Merge branch 'master' of https://github.com/r9y9/deepvoice3_pytorch

Author: r9y9

deepvoice3_pytorch/deepvoice3.py

@@ -305,13 +305,13 @@ def forward(self, encoder_out, inputs=None,
             w = self.key_position_rate
             # TODO: may be useful to have projection per attention layer
             if self.speaker_proj1 is not None:
-                w = w * F.sigmoid(self.speaker_proj1(speaker_embed)).view(-1)
+                w = w * torch.sigmoid(self.speaker_proj1(speaker_embed)).view(-1)
             text_pos_embed = self.embed_keys_positions(text_positions, w)
             keys = keys + text_pos_embed
         if frame_positions is not None:
             w = self.query_position_rate
             if self.speaker_proj2 is not None:
-                w = w * F.sigmoid(self.speaker_proj2(speaker_embed)).view(-1)
+                w = w * torch.sigmoid(self.speaker_proj2(speaker_embed)).view(-1)
             frame_pos_embed = self.embed_query_positions(frame_positions, w)
 
         # transpose only once to speed up attention layers
@@ -357,10 +357,10 @@ def forward(self, encoder_out, inputs=None,
         x = x.transpose(1, 2)
 
         # project to mel-spectorgram
-        outputs = F.sigmoid(x)
+        outputs = torch.sigmoid(x)
 
         # Done flag
-        done = F.sigmoid(self.fc(x))
+        done = torch.sigmoid(self.fc(x))
 
         return outputs, torch.stack(alignments), done, decoder_states
 
@@ -373,7 +373,7 @@ def incremental_forward(self, encoder_out, text_positions, speaker_embed=None,
         w = self.key_position_rate
         # TODO: may be useful to have projection per attention layer
         if self.speaker_proj1 is not None:
-            w = w * F.sigmoid(self.speaker_proj1(speaker_embed)).view(-1)
+            w = w * torch.sigmoid(self.speaker_proj1(speaker_embed)).view(-1)
         text_pos_embed = self.embed_keys_positions(text_positions, w)
         keys = keys + text_pos_embed
 
@@ -399,7 +399,7 @@ def incremental_forward(self, encoder_out, text_positions, speaker_embed=None,
             frame_pos = keys.data.new(B, 1).fill_(t + 1).long()
             w = self.query_position_rate
             if self.speaker_proj2 is not None:
-                w = w * F.sigmoid(self.speaker_proj2(speaker_embed)).view(-1)
+                w = w * torch.sigmoid(self.speaker_proj2(speaker_embed)).view(-1)
             frame_pos_embed = self.embed_query_positions(frame_pos, w)
 
             if test_inputs is not None:
@@ -457,8 +457,8 @@ def incremental_forward(self, encoder_out, text_positions, speaker_embed=None,
             ave_alignment = ave_alignment.div_(num_attention_layers)
 
             # Ooutput & done flag predictions
-            output = F.sigmoid(x)
-            done = F.sigmoid(self.fc(x))
+            output = torch.sigmoid(x)
+            done = torch.sigmoid(self.fc(x))
 
             decoder_states += [decoder_state]
             outputs += [output]
@@ -601,4 +601,4 @@ def forward(self, x, speaker_embed=None):
         # Back to B x T x C
         x = x.transpose(1, 2)
 
-        return F.sigmoid(x)
+        return torch.sigmoid(x)

deepvoice3_pytorch/modules.py

@@ -160,7 +160,7 @@ def _forward(self, x, speaker_embed, is_incremental):
             # Since conv layer assumes BCT, we need to transpose
             softsign = softsign if is_incremental else softsign.transpose(1, 2)
             a = a + softsign
-        x = a * F.sigmoid(b)
+        x = a * torch.sigmoid(b)
         return (x + residual) * math.sqrt(0.5) if self.residual else x
 
     def clear_buffer(self):
@@ -222,7 +222,7 @@ def _forward(self, x, is_incremental):
             return (x + residual) * math.sqrt(0.5)
         else:
             a, b = x.split(x.size(splitdim) // 2, dim=splitdim)
-            T = F.sigmoid(b)
+            T = torch.sigmoid(b)
             return (T * a + (1 - T) * residual)
 
     def clear_buffer(self):

deepvoice3_pytorch/nyanko.py

@@ -237,10 +237,10 @@ def forward(self, encoder_out, inputs=None,
         x = x.transpose(1, 2)
 
         # Mel
-        outputs = F.sigmoid(x)
+        outputs = torch.sigmoid(x)
 
         # Done prediction
-        done = F.sigmoid(self.fc(x))
+        done = torch.sigmoid(self.fc(x))
 
         # Adding extra dim for convenient
         alignments = alignments.unsqueeze(0)
@@ -310,8 +310,8 @@ def incremental_forward(self, encoder_out, text_positions,
             x = self.last_conv.incremental_forward(x)
 
             # Ooutput & done flag predictions
-            output = F.sigmoid(x)
-            done = F.sigmoid(self.fc(x))
+            output = torch.sigmoid(x)
+            done = torch.sigmoid(self.fc(x))
 
             decoder_states += [decoder_state]
             outputs += [output]