architecture.md

Architecture

Code structure

• All code goes in the /src/ directory
• Bundled package data
  • Binary models (pre-trained)
    • src/models/<model>/model.bin
  • Encoders for processing input
    • src/models/<model>/encoders/
  • Metadata about model metrics
    • src/models/<model>/metadata/results.csv
• Models
  • Each model is a python module under webcompat_ml.models
  • Each model has a CLI tool for training and classification
  • Each CLI is installed as a console script so the rest of the pipeline can use it as a standalone program.
• Utils
  • We keep track of some helpers functions used by the rest of the pipeline in the webcompat_ml.utils module.

Available models

Invalid

• Classifies issues as valid or invalid.
• Based on xgboost

Pipeline

Data processing

• Given a pandas DataFrame
  • Build and store the encoders for fields when loading data required for the model
  • Load the encoders defined in the building process

Training

• Given a CSV as input
  • Convert input to a pandas DataFrame
  • Load the encoders defined in the building phase
  • Process data using the encoders
  • Split the input to training/validation sets
  • Train the model with training data set
  • Validate the model against validation data set
  • Store model metrics

Classification

• Given a CSV as input and a trained model
  • Convert input to a pandas DataFrame
  • Load the encoders defined in the building phase
  • Process data using the encoders
  • Generate predictions