From 00c0c9cde637a6cd7e3938a77ff833db1440ebca Mon Sep 17 00:00:00 2001
From: Eren <erengolge@gmail.com>
Date: Mon, 17 Sep 2018 20:19:09 +0200
Subject: [PATCH] Padding with funcitonal interface to match TF "SAME"

---
 layers/tacotron.py | 16 +++++++++-------
 1 file changed, 9 insertions(+), 7 deletions(-)

diff --git a/layers/tacotron.py b/layers/tacotron.py
index 336f2d9e..41bd15d2 100644
--- a/layers/tacotron.py
+++ b/layers/tacotron.py
@@ -56,18 +56,20 @@ class BatchNormConv1d(nn.Module):
                  padding,
                  activation=None):
         super(BatchNormConv1d, self).__init__()
+        self.padding = padding
         self.conv1d = nn.Conv1d(
             in_channels,
             out_channels,
             kernel_size=kernel_size,
             stride=stride,
-            padding=padding,
+            padding=0,
             bias=False)
         # Following tensorflow's default parameters
         self.bn = nn.BatchNorm1d(out_channels, momentum=0.99, eps=1e-3)
         self.activation = activation
 
     def forward(self, x):
+        x = nn.functional.pad(x, self.padding)
         x = self.conv1d(x)
         if self.activation is not None:
             x = self.activation(x)
@@ -130,12 +132,12 @@ class CBHG(nn.Module):
                 conv_bank_features,
                 kernel_size=k,
                 stride=1,
-                padding=k // 2,
+                padding=[(k - 1) // 2, k // 2],
                 activation=self.relu) for k in range(1, K + 1)
         ])
-        # max pooling of conv bank
+        # max pooling of conv bank, padding with nn.functional 
         # TODO: try average pooling OR larger kernel size
-        self.max_pool1d = nn.MaxPool1d(kernel_size=2, stride=1, padding=1)
+        self.max_pool1d = nn.MaxPool1d(kernel_size=2, stride=1, padding=0)
         out_features = [K * conv_bank_features] + conv_projections[:-1]
         activations = [self.relu] * (len(conv_projections) - 1)
         activations += [None]
@@ -148,7 +150,7 @@ class CBHG(nn.Module):
                 out_size,
                 kernel_size=3,
                 stride=1,
-                padding=1,
+                padding=[1, 1],
                 activation=ac)
             layer_set.append(layer)
         self.conv1d_projections = nn.ModuleList(layer_set)
@@ -181,11 +183,11 @@ class CBHG(nn.Module):
         outs = []
         for conv1d in self.conv1d_banks:
             out = conv1d(x)
-            out = out[:, :, :T]
             outs.append(out)
         x = torch.cat(outs, dim=1)
         assert x.size(1) == self.conv_bank_features * len(self.conv1d_banks)
-        x = self.max_pool1d(x)[:, :, :T]
+        x = nn.functional.pad(x, [0, 1])
+        x = self.max_pool1d(x)
         for conv1d in self.conv1d_projections:
             x = conv1d(x)
         # (B, T_in, hid_feature)