[fix] Address Ravin's comments and fix range issues in row cut

Author: nabenabe0928

autoPyTorch/pipeline/components/training/data_loader/base_data_loader.py

@@ -64,10 +64,10 @@ def transform(self, X: Dict[str, Any]) -> Dict[str, Any]:
         underlying model and returns the transformed array.
 
         Args:
-            X (Dict[str, Any])): 'X' dictionary
+            X (Dict[str, Any])): fit dictionary
 
         Returns:
-            (Dict[str, Any]): the updated 'X' dictionary
+            (Dict[str, Any]): the updated fit dictionary
         """
         X.update({'train_data_loader': self.train_data_loader,
                   'val_data_loader': self.val_data_loader,

autoPyTorch/pipeline/components/training/trainer/GridCutMixTrainer.py

@@ -38,11 +38,11 @@ def data_preparation(self, X: np.ndarray, y: np.ndarray,
         lam = self.random_state.beta(alpha, beta)
         batch_size, _, W, H = X.shape
         device = torch.device('cuda' if X.is_cuda else 'cpu')
-        batch_indices = torch.randperm(batch_size).to(device)
+        permed_indices = torch.randperm(batch_size).to(device)
 
         r = self.random_state.rand(1)
         if beta <= 0 or r > self.alpha:
-            return X, {'y_a': y, 'y_b': y[batch_indices], 'lam': 1}
+            return X, {'y_a': y, 'y_b': y[permed_indices], 'lam': 1}
 
         # Draw parameters of a random bounding box
         # Where to cut basically
@@ -56,13 +56,13 @@ def data_preparation(self, X: np.ndarray, y: np.ndarray,
         bbx2 = np.clip(cx + cut_w // 2, 0, W)
         bby2 = np.clip(cy + cut_h // 2, 0, H)
 
-        X[:, :, bbx1:bbx2, bby1:bby2] = X[batch_indices, :, bbx1:bbx2, bby1:bby2]
+        X[:, :, bbx1:bbx2, bby1:bby2] = X[permed_indices, :, bbx1:bbx2, bby1:bby2]
 
         # Adjust lam
         pixel_size = W * H
         lam = 1 - ((bbx2 - bbx1) * (bby2 - bby1) / pixel_size)
 
-        y_a, y_b = y, y[batch_indices]
+        y_a, y_b = y, y[permed_indices]
 
         return X, {'y_a': y_a, 'y_b': y_b, 'lam': lam}
 

autoPyTorch/pipeline/components/training/trainer/RowCutMixTrainer.py

@@ -30,27 +30,29 @@ def data_preparation(self, X: np.ndarray, y: np.ndarray,
         lam = self.random_state.beta(alpha, beta)
         batch_size = X.shape[0]
         device = torch.device('cuda' if X.is_cuda else 'cpu')
-        batch_indices = torch.randperm(batch_size).to(device)
+        permed_indices = torch.randperm(batch_size).to(device)
 
         r = self.random_state.rand(1)
         if beta <= 0 or r > self.alpha:
-            return X, {'y_a': y, 'y_b': y[batch_indices], 'lam': 1}
+            return X, {'y_a': y, 'y_b': y[permed_indices], 'lam': 1}
 
-        row_size = X.shape[1]
-        row_indices = torch.tensor(
+        # batch_size (permutation of rows), col_size = X.shape
+        col_size = X.shape[1]
+        col_indices = torch.tensor(
             self.random_state.choice(
-                range(1, row_size),
-                max(1, int(row_size * lam)),
+                range(col_size),
+                max(1, int(col_size * lam)),
                 replace=False
             )
         )
 
-        X[:, row_indices] = X[batch_indices, :][:, row_indices]
+        # Replace selected columns with columns from another data point
+        X[:, col_indices] = X[permed_indices, :][:, col_indices]
 
         # Adjust lam
-        lam = 1 - len(row_indices) / X.shape[1]
+        lam = 1 - len(col_indices) / X.shape[1]
 
-        y_a, y_b = y, y[batch_indices]
+        y_a, y_b = y, y[permed_indices]
 
         return X, {'y_a': y_a, 'y_b': y_b, 'lam': lam}
 

autoPyTorch/pipeline/components/training/trainer/RowCutOutTrainer.py

@@ -46,8 +46,9 @@ def data_preparation(self, X: np.ndarray, y: np.ndarray,
             lam = 1
             return X, {'y_a': y_a, 'y_b': y_b, 'lam': lam}
 
-        row_size = X.shape[1]
-        row_indices = self.random_state.choice(range(1, row_size), max(1, int(row_size * self.patch_ratio)),
+        # (batch_size (permutation of rows), col_size) = X.shape
+        col_size = X.shape[1]
+        col_indices = self.random_state.choice(range(col_size), max(1, int(col_size * self.patch_ratio)),
                                                replace=False)
 
         if not isinstance(self.numerical_columns, typing.Iterable):
@@ -56,7 +57,7 @@ def data_preparation(self, X: np.ndarray, y: np.ndarray,
                                                   self.numerical_columns))
 
         numerical_indices = torch.tensor(self.numerical_columns)
-        categorical_indices = torch.tensor([idx for idx in row_indices if idx not in self.numerical_columns])
+        categorical_indices = torch.tensor([idx for idx in col_indices if idx not in self.numerical_columns])
 
         X[:, categorical_indices.long()] = self.CATEGORICAL_VALUE
         X[:, numerical_indices.long()] = self.NUMERICAL_VALUE