Skip to main content
Version: 1.0

ML in GCP

It offers four options to build ML models

  • BigQuery ML: uses sql query to create ml models
  • Pre-Builds APIs: programming options
  • Auto ML: no code options to build models on Vertex AI
  • Custom training: custom environment

ml_options

  • [source]: google training session

How to choose ?​

ml_options

Pre-Builds APIs​

  • Neutral Language API

  • Extract entities

    • Detect sentiments
    • Analyze syntax
    • Classify content
    • Vision API
      • Lab & web detection
      • Logo detection
      • Landmark detection
      • Crop hints
      • Explicit content detection

  • Cloud Speech-to-Text API

    • Three types of reorganization
      • Synch (REST and gRPC): 1 minute or less
      • Asynch (REST and gRPC): up to 480 minutes
      • Streaming (gRPC): for real-time speech
    • Sample rate between 8000 Hz and 48000 Hz
    • Recommended 16000 Hz
    • No need to resample existing audio

AutoML​

  • Transfer learning
  • Neural Architecture Search: find the optimal model for the relevant project.

Custom training​

  • Vertex AI Workbench
    • Pre-built container: Tensorflow, Pytorch, Scikit-learn, ...
    • Custom container

Vertex AI​

  1. Data preparation
  • Upload data: text, tabular, image, video,
  • Feature engineering
  1. Train model
  • Model training
  • Model evaluation

model_types

  • [source]: google training session
  1. Model serving
  • Model deployment
  • Model monitoring

BigQuery ML​

  • Extract data from dataset
  • Create the model (specify the type, hyperparameters)
  • Evaluate model
  • Predict