Add dropout to fix batch size mismatch and device mismatch in modular_attack

src/secmlt/adv/evasion/advlib_attacks/advlib_base.py

@@ -64,7 +64,7 @@ def _run(
         samples: torch.Tensor,
         labels: torch.Tensor,
     ) -> torch.Tensor:
-        if not isinstance(model, BasePytorchClassifier):
+        if not isinstance(model, BasePytorchClassifier) and not isinstance(model, EnsembleModel):
             msg = "Model type not supported."
             raise NotImplementedError(msg)
         device = model._get_device()
@@ -74,7 +74,7 @@ def _run(
             model=model,
             inputs=samples,
             labels=labels,
-            ε=self.epsilon,
+            eps=self.epsilon,
             targeted=(self.y_target is not None),
             loss_function=self.loss_function,
         )

src/secmlt/adv/evasion/modular_attack.py

@@ -232,14 +232,14 @@ def _run(
 
             # keep perturbation with highest loss
             best_delta.data = torch.where(
-                atleast_kd(losses.detach().cpu() < best_losses, len(samples.shape)),
-                delta.data,
-                best_delta.data,
+                atleast_kd(losses.detach().cpu() < best_losses.detach().cpu(), len(samples.shape)),
+                delta.detach().cpu().data,
+                best_delta.detach().cpu().data,
             )
             best_losses.data = torch.where(
-                losses.detach().cpu() < best_losses,
+                losses.detach().cpu() < best_losses.detach().cpu(),
                 losses.detach().cpu(),
-                best_losses.data,
+                best_losses.detach().cpu().data,
             )
-        x_adv, _ = self.manipulation_function(samples.data, best_delta.data)
+        x_adv, _ = self.manipulation_function(samples.detach().cpu().data, best_delta.data)
         return x_adv, best_delta

src/secmlt/trackers/trackers.py

@@ -68,6 +68,10 @@ def get(self) -> torch.Tensor:
         torch.Tensor
             History of tracked parameters.
         """
+        # Dropout of the last batch element if there is a size mismatch
+        reference_size = self.tracked[0].size()
+        if self.tracked[-1].size() != reference_size:
+            self.tracked.pop()
         return torch.stack(self.tracked, -1)
 
     def get_last_tracked(self) -> Union[None, torch.Tensor]: