use memory queue if r is smaller than queue size

layers/tacotron.py

@@ -395,7 +395,7 @@ class Decoder(nn.Module):
         return output, stop_token, self.attention_layer.attention_weights
 
     def _update_memory_queue(self, new_memory):
-        if self.memory_size > 0:
+        if self.memory_size > 0 and new_memory.shape[-1] < self.memory_size:
             self.memory_input = torch.cat([
                 self.memory_input[:, self.r * self.memory_dim:].clone(),
                 new_memory