forked from maum-ai/assem-vc
-
Notifications
You must be signed in to change notification settings - Fork 0
/
f0_encoder.py
32 lines (28 loc) · 1.21 KB
/
f0_encoder.py
1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
import torch.nn as nn
class ConvNorm(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size=1, stride=1,
padding=None, dilation=1, bias=True, w_init_gain='linear'):
super(ConvNorm, self).__init__()
if padding is None:
assert(kernel_size % 2 == 1)
padding = int(dilation * (kernel_size - 1) / 2)
self.conv = nn.Conv1d(in_channels, out_channels,
kernel_size=kernel_size, stride=stride,
padding=padding, dilation=dilation,
bias=bias)
nn.init.xavier_uniform_(
self.conv.weight, gain=nn.init.calculate_gain(w_init_gain))
def forward(self, signal):
conv_signal = self.conv(signal)
return conv_signal
class F0_Encoder(nn.Module):
def __init__(self, hp):
super().__init__()
self.prenet_f0 = ConvNorm(
1, hp.chn.prenet_f0,
kernel_size=hp.ker.prenet_f0,
padding=max(0, int(hp.ker.prenet_f0 / 2)),
bias=True, stride=1, dilation=1)
def forward(self, f0s):
f0s = self.prenet_f0(f0s)
return f0s