For truncated BPTT, use built-in detach

For GANs, detach variables of fake images when training discriminator.

tutorials/08 - Language Model/main-gpu.py

@@ -61,7 +61,7 @@ optimizer = torch.optim.Adam(model.parameters(), lr=learning_rate)
 
 # Truncated Backpropagation 
 def detach(states):
-    return [Variable(state.data) for state in states] 
+    return [state.detach() for state in states] 
 
 # Training
 for epoch in range(num_epochs):
@@ -119,4 +119,4 @@ with open(sample_path, 'w') as f:
             print('Sampled [%d/%d] words and save to %s'%(i+1, num_samples, sample_path))
 
 # Save the Trained Model
-torch.save(model.state_dict(), 'model.pkl')
+torch.save(model.state_dict(), 'model.pkl')