Add support for bert-sklearn #minor (#74)

docs/examples/plot_bert.py

@@ -0,0 +1,37 @@
+# -*- coding: utf-8 -*-
+"""
+=====================
+BERT sklearn
+=====================
+
+In order to use `bert-sklearn <https://github.com/charles9n/bert-sklearn>`_ with HiClass, some of scikit-learns checks need to be disabled.
+The reason is that BERT expects text as input for the features, but scikit-learn expects numerical features.
+Hence, the checks will fail.
+To disable scikit-learn's checks, we can simply use the parameter `bert=True` in the constructor of the local hierarchical classifier.
+"""
+from bert_sklearn import BertClassifier
+from hiclass import LocalClassifierPerParentNode
+
+# Define data
+X_train = X_test = [
+    "Batman",
+    "Rorschach",
+]
+Y_train = [
+    ["Action", "The Dark Night"],
+    ["Action", "Watchmen"],
+]
+
+# Use BERT for every node
+bert = BertClassifier()
+classifier = LocalClassifierPerParentNode(
+    local_classifier=bert,
+    bert=True,
+)
+
+# Train local classifier per node
+classifier.fit(X_train, Y_train)
+
+# Predict
+predictions = classifier.predict(X_test)
+print(predictions)

docs/requirements.txt

@@ -8,3 +8,4 @@ matplotlib==3.5.2
 pandas==1.4.2
 ray==1.13.0
 numpy<1.24
+git+https://github.com/charles9n/bert-sklearn.git@master

hiclass/HierarchicalClassifier.py

@@ -65,6 +65,7 @@ def __init__(
         edge_list: str = None,
         replace_classifiers: bool = True,
         n_jobs: int = 1,
+        bert: bool = False,
         classifier_abbreviation: str = "",
     ):
         """
@@ -87,6 +88,8 @@ def __init__(
         n_jobs : int, default=1
             The number of jobs to run in parallel. Only :code:`fit` is parallelized.
             If :code:`Ray` is installed it is used, otherwise it defaults to :code:`Joblib`.
+        bert : bool, default=False
+            If True, skip scikit-learn's checks and sample_weight passing for BERT.
         classifier_abbreviation : str, default=""
             The abbreviation of the local hierarchical classifier to be displayed during logging.
         """
@@ -95,6 +98,7 @@ def __init__(
         self.edge_list = edge_list
         self.replace_classifiers = replace_classifiers
         self.n_jobs = n_jobs
+        self.bert = bert
         self.classifier_abbreviation = classifier_abbreviation
 
     def fit(self, X, y, sample_weight=None):
@@ -130,9 +134,13 @@ def _pre_fit(self, X, y, sample_weight):
         # Check that X and y have correct shape
         # and convert them to np.ndarray if need be
 
-        self.X_, self.y_ = self._validate_data(
-            X, y, multi_output=True, accept_sparse="csr"
-        )
+        if not self.bert:
+            self.X_, self.y_ = self._validate_data(
+                X, y, multi_output=True, accept_sparse="csr"
+            )
+        else:
+            self.X_ = np.array(X)
+            self.y_ = np.array(y)
 
         if sample_weight is not None:
             self.sample_weight_ = _check_sample_weight(sample_weight, X)

hiclass/LocalClassifierPerLevel.py

@@ -47,6 +47,7 @@ def __init__(
         edge_list: str = None,
         replace_classifiers: bool = True,
         n_jobs: int = 1,
+        bert: bool = False,
     ):
         """
         Initialize a local classifier per level.
@@ -68,6 +69,8 @@ def __init__(
         n_jobs : int, default=1
             The number of jobs to run in parallel. Only :code:`fit` is parallelized.
             If :code:`Ray` is installed it is used, otherwise it defaults to :code:`Joblib`.
+        bert : bool, default=False
+            If True, skip scikit-learn's checks and sample_weight passing for BERT.
         """
         super().__init__(
             local_classifier=local_classifier,
@@ -76,6 +79,7 @@ def __init__(
             replace_classifiers=replace_classifiers,
             n_jobs=n_jobs,
             classifier_abbreviation="LCPL",
+            bert=bert,
         )
 
     def fit(self, X, y, sample_weight=None):
@@ -135,7 +139,10 @@ def predict(self, X):
         check_is_fitted(self)
 
