Details

    • Type: New Feature
    • Status: Open
    • Priority: Major
    • Resolution: Unresolved
    • Affects Version/s: None
    • Fix Version/s: 3.22.0.1
    • Component/s: Grid, StackedEnsemble
    • Labels:
      None

      Description

      Here is the initial design idea for the Stacked Ensemble grid search (which is mostly a search of the metalearner hyperparameters, but could also include other params like metalearner_nfolds).

      Python example:

      metalearner_grid_params_gbm = {'max_depth': [2,3,4], 'col_sample_rate': [0.2,0.5,0.7]}
      metalearner_grid_params_rf = {'ntrees': [200,300,400], 'col_sample_rate': [0.2,0.5,0.7]}
      
      
      # set up SE grid, use hyper_params to pass a new value called metalearner_params
      grid = H2OGridSearch(model=H2OStackedEnsembleEstimator,
                           hyper_params={'metalearner_grid_params': [{'algorithm': "GBM", 'params': metalearner_grid_params_gbm}, 
                                                               {'algorithm': "DRF", 'params': metalearner_grid_params_rf}]},
                           seed=1,
                           search_criteria={'strategy': 'RandomDiscrete', 'max_models': 36})
      
      grid.train(x=x, y=y, 
      	       training_frame=train, 
      	       seed=1,  #this is SE seed (not grid seed)
      	       base_models=[my_gbm, my_rf])  #pass along fixed SE params like base_models
      
      
      
      
      # Single model (for comparison)
      metalearner_gbm_params = {'max_depth': 2, 'col_sample_rate': 0.3}
      ensemble = H2OStackedEnsembleEstimator(base_models=[my_gbm, my_rf],
                                             metalearner_algorithm="GBM",
                                             metalearner_params=metalearner_gbm_params)
      ensemble.train(x=x, y=y, training_frame=train)
      

        Attachments

          Issue links

            Activity

              People

              • Assignee:
                michalk Michal Kurka
                Reporter:
                erin Erin LeDell
              • Votes:
                0 Vote for this issue
                Watchers:
                2 Start watching this issue

                Dates

                • Created:
                  Updated: