diff --git a/losses.py b/losses.py
index 1e9a080..ce05b0d 100755
--- a/losses.py
+++ b/losses.py
@@ -10,7 +10,19 @@ Shorthands for loss:
 - TripletLoss: htri
 - CenterLoss: cent
 """
-__all__ = ['CrossEntropyLabelSmooth', 'TripletLoss', 'CenterLoss']
+__all__ = ['DeepSupervision', 'CrossEntropyLabelSmooth', 'TripletLoss', 'CenterLoss']
+
+def DeepSupervision(criterion, xs, y):
+    """
+    Args:
+        criterion: loss function
+        xs: tuple of inputs
+        y: ground truth
+    """
+    loss = 0.
+    for x in xs:
+        loss += criterion(x, y)
+    return loss
 
 class CrossEntropyLabelSmooth(nn.Module):
     """Cross entropy loss with label smoothing regularizer.
diff --git a/models/HACNN.py b/models/HACNN.py
index dc043eb..f5a832c 100644
--- a/models/HACNN.py
+++ b/models/HACNN.py
@@ -188,10 +188,11 @@ class HACNN(nn.Module):
         feat_dim (int): feature dimension for each branch
         learn_region (bool): whether to learn region features (i.e. local branch)
     """
-    def __init__(self, num_classes, loss={'xent'}, nchannels=[128, 256, 384], feat_dim=512, learn_region=True, **kwargs):
+    def __init__(self, num_classes, loss={'xent'}, nchannels=[128, 256, 384], feat_dim=512, learn_region=True, use_gpu=True, **kwargs):
         super(HACNN, self).__init__()
         self.loss = loss
         self.learn_region = learn_region
+        self.use_gpu = use_gpu
 
         self.conv = ConvBlock(3, 32, 3, s=2, p=1)
 
@@ -267,6 +268,7 @@ class HACNN(nn.Module):
         theta = torch.zeros(theta_i.size(0), 2, 3)
         theta[:,:,:2] = scale_factors
         theta[:,:,-1] = theta_i
+        if self.use_gpu: theta = theta.cuda()
         return theta
 
     def forward(self, x):