fix augmentations order

Author: Demoren1

code/DARTS_baseline.py

@@ -84,8 +84,8 @@ def run(self):
         self.config.selected_archs = []
         shutil.rmtree(self.config.output_path, ignore_errors=True)
 
-        for _ in tqdm(range(self.config.n_ensemble_models), desc="Finding best architectures"):
-            train_loader, valid_loader, test_loader = self.get_data_loaders()
+        for idx, _ in enumerate(tqdm(range(self.config.n_ensemble_models), desc="Finding best architectures")):
+            train_loader, valid_loader, test_loader = self.get_data_loaders(seed=self.config.seed + idx * 10)
             self.config.selected_archs.append(self.get_best_models(train_loader, valid_loader))
 
         for idx, arch in enumerate(tqdm(self.config.selected_archs, desc="Training models")):

code/train_models.py

@@ -76,18 +76,18 @@ def __init__(self, config: TrainConfig):
             config.device if torch.cuda.is_available() else "cpu"
         )
 
-    def get_data_loaders(self, batch_size: int = None) -> Tuple[DataLoader, DataLoader]:
+    def get_data_loaders(self, batch_size: int = None, seed=None) -> Tuple[DataLoader, DataLoader]:
         """
         Create training and validation data loaders for the chosen dataset.
         """
         bs = batch_size if batch_size else self.config.batch_size_final
         train_transform = transforms.Compose(
             self.base_transform
             + [
+                transforms.RandomHorizontalFlip(),
+                transforms.RandomCrop(32, padding=4),
                 transforms.ToTensor(),
                 transforms.Normalize(self.MEAN, self.STD),
-                transforms.RandomCrop(32, padding=4),
-                transforms.RandomHorizontalFlip(),
             ]
         )
         test_transform = transforms.Compose(
@@ -123,6 +123,8 @@ def get_data_loaders(self, batch_size: int = None) -> Tuple[DataLoader, DataLoad
 
         num_samples = len(train_data)
         indices = list(range(num_samples))
+        np_seed = seed if seed else self.config.seed
+        np.random.seed(np_seed)
         np.random.shuffle(indices)
 
         if self.config.evaluate_ensemble_flag: