[optuna] some more tunings due to lgbm optuna tuning

configs/ref_lgbm.yaml

@@ -37,11 +37,12 @@ learning:
     n_estimators: 10
     random_state: 1
     max_depth: 20
+    subsample: 0.9
   optimisation:
     optuna_params:
-      n_trials: 100
+      n_trials: 1000
       step: 2
-      cv: 2
+      cv: 5
       verbose: 1000
       random_state: 1
       scoring: r2  # r2, neg_mean_absolute_error, etc..see note above
@@ -53,12 +54,17 @@ learning:
       learning_rate: FloatDistribution(0.0001, 10, log=True)
       num_leaves: IntDistribution(2, 256)
       subsample_for_bin: IntDistribution(10, 1000000, step=10)
-      colsample_bytree: FloatDistribution(0.01, 1.0, log=True)
+      subsample: FloatDistribution(0.4, 1.0)  # this is the same as bagging_fraction
+      subsample_freq: IntDistribution(1, 100)  # this is the same as bagging_freq
+      colsample_bytree: FloatDistribution(0.4, 1.0)
+      colsample_bynode: FloatDistribution(0.01, 1.0)
       reg_alpha: FloatDistribution(1e-8, 10, log=True)
       reg_lambda: FloatDistribution(1e-8, 10, log=True)
       min_child_samples: IntDistribution(2, 100)
       min_child_weight: FloatDistribution(0, 10)
       min_split_gain: FloatDistribution(1e-8, 1, log=True)
+      boosting_type: CategoricalDistribution(['gbdt', 'dart', 'rf'])
+
 #    hyperopt_params:
 #      max_evals: 5
 #      step: 2

setup.py

@@ -109,6 +109,7 @@ def run(self):
         'Pillow >= 8.1.2',
         "PyWavelets==1.2.0",
         "imageio==2.9.0",
+        "optuna==3.2.0",
     ],
     extras_require={
         'kmz': [

uncoverml/optimise/models.py

@@ -596,6 +596,7 @@ def __init__(
             n_jobs=n_jobs,
             silent=silent,
             importance_type=importance_type,
+            **kwargs
         )
 
     def predict(self, X, *args, **kwargs):

uncoverml/optuna_opt.py

@@ -72,7 +72,7 @@ def optimise_model(X, targets_all: Targets, conf: Config):
         param_distributions=search_space,
         scoring=scorer,
         cv=cv,
-        error_score=np.nan,
+        error_score='raise',
         random_state=rstate,
         return_train_score=True,
         ** conf.optuna_params