def Conv1d(in_channels,
out_channels,
kernel_size=1,
padding=0,
dilation=1,
bias=True,
weight_normalization=True,
dropout=0,
std_mul=1.0,
**kwargs):
m = conv.Conv1d(in_channels,
out_channels,
kernel_size=kernel_size,
padding=padding,
dilation=dilation,
bias=bias,
**kwargs)
if weight_normalization:
assert bias
std = math.sqrt(
(std_mul * (1.0 - dropout)) / (m.kernel_size[0] * in_channels))
m.weight.data.normal_(mean=0, std=std)
m.bias.data.zero_()
return nn.utils.weight_norm(m)
else:
return m
def Conv1d(in_channels,
out_channels,
kernel_size,
dropout=0,
std_mul=4.0,
**kwargs):
m = conv.Conv1d(in_channels, out_channels, kernel_size, **kwargs)
std = math.sqrt(
(std_mul * (1.0 - dropout)) / (m.kernel_size[0] * in_channels))
m.weight.data.normal_(mean=0, std=std)
m.bias.data.zero_()
return nn.utils.weight_norm(m)
def Conv1d1x1(in_channels, out_channels, bias=True, weight_normalization=True):
"""1-by-1 convolution layer
"""
if weight_normalization:
assert bias
return Conv1d(in_channels,
out_channels,
kernel_size=1,
padding=0,
dilation=1,
bias=bias,
std_mul=1.0)
else:
return conv.Conv1d(in_channels,
out_channels,
kernel_size=1,
padding=0,
dilation=1,
bias=bias)
def __init__(self,
residual_channels,
gate_channels,
kernel_size,
skip_out_channels=None,
cin_channels=-1,
gin_channels=-1,
dropout=1 - 0.95,
padding=None,
dilation=1,
causal=True,
bias=True,
weight_normalization=True,
*args,
**kwargs):
super(ResidualConv1dGLU, self).__init__()
self.dropout = dropout
if skip_out_channels is None:
skip_out_channels = residual_channels
if padding is None:
# no future time stamps available
if causal:
padding = (kernel_size - 1) * dilation
else:
padding = (kernel_size - 1) // 2 * dilation
self.causal = causal
if weight_normalization:
assert bias
self.conv = Conv1d(residual_channels,
gate_channels,
kernel_size,
padding=padding,
dilation=dilation,
bias=bias,
std_mul=1.0,
*args,
**kwargs)
else:
self.conv = conv.Conv1d(residual_channels,
gate_channels,
kernel_size,
padding=padding,
dilation=dilation,
bias=bias,
*args,
**kwargs)
# local conditioning
if cin_channels > 0:
self.conv1x1c = Conv1d1x1(
cin_channels,
gate_channels,
bias=bias,
weight_normalization=weight_normalization)
else:
self.conv1x1c = None
# global conditioning
if gin_channels > 0:
self.conv1x1g = Conv1d1x1(
gin_channels,
gate_channels,
bias=bias,
weight_normalization=weight_normalization)
else:
self.conv1x1g = None
# conv output is split into two groups
gate_out_channels = gate_channels // 2
self.conv1x1_out = Conv1d1x1(gate_out_channels,
residual_channels,
bias=bias,
weight_normalization=weight_normalization)
self.conv1x1_skip = Conv1d1x1(
gate_out_channels,
skip_out_channels,
bias=bias,
weight_normalization=weight_normalization)
self.tanh = nn.Tanh()
self.sigmoid = nn.Sigmoid()
def Conv1d(in_channels, out_channels, kernel_size, dropout=0, **kwargs):
m = conv.Conv1d(in_channels, out_channels, kernel_size, **kwargs)
nn.init.kaiming_normal_(m.weight, nonlinearity="relu")
if m.bias is not None:
nn.init.constant_(m.bias, 0)
return nn.utils.weight_norm(m)
def Conv1d(in_channels, out_channels, kernel_size, dropout=0, **kwargs):
m = conv.Conv1d(in_channels, out_channels, kernel_size, **kwargs)
return m