         # Input validation
-        X = check_array(X, accept_sparse="csr")
+        if not self.bert:
+            X = check_array(X, accept_sparse="csr")
+        else:
+            X = np.array(X)
 
         # Initialize array that holds predictions
         y = np.empty((X.shape[0], self.max_levels_), dtype=self.dtype_)
@@ -242,7 +249,10 @@ def _fit_classifier(self, level, separator):
         unique_y = np.unique(y)
         if len(unique_y) == 1 and self.replace_classifiers:
             classifier = ConstantClassifier()
-        classifier.fit(X, y, sample_weight)
+        if not self.bert:
+            classifier.fit(X, y, sample_weight)
+        else:
+            classifier.fit(X, y)
         return classifier
 
     @staticmethod

hiclass/LocalClassifierPerNode.py

@@ -42,6 +42,7 @@ def __init__(
         edge_list: str = None,
         replace_classifiers: bool = True,
         n_jobs: int = 1,
+        bert: bool = False,
     ):
         """
         Initialize a local classifier per node.
@@ -74,6 +75,8 @@ def __init__(
         n_jobs : int, default=1
             The number of jobs to run in parallel. Only :code:`fit` is parallelized.
             If :code:`Ray` is installed it is used, otherwise it defaults to :code:`Joblib`.
+        bert : bool, default=False
+            If True, skip scikit-learn's checks and sample_weight passing for BERT.
         """
         super().__init__(
             local_classifier=local_classifier,
@@ -82,6 +85,7 @@ def __init__(
             replace_classifiers=replace_classifiers,
             n_jobs=n_jobs,
             classifier_abbreviation="LCPN",
+            bert=bert,
         )
         self.binary_policy = binary_policy
 
@@ -145,7 +149,10 @@ def predict(self, X):
         check_is_fitted(self)
 
         # Input validation
-        X = check_array(X, accept_sparse="csr")
+        if not self.bert:
+            X = check_array(X, accept_sparse="csr")
+        else:
+            X = np.array(X)
 
         # Initialize array that holds predictions
         y = np.empty((X.shape[0], self.max_levels_), dtype=self.dtype_)
@@ -233,7 +240,10 @@ def _fit_classifier(self, node):
         unique_y = np.unique(y)
         if len(unique_y) == 1 and self.replace_classifiers:
             classifier = ConstantClassifier()
-        classifier.fit(X, y, sample_weight)
+        if not self.bert:
+            classifier.fit(X, y, sample_weight)
+        else:
+            classifier.fit(X, y)
         return classifier
 
     def _clean_up(self):

hiclass/LocalClassifierPerParentNode.py

@@ -40,6 +40,7 @@ def __init__(
         edge_list: str = None,
         replace_classifiers: bool = True,
         n_jobs: int = 1,
+        bert: bool = False,
     ):
         """
         Initialize a local classifier per parent node.
@@ -61,6 +62,8 @@ def __init__(
         n_jobs : int, default=1
             The number of jobs to run in parallel. Only :code:`fit` is parallelized.
             If :code:`Ray` is installed it is used, otherwise it defaults to :code:`Joblib`.
+        bert : bool, default=False
+            If True, skip scikit-learn's checks and sample_weight passing for BERT.
         """
         super().__init__(
             local_classifier=local_classifier,
@@ -69,6 +72,7 @@ def __init__(
             replace_classifiers=replace_classifiers,
             n_jobs=n_jobs,
             classifier_abbreviation="LCPPN",
+            bert=bert,
         )
 
     def fit(self, X, y, sample_weight=None):
@@ -128,7 +132,10 @@ def predict(self, X):
         check_is_fitted(self)
 
         # Input validation
-        X = check_array(X, accept_sparse="csr")
+        if not self.bert:
+            X = check_array(X, accept_sparse="csr")
+        else:
+            X = np.array(X)
 
