pack padded sequence and reverse

Author: wasimusu

config.py

@@ -11,10 +11,10 @@ def __init__(self):
         self.mode = "train"  # TODO: Why do we want this in config?
 
         # Device params
-        self.use_cuda = torch.cuda.is_available()
-        self.device = ('cuda' if self.use_cuda else 'cpu')
-        # self.device = 'cpu'
-        # self.use_cuda = False
+        # self.use_cuda = torch.cuda.is_available()
+        # self.device = ('cuda' if self.use_cuda else 'cpu')
+        self.device = 'cpu'
+        self.use_cuda = False
 
         # Global dimension params
         self.embedding_dim = 100

networks.py

@@ -57,14 +57,14 @@ def forward(self, inputs, mask):
 
         # Convert the numbers into embeddings
         inputs = self.embeddings(inputs.to('cpu'))
-        packed = inputs
+        # packed = inputs
 
         # Get the sorted version of inputs as required for pack_padded_sequence
-        # inputs_sorted = torch.index_select(inputs, 0, lens_argsort)
+        inputs_sorted = torch.index_select(inputs, 0, lens_argsort)
 
-        # packed = pack_padded_sequence(inputs_sorted, lens, batch_first=True)
+        packed = pack_padded_sequence(inputs_sorted, lens_sorted, batch_first=True)
         output, self.hidden = self.encoder(packed, self.hidden)
-        # output, _ = pad_packed_sequence(output, batch_first=True)
+        output, _ = pad_packed_sequence(output, batch_first=True)
 
         # Restore batch elements to original order
         # output = torch.index_select(output, 0, lens_argsort_argsort)
@@ -119,8 +119,7 @@ def forward(self, inputs, mask):
         output = torch.index_select(output, 0, lens_argsort_argsort.to('cpu'))
 
         # Make output contiguous for speed of future operations
-        # TODO: Try without and time to see if this actually speeds up
-        output = output.contiguous()
+        # output = output.contiguous()
 
         output = self.dropout(output)
         return output
@@ -161,8 +160,6 @@ def __init__(self, device,
                           num_layers=num_layers,
                           bidirectional=bidirectional)
 
-        # self.hidden = self.initHidden() # for GRU
-
     def forward(self, U, d_mask, target_span):
 
         batch_indices = torch.arange(self.batch_size, out=torch.LongTensor(self.batch_size))
@@ -226,10 +223,6 @@ def forward(self, U, d_mask, target_span):
             loss = cumulative_loss / self.max_dec_steps
         return loss, s_i, e_i
 
-    # def initHidden(self):
-    #     return torch.zeros(self.num_directions * self.num_layers, self.batch_size, self.hidden_size)
-
-
 class CoattentionNetwork(nn.Module):
     def __init__(self, device,
                  hidden_size,