deterministic simplex sampling feature added

analyticalmindsltd · Aug 21, 2022 · ef38e72 · ef38e72
1 parent 38ca6ab
commit ef38e72
Show file tree

Hide file tree

Showing 3 changed files with 61 additions and 6 deletions.
diff --git a/smote_variants/_version.py b/smote_variants/_version.py
@@ -6,4 +6,4 @@
 @author: gykovacs
 """
 
-__version__= '0.6.6'
+__version__= '0.6.7'
diff --git a/smote_variants/base/_simplexsampling.py b/smote_variants/base/_simplexsampling.py
@@ -281,6 +281,54 @@ def add_samples(*,
     samples_by_count = [base_points + diffweight*s for s in splits]
     return np.vstack(samples_by_count)
 
+def counts_to_vector(counts):
+    """
+    Expand a count vector to a
+
+    Args:
+        counts (np.array): count vector
+
+    Returns:
+        np.array: the expanded vector
+    """
+
+    return np.hstack([np.repeat(idx, count) for idx, count in enumerate(counts)])
+
+def deterministic_sample(choices, n_to_sample, p):
+    """
+    Take a deterministic sample
+
+    Args:
+        choices (list): the list of choices
+        n_to_sample (int): the number of samples to take
+        p (np.array): the distribution
+
+    Returns:
+        np.array: the choices
+    """
+
+    sample_counts = np.ceil(n_to_sample * p).astype(int)
+
+    n_to_remove = np.sum(sample_counts) - n_to_sample
+
+    if n_to_remove == 0:
+        return choices[counts_to_vector(sample_counts)]
+
+    non_zero_mask = sample_counts > 0
+
+    removal_indices = np.floor(np.linspace(0.0, np.sum(non_zero_mask), n_to_remove, endpoint=False)).astype(int)
+
+    tmp = sample_counts[non_zero_mask]
+    tmp[removal_indices] = tmp[removal_indices] - 1
+
+    sample_counts[non_zero_mask] = tmp
+
+    assert np.sum(sample_counts) == n_to_sample
+
+    samples = choices[counts_to_vector(sample_counts)]
+
+    return samples
+
 class SimplexSamplingMixin(RandomStateMixin):
     """
     The mixin class for all simplex sampling based techniques.
@@ -436,11 +484,17 @@ def simplices(self,
 
         weights = weights * node_weights
 
-        # sample the simplices
         choices = np.arange(all_simplices.shape[0])
-        selected_indices = self.random_state.choice(choices,
-                                                    n_to_sample,
-                                                    p=weights/np.sum(weights))
+
+        if self.simplex_sampling == 'random':
+            # sample the simplices
+            selected_indices = self.random_state.choice(choices,
+                                                        n_to_sample,
+                                                        p=weights/np.sum(weights))
+        elif self.simplex_sampling == 'deterministic':
+            selected_simplices = deterministic_sample(choices,
+                                                        n_to_sample,
+                                                        p=weights/np.sum(weights))
         return all_simplices[selected_indices]
 
     def add_gaussian_noise(self, samples):

diff --git a/tests/oversampling/test_simplex_deterministic.py b/tests/oversampling/test_simplex_deterministic.py
@@ -21,7 +21,8 @@ def test_simplex_deterministic(smote_class):
     Args:
         smote_class (class): an oversampler class.
     """
-    ss_params = {'within_simplex_sampling': 'deterministic'}
+    ss_params = {'within_simplex_sampling': 'deterministic',
+                 'simplex_sampling': 'deterministic'}
     X, y = smote_class(ss_params=ss_params).sample(dataset['data'],
                                                    dataset['target'])