Implemented BALD acquisition function within QueryInstanceQBC

alipy/query_strategy/query_labels.py

@@ -357,7 +357,7 @@ class QueryInstanceQBC(BaseIndexQuery):
         Method name. This class only implement query_by_bagging for now.
 
     disagreement: str
-        method to calculate disagreement of committees. should be one of ['vote_entropy', 'KL_divergence']
+        method to calculate disagreement of committees. should be one of ['vote_entropy', 'KL_divergence', 'BALD']
 
     References
     ----------
@@ -373,10 +373,10 @@ class QueryInstanceQBC(BaseIndexQuery):
     def __init__(self, X=None, y=None, method='query_by_bagging', disagreement='vote_entropy'):
         self._method = method
         super(QueryInstanceQBC, self).__init__(X, y)
-        if disagreement in ['vote_entropy', 'KL_divergence']:
+        if disagreement in ['vote_entropy', 'KL_divergence', 'BALD']:
             self._disagreement = disagreement
         else:
-            raise ValueError("disagreement must be one of ['vote_entropy', 'KL_divergence']")
+            raise ValueError("disagreement must be one of ['vote_entropy', 'KL_divergence', 'BALD']")
 
     def select(self, label_index, unlabel_index, model=None, batch_size=1, n_jobs=None):
         """Select indexes from the unlabel_index for querying.
@@ -438,8 +438,10 @@ def select(self, label_index, unlabel_index, model=None, batch_size=1, n_jobs=No
         # calc score
         if self._disagreement == 'vote_entropy':
             score = self.calc_vote_entropy([estimator.predict(unlabel_x) for estimator in est_arr])
-        else:
+        elif self._disagreement == 'KL_divergence':
             score = self.calc_avg_KL_divergence([estimator.predict_proba(unlabel_x) for estimator in est_arr])
+        else:
+            score = self.calc_BALD([estimator.predict_proba(unlabel_x) for estimator in est_arr])
         return unlabel_index[nlargestarg(score, batch_size)]
 
     def select_by_prediction_mat(self, unlabel_index, predict, batch_size=1):
@@ -598,6 +600,42 @@ def calc_avg_KL_divergence(cls, predict_matrices):
                 "A 2D probabilistic prediction matrix must be provided, with the shape like [n_samples, n_class]")
         return score
 
+    @classmethod
+    def calc_BALD(cls, predict_matrices):
+        """Calculate the mutual information between model predictions and model parameters
+        for measuring the level of disagreement in QBC.
+
+        Parameters
+        ----------
+        predict_matrices: list
+            The prediction matrix for each committee.
+            Each committee predict matrix should have the shape [n_samples, n_classes] for probabilistic output
+            or [n_samples] for class output.
+
+        Returns
+        -------
+        score: list
+            Score for each instance. Shape [n_samples]
+
+        References
+        ----------
+        [1] Kirsch, Andreas, Joost Van Amersfoort, and Yarin Gal. "Batchbald: Efficient
+        and diverse batch acquisition for deep bayesian active learning." Advances in neural
+        information processing systems 32 (2019): 7026-7037.
+        """
+        input_shape, committee_size = cls()._check_committee_results(predict_matrices)
+        if len(input_shape) == 2:
+            prob = np.array(predict_matrices)
+            pb = np.mean(prob, axis=0)
+            entropy1 = (-prob*np.log(prob)).sum(axis=2).mean(axis=0)
+            entropy2 = (-pb*np.log(pb)).sum(axis=1)
+            bald_scores = entropy2 - entropy1
+            score = bald_scores.tolist()
+        else:
+            raise Exception(
+                "A 2D probabilistic prediction matrix must be provided, with the shape like [n_samples, n_class]")
+        return score
+
 
 class QueryExpectedErrorReduction(BaseIndexQuery):
     """The Expected Error Reduction (ERR) algorithm.