Fix arguments

eval.py

@@ -22,7 +22,7 @@
 parser.add_argument('--dataset', default='test', help="Data split to use for evaluation: dev or test.")
 parser.add_argument('--batch_size', type=int, default=100, help="Batch size for evaluation.")
 parser.add_argument('--gold', default='', help="Optional: a file where to write gold summarizations. Default to not write.")
-parser.add_argument('--out', default='', help="Optional: a file where to write predictions.i Default to not write.")
+parser.add_argument('--out', default='', help="Optional: a file where to write predictions. Default to not write.")
 parser.add_argument('--use_bleu', action='store_true', help="Use BLEU instead of ROUGE metrics for scoring.")
 
 parser.add_argument('--seed', type=int, default=1234)

model/copy_model.py

@@ -31,7 +31,6 @@ def __init__(self, opt, emb_matrix=None):
         self.opt = opt
         self.emb_matrix = emb_matrix
         self.use_bg = opt.get('background', False)
-        self.attn_bg = opt.get('attn_background', False)
 
         print("Building Seq2Seq with Copy model ...")
         self.enc_hidden_dim = self.hidden_dim // 2
@@ -56,8 +55,7 @@ def __init__(self, opt, emb_matrix=None):
             self.bg_encoder = nn.LSTM(self.emb_dim, self.enc_hidden_dim, 1, \
                     bidirectional=True, batch_first=True, dropout=0) # when nlayer=1, rnn dropout does not apply
             self.bg_drop = nn.Dropout(self.dropout)
-            if self.attn_bg:
-                self.bg_attn_layer = BasicAttention(self.hidden_dim)
+            self.bg_attn_layer = BasicAttention(self.hidden_dim)
 
         self.SOS_tensor = torch.LongTensor([constant.SOS_ID])
         self.SOS_tensor = self.SOS_tensor.cuda() if self.use_cuda else self.SOS_tensor
@@ -205,9 +203,8 @@ def forward(self, src, tgt_in, bg):
             bg_lens = bg_mask.eq(0).long().sum(1)
             bg_out, bg_hidden = self.encode(bg_inputs, bg_lens, encoder=self.bg_encoder, sort=True)
             bg_h = bg_hidden[0]
-            if self.attn_bg:
-                # use attentional representation
-                _, bg_h, _ = self.bg_attn_layer(enc_hidden[0], bg_out, mask=bg_mask)
+            # use attentional representation
+            _, bg_h, _ = self.bg_attn_layer(enc_hidden[0], bg_out, mask=bg_mask)
             bg_h = self.bg_drop(bg_h)
         else:
             bg_h = None
@@ -249,9 +246,8 @@ def predict(self, src, bg, beam_size=5):
             bg_lens = bg_mask.eq(0).long().sum(1)
             bg_out, bg_hidden = self.encode(bg_inputs, bg_lens, encoder=self.bg_encoder, sort=True)
             bg_h = bg_hidden[0]
-            if self.attn_bg:
-                # use attentional representation
-                _, bg_h, _ = self.bg_attn_layer(enc_hidden[0], bg_out, mask=bg_mask)
+            # use attentional representation
+            _, bg_h, _ = self.bg_attn_layer(enc_hidden[0], bg_out, mask=bg_mask)
             bg_h = self.bg_drop(bg_h)
         else:
             bg_h = None

train.py

@@ -41,7 +41,6 @@
 parser.add_argument('--cov_loss_epoch', type=int, default=0, help='Add coverage loss starting from this epoch.')
 
 parser.add_argument('--background', action='store_true', help='Use background information for decoder.')
-parser.add_argument('--attn_background', action='store_true', help='Use attentional background encoder.')
 parser.add_argument('--concat_background', action='store_true', help='Simply concat background to findings.')
 
 parser.add_argument('--use_bleu', action='store_true', help='Use BLEU as the metric. By default use ROUGE.')