change the bitwise for masking and small fixes

layers/common_layers.py

@@ -234,7 +234,7 @@ class Attention(nn.Module):
                 query, processed_inputs)
         # apply masking
         if mask is not None:
-            attention.data.masked_fill_(torch.bitwise_not(mask), self._mask_value)
+            attention.data.masked_fill_(~mask, self._mask_value)
         # apply windowing - only in eval mode
         if not self.training and self.windowing:
             attention = self.apply_windowing(attention, inputs)

layers/tacotron.py

@@ -315,7 +315,7 @@ class Decoder(nn.Module):
         # learn init values instead of zero init.
         self.stopnet = StopNet(256 + memory_dim * self.r_init)
 
-    def _set_r(self, new_r):
+    def set_r(self, new_r):
         self.r = new_r
 
     def _reshape_memory(self, memory):