[Week 1-2] ML Project Life Cycle

( reference : Introduction to Machine Learning in Production )

ML Project Life Cycle

[1] Welcome

After training ML model… now what?

\(\rightarrow\) useless, unless how to put them into production!

Goal

• not just build a ML model,
• but also put them into production

\(\rightarrow\) learn entire life cycle of ML project

Example )

• use CV to inspect phone defects

  ( put a bounding box around them )

• with data set of scratched phones, maybe able to train a model

But…. what to do to put this into production deployment?

Process

• prediction server is to accept API calls,

  ( prediction server = cloud, edge device … )

• receive an image,

• make a decision as to whether or

Then…what’s the problem??

• trained model might have worked well on your test set
• BUT…. real life production deployment might gives the model much darker images! ( = data drift / concept drift )

Much more problem in real world !!

\(\rightarrow\) will learn lot of important practical things for building ML systems that work not just in the lab, “but in a production deployment environment”

Another challenge : takes a lot more than ML code

• ML model codel = just only 5-10%

[2] Steps of an ML Project

[3] Case Study : speech recognition

step 1) scoping

• define task (project) = Speech Recongition
  • X = voice
  • y = text
• decide key metrics
  • ex) accuracy, latency, throughput
• estimate resources & timeline

step 2) data

• define data
  • Q) is the data labeled “consistently” ?
    • ex) with same voice, the label might be..
      • a) “Um, the weather is”
      • b) “Um… the weather is”
      • c) “The weather is”
  • Q) how much silence before/after the clip?
  • Q) how to perform “volume normalization”?

step 3) Modeling

ML model = (a) + (b) + (c)

• (a) code
• (b) hyperparameters
• (c) data

Research/Academia’s task

• “DATA” is fixed & change “code” & “hyperparameters”

Product team’ task

• “CODE” is fixed & change “data” & “hyperparameters”

step 4) Deployment

[4] Course Outline

review : **ML project life cycle. **

preview : will learn…

• starting from the end goal “deployment”, and then move toward “scoping”
• ( deployment \(\rightarrow\) modeling \(\rightarrow\) data \(\rightarrow\) scoping )