Constant queue size for autoregression window

layers/tacotron.py

@@ -309,7 +309,7 @@ class Decoder(nn.Module):
         self.memory_size = memory_size if memory_size > 0 else r
         self.memory_dim = memory_dim
         # memory -> |Prenet| -> processed_memory
-        self.prenet = Prenet(memory_dim * memory_dim * self.memory_size, out_features=[256, 128])
+        self.prenet = Prenet(memory_dim * self.memory_size, out_features=[256, 128])
         # processed_inputs, processed_memory -> |Attention| -> Attention, attention, RNN_State
         self.attention_rnn = AttentionRNNCell(
             out_dim=128,

models/tacotron.py

@@ -13,6 +13,7 @@ class Tacotron(nn.Module):
                  mel_dim=80,
                  r=5,
                  padding_idx=None, 
+                 memory_size=5,
                  attn_windowing=False):
         super(Tacotron, self).__init__()
         self.r = r
@@ -23,7 +24,7 @@ class Tacotron(nn.Module):
         print(" | > Number of characters : {}".format(num_chars))
         self.embedding.weight.data.normal_(0, 0.3)
         self.encoder = Encoder(embedding_dim)
-        self.decoder = Decoder(256, mel_dim, r, attn_windowing)
+        self.decoder = Decoder(256, mel_dim, r, memory_size, attn_windowing)
         self.postnet = PostCBHG(mel_dim)
         self.last_linear = nn.Sequential(
             nn.Linear(self.postnet.cbhg.gru_features * 2, linear_dim),

train.py

@@ -357,7 +357,7 @@ def evaluate(model, criterion, criterion_st, ap, current_step):
 
 def main(args):
     num_chars = len(phonemes) if c.use_phonemes else len(symbols)
-    model = Tacotron(num_chars, c.embedding_size, ap.num_freq, ap.num_mels, c.r)
+    model = Tacotron(num_chars, c.embedding_size, ap.num_freq, ap.num_mels, c.r, c.memory_size)
     print(" | > Num output units : {}".format(ap.num_freq), flush=True)
 
     optimizer = optim.Adam(model.parameters(), lr=c.lr, weight_decay=0)