ML flow config

config.yml

default:
  # Common ----
  experiment_dir: '.' # relative from project directory as root
  data_dir: '../data' # relative from @experiment_dir as root
  cache_dir: 'cache' # relative from @experiment_dir as root
  output_dir: 'output' # relative from @experiment_dir as root
  experiment_version: '2020-01-01'
  
  stages:
    # Downloading data stage config ----
    downloading:
      data_version: '2020-01-01'
      metadata:
        target_columns: []
        hashing_columns: []
    
    # Preprocessing data stage config ----
    preprocessing:
      data_version: '2020-01-01'
      metadata:
        unique_id: 'id'
        binary: []
        datetime: []
        timestamp: []
        extra: []
    
    # Feature engineering stage config ----
    feature_engineering:
      data_version: '2020-01-01'
      metadata:
        unique_id: 'id'
        factors: []
        extra: []
    
    # Train model stage config ----
    train_model:
      data_version: '2020-02-02'
      metadata:
        unique_id: 'id'
        extra: ['f2']
        label: 'target'
        weight: 'value'
        timestamp: 'time'
      
      factors_encoding:
        type: 'TargetEncoding'
        min_leveles_n: 3
        folds_n: 8
        fold_column: 'te_fold_n'
      
      early_stopping:
        rounds_n: 3
        metric: 'logloss'
      
      folding:
        folds_n: 8
        fold_column: 'fold_n' # must be null when folds_n has value
        fold_assignment: NULL # must be null when fold_column has value
      
      auto_ml:
        include_algos: ['GBM', 'DRF'] # see available algos https://docs.h2o.ai/h2o/latest-stable/h2o-docs/data-science/algo-params/include_algos.html
        max_duration_secs: 3600 # limit the runtime in secs
        max_models_n: 1000 # max trained models number
        rounds_n: 5 # search will stop after not improving much over the best N random models
        
      cross_validation:
        top_n_models: 5
        
      final_model:
        model_version: 'auto'
      
      seed: 314
      
    # New data predictions stage config ----
    predict_new_data:
      data_version: '2020-02-02'
      model_version: '1601976544.59666'
      metadata:
        unique_id: 'id'
        weight: 'W'
        label: 'label'
      threshold_searching_strategy:
        name: 'quantile'
        value: 0.8
  
  # Misc ----
  verbose: 1