# W3 Automated Machine Learning

# Why use AutoML

• Ability to reduce time-to-market of the product as a result of lesser iterations of creating the model.
• Lack of particular ML skillsets in teams is not a concern
• Ability to iterate & experiment quickly
• Ability to optimize scarce resources and skillsets
• Lets experts focus on harder tasks which involves domain knowledge

# AutoML workflow

• AutoML aims at automating the process of building models
• Steps of workflow
  • You provide a labelled dataset from which it detects the type of problem to solve - regression, classification, etc
  • It then selects an algorithm
  • It applies transformations and preprocessing
  • Selects various hyperparameters and configs to train & test the models

# SageMaker Autopilot

• Fully transparent and shares code and notebooks for all the processing which are reproducible
• Steps
  • Upload dataset to S3
  • Provide Autopilot with the target variable
  • It goes through the entire AutoML workflow
  • It returns 2 notebooks - the data exploration (what it learned and potential issues with data) and candidate generation notebook (each preprocessing step, algorithm and hyperparameter choices)
• SDK's available
  • AWS CLI
  • AWS SDK
  • Amazon SageMaker
  • SageMaker Studio

# TFIDF vectorizer for text

• tf(t, d) = \frac{f_{t, d}}{\sum_{t' \in d}f_{t', d}}
• idf(t, D) = log(\frac{|D|}{|{d \in D : t \in d}|}) (scales terms based on frequency)
• tf-idf(t, d, D) = tf(t, d) * idf(t, D)
• Where t = term, d = document, D = corpus

# Autopilot results

• Data transformation and job config code
• Data exploration and candidate notebooks
• Transformed data (train, val data)
• Models
• Metrics report

# Model hosting

• Involves a stack containing a proxy, web server, serving code and model
• Using autopilot, choose the instance counts, docker container for inference and it takes care of creating endpoints.
• PipelineModel contains various containers
  • Data Transformation - built from model that was trained to transform data
  • Algorithm - built from trained model selecting best algorithm to predict
  • Inverse label Transformer - converts numerical intermediate prediction to labels
• All of above hosted on same endpoint using inference model