Merge pull request #23 from NVIDIA/attention_full_mel

model.py: attending to full mel instead of prenet and dropout mel

Author: rafaelvalle

model.py

@@ -221,7 +221,7 @@ def __init__(self, hparams):
             [hparams.prenet_dim, hparams.prenet_dim])
 
         self.attention_rnn = nn.LSTMCell(
-            hparams.prenet_dim + hparams.encoder_embedding_dim,
+            hparams.decoder_rnn_dim + hparams.encoder_embedding_dim,
             hparams.attention_rnn_dim)
 
         self.attention_layer = Attention(
@@ -230,7 +230,7 @@ def __init__(self, hparams):
             hparams.attention_location_kernel_size)
 
         self.decoder_rnn = nn.LSTMCell(
-            hparams.attention_rnn_dim + hparams.encoder_embedding_dim,
+            hparams.prenet_dim + hparams.encoder_embedding_dim,
             hparams.decoder_rnn_dim, 1)
 
         self.linear_projection = LinearNorm(
@@ -351,8 +351,7 @@ def decode(self, decoder_input):
         attention_weights:
         """
 
-        decoder_input = self.prenet(decoder_input)
-        cell_input = torch.cat((decoder_input, self.attention_context), -1)
+        cell_input = torch.cat((self.decoder_hidden, self.attention_context), -1)
         self.attention_hidden, self.attention_cell = self.attention_rnn(
             cell_input, (self.attention_hidden, self.attention_cell))
 
@@ -364,8 +363,8 @@ def decode(self, decoder_input):
             attention_weights_cat, self.mask)
 
         self.attention_weights_cum += self.attention_weights
-        decoder_input = torch.cat(
-            (self.attention_hidden, self.attention_context), -1)
+        prenet_output = self.prenet(decoder_input)
+        decoder_input = torch.cat((prenet_output, self.attention_context), -1)
         self.decoder_hidden, self.decoder_cell = self.decoder_rnn(
             decoder_input, (self.decoder_hidden, self.decoder_cell))