Implementing the MLM Sampler (#134)

astartes/samplers/interpolation/mlm.py

@@ -0,0 +1,46 @@
+from typing import overload
+
+import numpy as np
+
+from astartes.samplers.interpolation import KennardStone
+
+
+class MLM(KennardStone):
+    # could be convenient to know size of train and test during init...
+    @overload
+    def get_sample_idxs(self, n_samples):
+        """Overload the KennardStone method to permute 10% of samples from train
+
+        Args:
+            n_samples (int): Number of samples to retrieve.
+
+        Returns:
+            np.array: The selected indices
+        """
+        if self._current_sample_idx == 0:  # permute indexes on the first call
+            train_idxs = self._samples_idxs[
+                self._current_sample_idx : self._current_sample_idx + n_samples
+            ]
+            other_idxs = self._samples_idxs[self._current_sample_idx + n_samples : -1]
+
+            # set RNG
+            rng = np.random.default_rng(seed=self.get_config("random_state"))
+            n_to_permute = np.floor(0.1 * len(train_idxs))
+            train_permute_idxs = rng.choice(train_idxs, n_to_permute)
+            remaining_train_idxs = [
+                i for i in train_idxs if i not in train_permute_idxs
+            ]
+            other_permute_idxs = rng.choice(other_idxs, n_to_permute)
+            remaining_other_idxs = [
+                i for i in other_idxs if i not in other_permute_idxs
+            ]
+            # reassamble the indexes
+            self._samples_idxs = np.hstack(
+                (
+                    remaining_train_idxs,
+                    other_permute_idxs,
+                    remaining_other_idxs,
+                    train_permute_idxs,
+                )
+            )
+        return super().get_sample_idxs(n_samples)