Attn masking

layers/attention.py

@@ -1,6 +1,7 @@
 import torch
 from torch import nn
 from torch.nn import functional as F
+from utils.generic_utils import sequence_mask
 
 
 class BahdanauAttention(nn.Module):
@@ -91,8 +92,17 @@ class AttentionRNNCell(nn.Module):
                 'b' (Bahdanau) or 'ls' (Location Sensitive).".format(align_model))
 
 
-    def forward(self, memory, context, rnn_state, annotations,
-                attention_vec, mask=None, annotations_lengths=None):
+    def forward(self, memory, context, rnn_state, annots,
+                atten, annot_lens=None):
+        """
+        Shapes:
+            - memory: (batch, 1, dim) or (batch, dim)
+            - context: (batch, dim)
+            - rnn_state: (batch, out_dim)
+            - annots: (batch, max_time, annot_dim)
+            - atten: (batch, max_time)
+            - annot_lens: (batch,)
+        """
         # Concat input query and previous context context
         rnn_input = torch.cat((memory, context), -1)
         # Feed it to RNN
@@ -102,18 +112,18 @@ class AttentionRNNCell(nn.Module):
         # (batch, max_time)
         # e_{ij} = a(s_{i-1}, h_j)
         if self.align_model is 'b':
-            alignment = self.alignment_model(annotations, rnn_output)
+            alignment = self.alignment_model(annots, rnn_output)
         else:
-            alignment = self.alignment_model(annotations, rnn_output, attention_vec)
-        # TODO: needs recheck.
-        if mask is not None:
-            mask = mask.view(query.size(0), -1)
-            alignment.data.masked_fill_(mask, self.score_mask_value)
+            alignment = self.alignment_model(annots, rnn_output, atten)
+        if annot_lens is not None:
+            mask = sequence_mask(annot_lens)
+            mask = mask.view(memory.size(0), -1)
+            alignment.masked_fill_(1 - mask, -float("inf"))
         # Normalize context weight
         alignment = F.softmax(alignment, dim=-1)
         # Attention context vector
         # (batch, 1, dim)
         # c_i = \sum_{j=1}^{T_x} \alpha_{ij} h_j
-        context = torch.bmm(alignment.unsqueeze(1), annotations)
+        context = torch.bmm(alignment.unsqueeze(1), annots)
         context = context.squeeze(1)
         return rnn_output, context, alignment

layers/losses.py

@@ -1,26 +1,53 @@
 import torch
 from torch.nn import functional
 from torch import nn
+from utils.generic_utils import sequence_mask
 
 
-# from https://gist.github.com/jihunchoi/f1434a77df9db1bb337417854b398df1
-def _sequence_mask(sequence_length, max_len=None):
-    if max_len is None:
-        max_len = sequence_length.data.max()
-    batch_size = sequence_length.size(0)
-    seq_range = torch.arange(0, max_len).long()
-    seq_range_expand = seq_range.unsqueeze(0).expand(batch_size, max_len)
-    if sequence_length.is_cuda:
-        seq_range_expand = seq_range_expand.cuda()
-    seq_length_expand = (sequence_length.unsqueeze(1)
-                         .expand_as(seq_range_expand))
-    return seq_range_expand < seq_length_expand
-
-
-class L2LossMasked(nn.Module):
+class L1LossMasked(nn.Module):
 
     def __init__(self):
-        super(L2LossMasked, self).__init__()
+        super(L1LossMasked, self).__init__()
+
+    def forward(self, input, target, length):
+        """
+        Args:
+            input: A Variable containing a FloatTensor of size
+                (batch, max_len, dim) which contains the
+                unnormalized probability for each class.
+            target: A Variable containing a LongTensor of size
+                (batch, max_len, dim) which contains the index of the true
+                class for each corresponding step.
+            length: A Variable containing a LongTensor of size (batch,)
+                which contains the length of each data in a batch.
+        Returns:
+            loss: An average loss value masked by the length.
+        """
+        input = input.contiguous()
+        target = target.contiguous()
+
+        # logits_flat: (batch * max_len, dim)
+        input = input.view(-1, input.shape[-1])
+        # target_flat: (batch * max_len, dim)
+        target_flat = target.view(-1, target.shape[-1])
+        # losses_flat: (batch * max_len, dim)
+        losses_flat = functional.l1_loss(input, target_flat, size_average=False,
+                                         reduce=False)
+        # losses: (batch, max_len, dim)
+        losses = losses_flat.view(*target.size())
+
+        # mask: (batch, max_len, 1)
+        mask = sequence_mask(sequence_length=length,
+                              max_len=target.size(1)).unsqueeze(2)
+        losses = losses * mask.float()
+        loss = losses.sum() / (length.float().sum() * float(target.shape[2]))
+        return loss
+
+
+class MSELossMasked(nn.Module):
+
+    def __init__(self):
+        super(MSELossMasked, self).__init__()
 
     def forward(self, input, target, length):
         """
@@ -48,9 +75,9 @@ class L2LossMasked(nn.Module):
                                          reduce=False)
         # losses: (batch, max_len, dim)
         losses = losses_flat.view(*target.size())
-        
+
         # mask: (batch, max_len, 1)
-        mask = _sequence_mask(sequence_length=length,
+        mask = sequence_mask(sequence_length=length,
                               max_len=target.size(1)).unsqueeze(2)
         losses = losses * mask.float()
         loss = losses.sum() / (length.float().sum() * float(target.shape[2]))

layers/tacotron.py

@@ -213,7 +213,7 @@ class Decoder(nn.Module):
         self.proj_to_mel = nn.Linear(256, memory_dim * r)
         self.stopnet = StopNet(r, memory_dim)
 
-    def forward(self, inputs, memory=None):
+    def forward(self, inputs, memory=None, input_lens=None):
         """
         Decoder forward step.
 
@@ -225,6 +225,7 @@ class Decoder(nn.Module):
             memory (None): Decoder memory (autoregression. If None (at eval-time),
               decoder outputs are used as decoder inputs. If None, it uses the last
               output as the input.
+            input_lens (None): Time length of each input in batch.
 
         Shapes:
             - inputs: batch x time x encoder_out_dim
@@ -273,7 +274,8 @@ class Decoder(nn.Module):
                                        # attention_cum.unsqueeze(1)),
                                       # dim=1)
             attention_rnn_hidden, current_context_vec, attention = self.attention_rnn(
-                processed_memory, current_context_vec, attention_rnn_hidden, inputs, attention)
+                processed_memory, current_context_vec, attention_rnn_hidden,
+                inputs, attention, input_lens)
             # attention_cum += attention
             # Concat RNN output and attention context vector
             decoder_input = self.project_to_decoder_in(