ai.h2o.automl.modeling.GBMStepsProvider Maven / Gradle / Ivy
The newest version!
package ai.h2o.automl.modeling;
import ai.h2o.automl.*;
import ai.h2o.automl.ModelSelectionStrategies.KeepBestN;
import ai.h2o.automl.events.EventLogEntry;
import hex.Model;
import hex.tree.SharedTreeModel;
import hex.tree.gbm.GBMModel;
import hex.tree.gbm.GBMModel.GBMParameters;
import water.Job;
import water.Key;
import java.util.*;
public class GBMStepsProvider
implements ModelingStepsProvider
, ModelParametersProvider {
public static class GBMSteps extends ModelingSteps {
static final String NAME = Algo.GBM.name();
static GBMParameters prepareModelParameters() {
GBMParameters params = new GBMParameters();
params._score_tree_interval = 5;
params._histogram_type = SharedTreeModel.SharedTreeParameters.HistogramType.AUTO;
return params;
}
static abstract class GBMModelStep extends ModelingStep.ModelStep {
GBMModelStep(String id, AutoML autoML) {
super(NAME, Algo.GBM, id, autoML);
}
public GBMParameters prepareModelParameters() {
GBMParameters params = GBMSteps.prepareModelParameters();
params._ntrees = 10000;
params._sample_rate = 0.8;
params._col_sample_rate = 0.8;
params._col_sample_rate_per_tree = 0.8;
return params;
}
}
static abstract class GBMGridStep extends ModelingStep.GridStep {
public GBMGridStep(String id, AutoML autoML) {
super(NAME, Algo.GBM, id, autoML);
}
public GBMParameters prepareModelParameters() {
GBMParameters params = GBMSteps.prepareModelParameters();
params._ntrees = 10000;
return params;
}
}
static abstract class GBMExploitationStep extends ModelingStep.SelectionStep {
protected GBMModel getBestGBM() {
for (Model model : getTrainedModels()) {
if (model instanceof GBMModel) {
return (GBMModel) model;
}
}
return null;
}
@Override
public boolean canRun() {
return super.canRun() && getBestGBM() != null;
}
public GBMExploitationStep(String id, AutoML autoML) {
super(NAME, Algo.GBM, id, autoML);
if (autoML.getBuildSpec().build_models.exploitation_ratio > 0)
_ignoredConstraints = new AutoML.Constraint[] { AutoML.Constraint.MODEL_COUNT };
}
}
private final ModelingStep[] defaults = new GBMModelStep[] {
new GBMModelStep("def_1", aml()) {
@Override
public GBMParameters prepareModelParameters() {
GBMParameters params = super.prepareModelParameters();
params._max_depth = 6;
params._min_rows = 1;
return params;
}
},
new GBMModelStep("def_2", aml()) {
@Override
public GBMParameters prepareModelParameters() {
GBMParameters params = super.prepareModelParameters();
params._max_depth = 7;
params._min_rows = 10;
return params;
}
},
new GBMModelStep("def_3", aml()) {
@Override
public GBMParameters prepareModelParameters() {
GBMParameters params = super.prepareModelParameters();
params._max_depth = 8;
params._min_rows = 10;
return params;
}
},
new GBMModelStep("def_4", aml()) {
@Override
public GBMParameters prepareModelParameters() {
GBMParameters params = super.prepareModelParameters();
params._max_depth = 10;
params._min_rows = 10;
return params;
}
},
new GBMModelStep("def_5", aml()) {
@Override
public GBMParameters prepareModelParameters() {
GBMParameters params = super.prepareModelParameters();
params._max_depth = 15;
params._min_rows = 100;
return params;
}
},
};
static class DefaultGBMGridStep extends GBMGridStep {
public DefaultGBMGridStep(String id, AutoML autoML) {
super(id, autoML);
}
@Override
public Map prepareSearchParameters() {
Map searchParams = new HashMap<>();
searchParams.put("_max_depth", new Integer[]{3, 4, 5, 6, 7, 8, 9, 10, 11, 12, 13, 14, 15, 16, 17});
searchParams.put("_min_rows", new Integer[]{1, 5, 10, 15, 30, 100});
// searchParams.put("_learn_rate", new Double[]{0.001, 0.005, 0.008, 0.01, 0.05, 0.08, 0.1, 0.5, 0.8});
searchParams.put("_sample_rate", new Double[]{0.50, 0.60, 0.70, 0.80, 0.90, 1.00});
searchParams.put("_col_sample_rate", new Double[]{ 0.4, 0.7, 1.0});
searchParams.put("_col_sample_rate_per_tree", new Double[]{ 0.4, 0.7, 1.0});
searchParams.put("_min_split_improvement", new Double[]{1e-4, 1e-5});
return searchParams;
}
}
private final ModelingStep[] grids = new GBMGridStep[] {
new DefaultGBMGridStep("grid_1", aml()),
/*
new DefaultGBMGridStep("grid_1_resume", aml()) {
@Override
protected void setSearchCriteria(RandomDiscreteValueSearchCriteria searchCriteria, Model.Parameters baseParms) {
super.setSearchCriteria(searchCriteria, baseParms);
searchCriteria.set_stopping_rounds(0);
}
@Override
@SuppressWarnings("unchecked")
protected Job startJob() {
Key[] resumedGrid = aml().getResumableKeys(_provider, "grid_1");
if (resumedGrid.length == 0) return null;
return hyperparameterSearch(resumedGrid[0], prepareModelParameters(), prepareSearchParameters());
}
}
*/
};
private final ModelingStep[] exploitation = new ModelingStep[] {
new GBMExploitationStep("lr_annealing", aml()) {
Key resultKey = null;
@Override
protected Job startTraining(Key result, double maxRuntimeSecs) {
resultKey = result;
GBMModel bestGBM = getBestGBM();
aml().eventLog().info(EventLogEntry.Stage.ModelSelection, "Retraining best GBM with learning rate annealing: "+bestGBM._key);
GBMParameters params = (GBMParameters) bestGBM._input_parms.clone();
params._max_runtime_secs = 0; // reset max runtime
params._learn_rate_annealing = 0.99;
initTimeConstraints(params, maxRuntimeSecs);
setStoppingCriteria(params, new GBMParameters());
return asModelsJob(startModel(Key.make(result+"_model"), params), result);
}
@Override
protected ModelSelectionStrategy getSelectionStrategy() {
return (originalModels, newModels) ->
new KeepBestN<>(1, () -> makeTmpLeaderboard(Objects.toString(resultKey, _provider+"_"+_id)))
.select(new Key[] { getBestGBM()._key }, newModels);
}
}
};
public GBMSteps(AutoML autoML) {
super(autoML);
}
@Override
public String getProvider() {
return NAME;
}
@Override
protected ModelingStep[] getDefaultModels() {
return defaults;
}
@Override
protected ModelingStep[] getGrids() {
return grids;
}
@Override
protected ModelingStep[] getOptionals() {
return exploitation;
}
}
@Override
public String getName() {
return GBMSteps.NAME;
}
@Override
public GBMSteps newInstance(AutoML aml) {
return new GBMSteps(aml);
}
@Override
public GBMParameters newDefaultParameters() {
return new GBMParameters();
}
}