gzip File: compress model to save storage space

• The packaged ML model can be easily loaded by the server.

- ML Model as Standalone Model:

• Every pre-processing step and feature engineering is replicated on the server-side for transforming raw input.

• The server loads the ML model only.

• The transformed input is put in the model and the model outputs prediction in the required format.

- Advantage of static model:

• Static Model is cheaper to create and maintain than Dynamic Model which uses online training methodology.

- Drawback of static model:

• Many information sources through which data is extracted and trained on, change over time, therefore, the input data must be monitored.

References

• https://medium.com/analytics-and-data/overview-of-the-different-approaches-to-putting-machinelearning-ml-models-in-production-c699b34abf86

• https://developers.google.com/machine-learning/crash-course/static-vs-dynamic-training/video-lecture

• https://www.kdnuggets.com/2017/09/putting-machine-learning-production.html

• https://cloudxlab.com/blog/deploying-machine-learning-model-in-production/

• https://christophergs.github.io/machine%20learning/2019/03/17/how-to-deploy-machine-learning-models/#tools