coqui-ai
/
TTS
mirror of https://github.com/coqui-ai/TTS.git
1
0
Fork 0
Code Issues Packages Projects Releases Wiki Activity

Padding with funcitonal interface to match TF "SAME"

pull/10/head
Eren 2018-09-17 20:19:09 +02:00
parent 4a741c64b3
commit 00c0c9cde6
1 changed files with 9 additions and 7 deletions
Show Stats Download Patch File Download Diff File Expand all files Collapse all files

16
layers/tacotron.py
View File

@ -56,18 +56,20 @@ class BatchNormConv1d(nn.Module):
padding, padding,
activation=None): activation=None):
super(BatchNormConv1d, self).__init__() super(BatchNormConv1d, self).__init__()
self.padding = padding
self.conv1d = nn.Conv1d( self.conv1d = nn.Conv1d(
in_channels, in_channels,
out_channels, out_channels,
kernel_size=kernel_size, kernel_size=kernel_size,
stride=stride, stride=stride,
padding=padding, padding=0,
bias=False) bias=False)
# Following tensorflow's default parameters # Following tensorflow's default parameters
self.bn = nn.BatchNorm1d(out_channels, momentum=0.99, eps=1e-3) self.bn = nn.BatchNorm1d(out_channels, momentum=0.99, eps=1e-3)
self.activation = activation self.activation = activation
def forward(self, x): def forward(self, x):
x = nn.functional.pad(x, self.padding)
x = self.conv1d(x) x = self.conv1d(x)
if self.activation is not None: if self.activation is not None:
x = self.activation(x) x = self.activation(x)
@ -130,12 +132,12 @@ class CBHG(nn.Module):
conv_bank_features, conv_bank_features,
kernel_size=k, kernel_size=k,
stride=1, stride=1,
padding=k // 2, padding=[(k - 1) // 2, k // 2],
activation=self.relu) for k in range(1, K + 1) 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 # 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] out_features = [K * conv_bank_features] + conv_projections[:-1]
activations = [self.relu] * (len(conv_projections) - 1) activations = [self.relu] * (len(conv_projections) - 1)
activations += [None] activations += [None]
@ -148,7 +150,7 @@ class CBHG(nn.Module):
out_size, out_size,
kernel_size=3, kernel_size=3,
stride=1, stride=1,
padding=1, padding=[1, 1],
activation=ac) activation=ac)
layer_set.append(layer) layer_set.append(layer)
self.conv1d_projections = nn.ModuleList(layer_set) self.conv1d_projections = nn.ModuleList(layer_set)
@ -181,11 +183,11 @@ class CBHG(nn.Module):
outs = [] outs = []
for conv1d in self.conv1d_banks: for conv1d in self.conv1d_banks:
out = conv1d(x) out = conv1d(x)
out = out[:, :, :T]
outs.append(out) outs.append(out)
x = torch.cat(outs, dim=1) x = torch.cat(outs, dim=1)
assert x.size(1) == self.conv_bank_features * len(self.conv1d_banks) 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: for conv1d in self.conv1d_projections:
x = conv1d(x) x = conv1d(x)
# (B, T_in, hid_feature) # (B, T_in, hid_feature)