Fix determinism for few preprocessing techniques (#1172)

* Fix determinism for few preprocessing techniques

* Code formatting

* Fix: model deserialization when mutual_info score is used for select_percentile/rates

Co-authored-by: Rohit Agarwal <rohit.agarwal4@aexp.com>

autosklearn/pipeline/components/data_preprocessing/rescaling/quantile_transformer.py

@@ -23,7 +23,8 @@ def __init__(self, n_quantiles: int, output_distribution: str,
         self.preprocessor = QuantileTransformer(
             n_quantiles=n_quantiles,
             output_distribution=output_distribution,
-            copy=False
+            copy=False,
+            random_state=random_state
         )
 
     @staticmethod

autosklearn/pipeline/components/feature_preprocessing/select_percentile_classification.py

@@ -1,6 +1,7 @@
 from ConfigSpace.configuration_space import ConfigurationSpace
 from ConfigSpace.hyperparameters import \
     UniformFloatHyperparameter, CategoricalHyperparameter, Constant
+from functools import partial
 
 from autosklearn.pipeline.components.base import AutoSklearnPreprocessingAlgorithm
 from autosklearn.pipeline.components.feature_preprocessing.select_percentile import \
@@ -27,7 +28,8 @@ def __init__(self, percentile, score_func="chi2", random_state=None):
         elif score_func == "f_classif":
             self.score_func = sklearn.feature_selection.f_classif
         elif score_func == "mutual_info":
-            self.score_func = sklearn.feature_selection.mutual_info_classif
+            self.score_func = partial(sklearn.feature_selection.mutual_info_classif,
+                                      random_state=self.random_state)
         else:
             raise ValueError("score_func must be in ('chi2, 'f_classif', 'mutual_info'), "
                              "but is: %s" % score_func)

autosklearn/pipeline/components/feature_preprocessing/select_percentile_regression.py

@@ -1,5 +1,6 @@
 from ConfigSpace.configuration_space import ConfigurationSpace
 from ConfigSpace.hyperparameters import UniformFloatHyperparameter, CategoricalHyperparameter
+from functools import partial
 
 from autosklearn.pipeline.components.base import AutoSklearnPreprocessingAlgorithm
 from autosklearn.pipeline.components.feature_preprocessing.select_percentile import \
@@ -24,7 +25,8 @@ def __init__(self, percentile, score_func="f_regression", random_state=None):
         if score_func == "f_regression":
             self.score_func = sklearn.feature_selection.f_regression
         elif score_func == "mutual_info":
-            self.score_func = sklearn.feature_selection.mutual_info_regression
+            self.score_func = partial(sklearn.feature_selection.mutual_info_regression,
+                                      random_state=self.random_state)
         else:
             raise ValueError("Don't know this scoring function: %s" % score_func)
 

autosklearn/pipeline/components/feature_preprocessing/select_rates_classification.py

@@ -2,6 +2,7 @@
 from ConfigSpace.hyperparameters import UniformFloatHyperparameter, \
     CategoricalHyperparameter
 from ConfigSpace import NotEqualsCondition
+from functools import partial
 
 from autosklearn.pipeline.components.base import \
     AutoSklearnPreprocessingAlgorithm
@@ -22,7 +23,8 @@ def __init__(self, alpha, mode='fpr',
         elif score_func == "f_classif":
             self.score_func = sklearn.feature_selection.f_classif
         elif score_func == "mutual_info_classif":
-            self.score_func = sklearn.feature_selection.mutual_info_classif
+            self.score_func = partial(sklearn.feature_selection.mutual_info_classif,
+                                      random_state=self.random_state)
             # mutual info classif constantly crashes without mode percentile
             self.mode = 'percentile'
         else:

autosklearn/pipeline/components/feature_preprocessing/select_rates_regression.py

@@ -2,6 +2,7 @@
 from ConfigSpace.hyperparameters import UniformFloatHyperparameter, \
     CategoricalHyperparameter
 from ConfigSpace import NotEqualsCondition
+from functools import partial
 
 from autosklearn.pipeline.components.base import \
     AutoSklearnPreprocessingAlgorithm
@@ -20,7 +21,8 @@ def __init__(self, alpha, mode='percentile',
         if score_func == "f_regression":
             self.score_func = sklearn.feature_selection.f_regression
         elif score_func == "mutual_info_regression":
-            self.score_func = sklearn.feature_selection.mutual_info_regression
+            self.score_func = partial(sklearn.feature_selection.mutual_info_regression,
+                                      random_state=self.random_state)
             # Mutual info consistently crashes if percentile is not the mode
             self.mode = 'percentile'
         else:

autosklearn/pipeline/components/feature_preprocessing/truncatedSVD.py

@@ -17,7 +17,7 @@ def fit(self, X, Y):
         self.target_dim = int(self.target_dim)
         target_dim = min(self.target_dim, X.shape[1] - 1)
         self.preprocessor = sklearn.decomposition.TruncatedSVD(
-            target_dim, algorithm='randomized')
+            target_dim, algorithm='randomized',  random_state=self.random_state)
         # TODO: remove when migrating to sklearn 0.16
         # Circumvents a bug in sklearn
         # https://github.com/scikit-learn/scikit-learn/commit/f08b8c8e52663167819f242f605db39f3b5a6d0c