add forbidden condition cyclic lr

Author: ravinkohli

autoPyTorch/pipeline/tabular_classification.py

@@ -289,6 +289,29 @@ def _get_hyperparameter_search_space(self,
                                 raise ValueError("Cannot find a legal default configuration")
                             cs.get_hyperparameter('network_embedding:__choice__').default_value = default
 
+        # Disable CyclicLR until todo is completed.
+        if 'lr_scheduler' in self.named_steps.keys() and 'trainer' in self.named_steps.keys():
+            trainers = cs.get_hyperparameter('trainer:__choice__').choices
+            for trainer in trainers:
+                available_schedulers = self.named_steps['lr_scheduler'].get_available_components(
+                    dataset_properties=dataset_properties,
+                    exclude=exclude if bool(exclude) else None,
+                    include=include if bool(include) else None)
+                # TODO: update cyclic lr to use n_restarts and adjust according to batch size
+                cyclic_lr_name = 'CyclicLR'
+                if cyclic_lr_name in available_schedulers:
+                    # disable snapshot ensembles and stochastic weight averaging
+                    cs.add_forbidden_clause(ForbiddenAndConjunction(
+                        ForbiddenEqualsClause(cs.get_hyperparameter(
+                            f'trainer:{trainer}:use_snapshot_ensemble'), True),
+                        ForbiddenEqualsClause(cs.get_hyperparameter('lr_scheduler:__choice__'), cyclic_lr_name)
+                    ))
+                    cs.add_forbidden_clause(ForbiddenAndConjunction(
+                        ForbiddenEqualsClause(cs.get_hyperparameter(
+                            f'trainer:{trainer}:use_stochastic_weight_averaging'), True),
+                        ForbiddenEqualsClause(cs.get_hyperparameter('lr_scheduler:__choice__'), cyclic_lr_name)
+                    ))
+
         self.configuration_space = cs
         self.dataset_properties = dataset_properties
         return cs

autoPyTorch/pipeline/tabular_regression.py

@@ -238,6 +238,29 @@ def _get_hyperparameter_search_space(self,
                                 raise ValueError("Cannot find a legal default configuration")
                             cs.get_hyperparameter('network_embedding:__choice__').default_value = default
 
+        # Disable CyclicLR until todo is completed.
+        if 'lr_scheduler' in self.named_steps.keys() and 'trainer' in self.named_steps.keys():
+            trainers = cs.get_hyperparameter('trainer:__choice__').choices
+            for trainer in trainers:
+                available_schedulers = self.named_steps['lr_scheduler'].get_available_components(
+                    dataset_properties=dataset_properties,
+                    exclude=exclude if bool(exclude) else None,
+                    include=include if bool(include) else None)
+                # TODO: update cyclic lr to use n_restarts and adjust according to batch size
+                cyclic_lr_name = 'CyclicLR'
+                if cyclic_lr_name in available_schedulers:
+                    # disable snapshot ensembles and stochastic weight averaging
+                    cs.add_forbidden_clause(ForbiddenAndConjunction(
+                        ForbiddenEqualsClause(cs.get_hyperparameter(
+                            f'trainer:{trainer}:use_snapshot_ensemble'), True),
+                        ForbiddenEqualsClause(cs.get_hyperparameter('lr_scheduler:__choice__'), cyclic_lr_name)
+                    ))
+                    cs.add_forbidden_clause(ForbiddenAndConjunction(
+                        ForbiddenEqualsClause(cs.get_hyperparameter(
+                            f'trainer:{trainer}:use_stochastic_weight_averaging'), True),
+                        ForbiddenEqualsClause(cs.get_hyperparameter('lr_scheduler:__choice__'), cyclic_lr_name)
+                    ))
+
         self.configuration_space = cs
         self.dataset_properties = dataset_properties
         return cs