FIXES #558 - use less memory for ensemble builder, allow setting memory limit

Author: mfeurer

autosklearn/automl.py

@@ -66,6 +66,7 @@ def __init__(self,
                  initial_configurations_via_metalearning=25,
                  ensemble_size=1,
                  ensemble_nbest=1,
+                 ensemble_memory_limit=1000,
                  seed=1,
                  ml_memory_limit=3072,
                  metadata_directory=None,
@@ -94,6 +95,7 @@ def __init__(self,
             initial_configurations_via_metalearning
         self._ensemble_size = ensemble_size
         self._ensemble_nbest = ensemble_nbest
+        self._ensemble_memory_limit = ensemble_memory_limit
         self._seed = seed
         self._ml_memory_limit = ml_memory_limit
         self._data_memory_limit = None
@@ -635,18 +637,21 @@ def _get_ensemble_process(self, time_left_for_ensembles,
         else:
             self._ensemble_size = ensemble_size
 
-        return EnsembleBuilder(backend=self._backend,
-                               dataset_name=dataset_name,
-                               task_type=task,
-                               metric=metric,
-                               limit=time_left_for_ensembles,
-                               ensemble_size=ensemble_size,
-                               ensemble_nbest=ensemble_nbest,
-                               seed=self._seed,
-                               shared_mode=self._shared_mode,
-                               precision=precision,
-                               max_iterations=max_iterations,
-                               read_at_most=np.inf)
+        return EnsembleBuilder(
+            backend=self._backend,
+            dataset_name=dataset_name,
+            task_type=task,
+            metric=metric,
+            limit=time_left_for_ensembles,
+            ensemble_size=ensemble_size,
+            ensemble_nbest=ensemble_nbest,
+            seed=self._seed,
+            shared_mode=self._shared_mode,
+            precision=precision,
+            max_iterations=max_iterations,
+            read_at_most=np.inf,
+            memory_limit=self._ensemble_memory_limit,
+        )
 
     def _load_models(self):
         if self._shared_mode:

autosklearn/ensembles/ensemble_selection.py

@@ -74,7 +74,12 @@ def _fast(self, predictions, labels):
             if s == 0:
                 weighted_ensemble_prediction = np.zeros(predictions[0].shape)
             else:
-                ensemble_prediction = np.mean(np.array(ensemble), axis=0)
+                # Memory-efficient averaging!
+                ensemble_prediction = np.zeros(ensemble[0].shape)
+                for pred in ensemble:
+                    ensemble_prediction += pred
+                ensemble_prediction /= s
+
                 weighted_ensemble_prediction = (s / float(s + 1)) * \
                                                ensemble_prediction
             fant_ensemble_prediction = np.zeros(weighted_ensemble_prediction.shape)

autosklearn/estimators.py

@@ -14,6 +14,7 @@ def __init__(self,
                  initial_configurations_via_metalearning=25,
                  ensemble_size=50,
                  ensemble_nbest=50,
+                 ensemble_memory_limit=1024,
                  seed=1,
                  ml_memory_limit=3072,
                  include_estimators=None,
@@ -63,6 +64,11 @@ def __init__(self,
             ensemble. Implements `Model Library Pruning` from `Getting the
             most out of ensemble selection`.
 
+        ensemble_memory_limit : int, optional (1024)
+            Memory limit in MB for the ensemble building process.
+            `auto-sklearn` will reduce the number of considered models
+            (``ensemble_nbest``) if the memory limit is reached.
+
         seed : int, optional (default=1)
             Used to seed SMAC. Will determine the output file names.
 
@@ -157,16 +163,16 @@ def __init__(self,
               optimization/validation set, which would later on be used to build
               an ensemble.
             * ``'model'`` : do not save any model files
-              
+
         smac_scenario_args : dict, optional (None)
             Additional arguments inserted into the scenario of SMAC. See the
             `SMAC documentation <https://automl.github.io/SMAC3/stable/options.html?highlight=scenario#scenario>`_
             for a list of available arguments.
-            
+
         get_smac_object_callback : callable
             Callback function to create an object of class
             `smac.optimizer.smbo.SMBO <https://automl.github.io/SMAC3/stable/apidoc/smac.optimizer.smbo.html>`_.
-            The function must accept the arguments ``scenario_dict``, 
+            The function must accept the arguments ``scenario_dict``,
             ``instances``, ``num_params``, ``runhistory``, ``seed`` and ``ta``.
             This is an advanced feature. Use only if you are familiar with
             `SMAC <https://automl.github.io/SMAC3/stable/index.html>`_.
@@ -191,6 +197,7 @@ def __init__(self,
         self.initial_configurations_via_metalearning = initial_configurations_via_metalearning
         self.ensemble_size = ensemble_size
         self.ensemble_nbest = ensemble_nbest
+        self.ensemble_memory_limit = ensemble_memory_limit
         self.seed = seed
         self.ml_memory_limit = ml_memory_limit
         self.include_estimators = include_estimators
@@ -236,6 +243,7 @@ def build_automl(self):
             self.initial_configurations_via_metalearning,
             ensemble_size=self.ensemble_size,
             ensemble_nbest=self.ensemble_nbest,
+            ensemble_memory_limit=self.ensemble_memory_limit,
             seed=self.seed,
             ml_memory_limit=self.ml_memory_limit,
             include_estimators=self.include_estimators,