change order BN + ReLU to ReLU + BN for SS

TTS/tts/layers/generic/res_conv_bn.py

@@ -27,8 +27,8 @@ class ConvBN(nn.Module):
     def forward(self, x):
         o = self.conv1d(x)
         o = self.pad(o)
-        o = self.norm(o)
         o = nn.functional.relu(o)
+        o = self.norm(o)
         return o
 
 

TTS/tts/layers/speedy_speech/decoder.py

@@ -25,7 +25,7 @@ class Decoder(nn.Module):
 
         self.post_conv = nn.Conv1d(hidden_channels, hidden_channels, 1)
         self.post_net = nn.Sequential(
-            ConvBNBlock(hidden_channels, 4, 1, num_conv_blocks=2),
+            ConvBNBlock(hidden_channels, residual_conv_bn_params['kernel_size'], 1, num_conv_blocks=2),
             nn.Conv1d(hidden_channels, out_channels, 1),
         )
 
@@ -33,4 +33,4 @@ class Decoder(nn.Module):
         # TODO: implement multi-speaker
         o = self.decoder(x, x_mask)
         o = self.post_conv(o) + x
-        return self.post_net(o)
+        return self.post_net(o) * x_mask

TTS/tts/layers/speedy_speech/encoder.py

@@ -154,8 +154,8 @@ class Encoder(nn.Module):
             o = self.pre(x) * x_mask
         o = self.encoder(o, x_mask)
         o = self.post_conv(o + x)
-        o = self.post_bn(o)
         o = F.relu(o)
+        o = self.post_bn(o)
         o = self.post_conv2(o)
         # [B, C, T]
         return o * x_mask