diff --git a/codes/models/classifiers/cifar_resnet_branched.py b/codes/models/classifiers/cifar_resnet_branched.py
index f5a53719..e314b22e 100644
--- a/codes/models/classifiers/cifar_resnet_branched.py
+++ b/codes/models/classifiers/cifar_resnet_branched.py
@@ -11,6 +11,7 @@
 import torch
 import torch.nn as nn
 
+from models.switched_conv.switched_conv_hard_routing import HardRoutingGate
 from trainer.networks import register_model
 
 
@@ -111,7 +112,7 @@ class ResNetTail(nn.Module):
 
 class ResNet(nn.Module):
 
-    def __init__(self, block, num_block, num_classes=100, num_tails=8):
+    def __init__(self, block, num_block, num_classes=100, num_tails=8, dropout_rate=.2):
         super().__init__()
         self.in_channels = 32
         self.conv1 = nn.Sequential(
@@ -123,6 +124,9 @@ class ResNet(nn.Module):
         self.conv3_x = self._make_layer(block, 64, num_block[1], 2)
         self.tails = nn.ModuleList([ResNetTail(block, num_block, 256) for _ in range(num_tails)])
         self.selector = ResNetTail(block, num_block, num_tails)
+        self.selector_gate = nn.Linear(256, 1)
+        self.gate = HardRoutingGate(num_tails, hard_en=True)
+        self.dropout_rate = dropout_rate
         self.final_linear = nn.Linear(256, num_classes)
 
     def _make_layer(self, block, out_channels, num_blocks, stride):
@@ -144,8 +148,16 @@ class ResNet(nn.Module):
         keys = torch.stack(keys, dim=1)
 
         query = self.selector(output).unsqueeze(2)
-        attn = torch.nn.functional.softmax(query * keys, dim=1)
-        values = self.final_linear(attn * keys)
+        selector = self.selector_gate(query * keys).squeeze(-1)
+        if self.training and self.dropout_rate > 0:
+            bs, br = selector.shape
+            drop = torch.rand((bs, br), device=x.device) > self.dropout_rate
+            # Ensure that there is always at least one switch left un-dropped out
+            fix_blank = (drop.sum(dim=1, keepdim=True) == 0).repeat(1, br)
+            drop = drop.logical_or(fix_blank)
+            selector = drop * selector
+        selector = self.gate(selector)
+        values = self.final_linear(selector.unsqueeze(-1) * keys)
 
         return values.sum(dim=1)
 
@@ -181,5 +193,8 @@ def resnet152():
 
 if __name__ == '__main__':
     model = ResNet(BasicBlock, [2,2,2,2])
-    print(model(torch.randn(2,3,32,32), torch.LongTensor([4,7])).shape)
+    v = model(torch.randn(2,3,32,32), torch.LongTensor([4,7]))
+    print(v.shape)
+    l = nn.MSELoss()(v, torch.randn_like(v))
+    l.backward()
 
diff --git a/codes/models/switched_conv/switched_conv_hard_routing.py b/codes/models/switched_conv/switched_conv_hard_routing.py
index 5eecb313..73e9ad09 100644
--- a/codes/models/switched_conv/switched_conv_hard_routing.py
+++ b/codes/models/switched_conv/switched_conv_hard_routing.py
@@ -63,7 +63,9 @@ class SwitchedConvHardRoutingFunction(torch.autograd.Function):
 class RouteTop1(torch.autograd.Function):
     @staticmethod
     def forward(ctx, input):
-        mask = torch.nn.functional.one_hot(input.argmax(dim=1), num_classes=input.shape[1]).permute(0,3,1,2)
+        mask = torch.nn.functional.one_hot(input.argmax(dim=1), num_classes=input.shape[1])
+        if len(input.shape) > 2:
+            mask = mask.permute(0, 3, 1, 2)  # TODO: Make this more extensible.
         out = torch.ones_like(input)
         out[mask != 1] = 0
         ctx.save_for_backward(mask, input.clone())
@@ -116,7 +118,8 @@ class SwitchNorm(nn.Module):
         self.register_buffer("accumulator", torch.zeros(accumulator_size, group_size))
 
     def add_norm_to_buffer(self, x):
-        flat = x.sum(dim=[0, 2, 3])
+        flatten_dims = [0] + [k+2 for k in range(len(x)-2)]
+        flat = x.sum(dim=flatten_dims)
         norm = flat / torch.mean(flat)
 
         self.accumulator[self.accumulator_index] = norm.detach().clone()
@@ -126,9 +129,9 @@ class SwitchNorm(nn.Module):
             if self.accumulator_filled <= 0:
                 self.accumulator_filled += 1
 
-    # Input into forward is a switching tensor of shape (batch,groups,width,height)
+    # Input into forward is a switching tensor of shape (batch,groups,<misc>)
     def forward(self, x: torch.Tensor, update_attention_norm=True):
-        assert len(x.shape) == 4
+        assert len(x.shape) >= 2
 
         # Push the accumulator to the right device on the first iteration.
         if self.accumulator.device != x.device:
@@ -148,7 +151,10 @@ class SwitchNorm(nn.Module):
             norm = torch.mean(self.accumulator, dim=0)
         else:
             norm = torch.ones(self.group_size, device=self.accumulator.device)
-        x = x / norm.view(1,-1,1,1)
+
+        while len(x.shape) < len(norm.shape):
+            norm = norm.unsqueeze(-1)
+        x = x / norm
 
         # Need to re-normalize x so that the groups dimension sum to 1, just like when it was fed in.
         return x / x.sum(dim=1, keepdim=True)