Call loss consistently

When unrolling models, we were not unpacking the tuple returned by
model.loss like we were in other methods. This keeps things consisent.

Author: yngtodd

common/darts/architecture.py

@@ -44,11 +44,10 @@ def comp_unrolled_model(self, data, target, eta, optimizer):
         model_unrolled
         """
         # forward to get loss
-        loss = self.model.loss(data, target)
+        _, loss = self.model.loss(data, target, reduce='mean')
         # flatten current weights
         theta = F.flatten(self.model.parameters()).detach()
         # theta: torch.Size([1930618])
-        # print('theta:', theta.shape)
         try:
             # fetch momentum data from theta optimizer
             moment = F.flatten(optimizer.state[v]['momentum_buffer'] for v in self.model.parameters())
@@ -57,6 +56,7 @@ def comp_unrolled_model(self, data, target, eta, optimizer):
             moment = torch.zeros_like(theta)
 
         # flatten all gradients
+        gradient= autograd.grad(loss, self.model.parameters(), allow_unused=True)
         dtheta = F.flatten(autograd.grad(loss, self.model.parameters())).data
         # indeed, here we implement a simple SGD with momentum and weight decay
         # theta = theta - eta * (moment + weight decay + dtheta)

examples/darts/uno/default_model.txt

@@ -1,6 +1,6 @@
 [Global_Params]
 model_name = 'darts_uno'
-unrolled = False 
+unrolled = True
 data_url = 'http://ftp.mcs.anl.gov/pub/candle/public/benchmarks/Pilot1/uno/' 
 savepath = '.'
 log_interval = 10