         # Initialize array that holds predictions
         y = np.empty((X.shape[0], self.max_levels_), dtype=self.dtype_)
@@ -203,7 +210,10 @@ def _fit_classifier(self, node):
         unique_y = np.unique(y)
         if len(unique_y) == 1 and self.replace_classifiers:
             classifier = ConstantClassifier()
-        classifier.fit(X, y, sample_weight)
+        if not self.bert:
+            classifier.fit(X, y, sample_weight)
+        else:
+            classifier.fit(X, y)
         return classifier
 
     def _fit_digraph(self, local_mode: bool = False, use_joblib: bool = False):

tests/test_HierarchicalClassifier.py

@@ -214,3 +214,13 @@ def test_make_leveled_non_iterable_y(noniterable_y):
 def test_fit_classifier():
     with pytest.raises(NotImplementedError):
         HierarchicalClassifier._fit_classifier(None, None)
+
+
+def test_pre_fit_bert():
+    classifier = HierarchicalClassifier()
+    classifier.logger_ = logging.getLogger("HC")
+    classifier.bert = True
+    x = [[0, 1], [2, 3]]
+    y = [["a", "b"], ["c", "d"]]
+    sample_weight = None
+    classifier._pre_fit(x, y, sample_weight)

tests/test_LocalClassifierPerLevel.py

@@ -10,6 +10,7 @@
 from sklearn.utils.estimator_checks import parametrize_with_checks
 from sklearn.utils.validation import check_is_fitted
 from hiclass import LocalClassifierPerLevel
+from hiclass.ConstantClassifier import ConstantClassifier
 
 
 @parametrize_with_checks([LocalClassifierPerLevel()])
@@ -180,3 +181,17 @@ def test_empty_levels(empty_levels):
         lcppn.root_,
     ]
     assert_array_equal(ground_truth, predictions)
+
+
+def test_fit_bert():
+    bert = ConstantClassifier()
+    lcpn = LocalClassifierPerLevel(
+        local_classifier=bert,
+        bert=True,
+    )
+    X = ["Text 1", "Text 2"]
+    y = ["a", "a"]
+    lcpn.fit(X, y)
+    check_is_fitted(lcpn)
+    predictions = lcpn.predict(X)
+    assert_array_equal(y, predictions)

tests/test_LocalClassifierPerNode.py

@@ -12,6 +12,7 @@
 
 from hiclass import LocalClassifierPerNode
 from hiclass.BinaryPolicy import ExclusivePolicy
+from hiclass.ConstantClassifier import ConstantClassifier
 
 
 @parametrize_with_checks([LocalClassifierPerNode()])
@@ -242,3 +243,17 @@ def test_empty_levels(empty_levels):
         lcppn.root_,
     ]
     assert_array_equal(ground_truth, predictions)
+
+
+def test_fit_bert():
+    bert = ConstantClassifier()
+    lcpn = LocalClassifierPerNode(
+        local_classifier=bert,
+        bert=True,
+    )
+    X = ["Text 1", "Text 2"]
+    y = ["a", "a"]
+    lcpn.fit(X, y)
+    check_is_fitted(lcpn)
+    predictions = lcpn.predict(X)
+    assert_array_equal(y, predictions)

tests/test_LocalClassifierPerParentNode.py

@@ -12,6 +12,7 @@
 from sklearn.utils.validation import check_is_fitted
 
 from hiclass import LocalClassifierPerParentNode
+from hiclass.ConstantClassifier import ConstantClassifier
 
 
 @parametrize_with_checks([LocalClassifierPerParentNode()])
@@ -226,3 +227,17 @@ def test_empty_levels(empty_levels):
         lcppn.root_,
     ]
     assert_array_equal(ground_truth, predictions)
+
+
+def test_bert():
+    bert = ConstantClassifier()
+    lcpn = LocalClassifierPerParentNode(
+        local_classifier=bert,
+        bert=True,
+    )
+    X = ["Text 1", "Text 2"]
+    y = ["a", "a"]
+    lcpn.fit(X, y)
+    check_is_fitted(lcpn)
+    predictions = lcpn.predict(X)
+    assert_array_equal(y, predictions)