Steps to deploy trained machine leaning model into production

1
votes

I am new to ML world, when reading about building model with training data and finally test the data to suit requirements, till this point I am able to follow, my questions is once tested model is ready to deply

  1. Do I need to train / re-train the model after production deployment? if so what is the practice?
  2. Is there way to persists the hypothesis, so that model can predict using result which are persisted?
  3. Is it good practice to re-train the model every day or week or month?

Assuming spark MLib is used to build the model

  • let me add more details. When I train the model, argument sake it generates score based train data in pre-production environment, can I save data with score in distributed storage for later predictions instead of re-train the model. I want to understand steps 1) When I take my well tested model from pre-production to production environment, do I need to have train data and new data set on which I would like to predict the results?

I am very new to machine learning, would like to understand how model code released to next environment, what steps I need to consider

thanks for clarification, which means I need to deploy model object into production then use production data (features) to calculate target data value?, can you please refer me to any book or information where I can get inside into how to build, validate and deploy supervised algorithms.

again, thanks for your time to educate me.

1
apache-sparkmachine-learning
The 'ML' tag is for the programming language ML, not machine learning. – molbdnilo

1 Answers

1
votes

Do I need to train / re-train the model after production deployment? if so what is the practice?

Not necessarily but you probably will. A lot depends on what's being modeled and how stable it is over time.

Is there way to persists the hypothesis, so that model can predict using result which are persisted?

  • ?? Not sure I understand the question but most production models have logging systems attached to analysis / reporting / vizualization software to help keep track of model performance and help decide when to retrain.

Is it good practice to re-train the model every day or week or month?

  • Kind of depends on performance and resource constraints. If you have a small number of models to retrain / score with and retraining won't affect SLA's, it's probably not a bad idea but when there are limited computing resources the answer might change.