move vit functionality into derandomised smoothing toolset

Signed-off-by: GiulioZizzo <[email protected]>

art/estimators/certification/derandomized_smoothing/vision_transformers/pytorch.py

@@ -106,6 +106,7 @@ def __init__(
         :param device_type: Type of device on which the classifier is run, either `gpu` or `cpu`.
         """
         import timm
+        import torch
         from timm.models.vision_transformer import VisionTransformer
         from art.estimators.certification.derandomized_smoothing.vision_transformers.vit import ArtViT
 
@@ -218,6 +219,7 @@ def get_models(cls, generate_from_null: bool = False) -> List[str]:
         :return: A list of compatible models
         """
         import timm
+        import torch
 
         supported_models = [
             "vit_base_patch8_224",
@@ -330,7 +332,7 @@ def update_batchnorm(self, x: np.ndarray, batch_size: int, nb_epochs: int = 1) -
         :param batch_size: Size of batches.
         :param nb_epochs: How many times to forward pass over the input data
         """
-
+        import torch
         self.model.train()
 
         ind = np.arange(len(x))