if RNN's hx is None, requires_grad=False (pytorch#2274)

donglixp · soumith · commit f6585e80d7f5 · 2017-08-03T05:29:50.000+05:30
When the initial hidden states of RNN are ``None'', we don't need to compute their gradients.
diff --git a/torch/nn/modules/rnn.py b/torch/nn/modules/rnn.py
@@ -135,7 +135,7 @@ def forward(self, input, hx=None):
             hx = torch.autograd.Variable(input.data.new(self.num_layers *
                                                         num_directions,
                                                         max_batch_size,
-                                                        self.hidden_size).zero_())
+                                                        self.hidden_size).zero_(), requires_grad=False)
             if self.mode == 'LSTM':
                 hx = (hx, hx)
 
