Attempt to fix nan

codes/models/classifiers/cifar_resnet_branched.py

@@ -158,6 +158,7 @@ class DropoutNorm(SwitchNorm):
 
         # Compute the dropout probabilities. This module is a no-op before the accumulator is initialized.
         if self.accumulator_filled > 0:
+            with torch.no_grad():
                 probs = torch.mean(self.accumulator, dim=0) * self.dropout_rate
                 bs, br = x.shape[:2]
                 drop = torch.rand((bs, br), device=x.device) > probs.unsqueeze(0)
@@ -172,7 +173,7 @@ class DropoutNorm(SwitchNorm):
 class HardRoutingGate(nn.Module):
     def __init__(self, breadth, dropout_rate=.8):
         super().__init__()
-        self.norm = DropoutNorm(breadth, dropout_rate, accumulator_size=2)
+        self.norm = DropoutNorm(breadth, dropout_rate, accumulator_size=128)
 
     def forward(self, x):
         soft = self.norm(nn.functional.softmax(x, dim=1))