'comment'

models.py

@@ -59,24 +59,18 @@ def forward(self, cw_idxs, qw_idxs, cwf):
 
         c_emb = self.emb(cw_idxs)         # (batch_size, c_len, hidden_size)
         q_emb = self.emb(qw_idxs)         # (batch_size, q_len, hidden_size)
-
 
-        # s = c_emb.shape
-        # cf_emb = torch.zeros(s[0],s[1],1,device='cuda')
+        #add word feature for content        
         cwf = torch.unsqueeze(cwf, dim = 2)
-
         cwf = cwf.type(torch.cuda.FloatTensor)
         cwf.to('cuda')
-
         ct_emb = torch.cat((c_emb, cwf), dim = 2)
+
+        #for dimension consistent, add a feature of all-zero for questions
         s = q_emb.shape
         qf_emb = torch.zeros(s[0],s[1],1,device='cuda')
         qt_emb = torch.cat((q_emb, qf_emb), dim = 2)
 
-        # for index in range(len(cw_idxs)):
-        #     for i, word_id in enumerate(cw_idxs[index]):
-        #         if word_id in qw_idxs[index]:
-        #             ct_emb[index][i][-1] = 1
 
         c_enc = self.enc(ct_emb, c_len)    # (batch_size, c_len, 2 * hidden_size)
         q_enc = self.enc(qt_emb, q_len)    # (batch_size, q_len, 2 * hidden_size)

util.py

@@ -82,6 +82,7 @@ def __init__(self, data_path, use_v2=True):
         #             self.context_word_features[i][j][0] = 1
 
 
+    #compute word features(exact match)
     def compute_context_word_features(self, idx):
         s = self.context_idxs[idx].shape
         context_word_features = torch.zeros(s[0])
@@ -95,8 +96,6 @@ def compute_context_word_features(self, idx):
         return context_word_features
 
 
-
-
     def __getitem__(self, idx):
         idx = self.valid_idxs[idx]
         example = (self.context_idxs[idx],
@@ -106,6 +105,7 @@ def __getitem__(self, idx):
                    self.y1s[idx],
                    self.y2s[idx],
                    self.ids[idx],
+                   #get word features
                    self.compute_context_word_features(idx)
                    )
 
@@ -153,6 +153,7 @@ def merge_2d(matrices, dtype=torch.int64, pad_value=0):
         return padded
 
     # Group by tensor type
+    #add cwf for word features for content
     context_idxs, context_char_idxs, \
         question_idxs, question_char_idxs, \
         y1s, y2s, ids, cwf = zip(*examples)