def __init__(self):
super(ConditionalDiscriminatorBlock, self).__init__()
self.c_net = nn.Sequential(
# (N, 80, Tmel)
ll.CustomConv1d(80,
256,
kernel_size=1,
stride=1,
padding='same',
lrelu=True),
# (N, 256, Tmel)
)
self.net = nn.ModuleList([
# (N, n_mags, Tmel*4)
ll.CustomConv1d(args.n_mags,
64,
kernel_size=5,
stride=1,
padding='same',
lrelu=True),
mm.ResidualBlock1d(64, 128),
mm.ResidualBlock1d(128, 256),
nn.AvgPool1d(3, 2, padding=1), # (N, 256, Tmel*2)
mm.ResidualBlock1d(256, 256),
nn.AvgPool1d(3, 2, padding=1), # (N, 256, Tmel)
mm.ResidualBlock1d(256, 256)
])
self.postnet = nn.ModuleList([
mm.ResidualBlock1d(256, 512),
nn.AvgPool1d(3, 2, padding=1), # (N, 256, Tmel//2)
mm.ResidualBlock1d(512, 512),
nn.AvgPool1d(3, 2, padding=1), # (N, 256, Tmel//4)
mm.ResidualBlock1d(512, 256),
mm.ResidualBlock1d(256, 1)
])
def __init__(self):
super(ConditionalDiscriminatorBlock, self).__init__()
self.c_net = nn.Sequential(
# (N, 80, Tmel)
ll.CustomConv1d(80,
256,
kernel_size=1,
stride=1,
padding='same',
lrelu=True),
# (N, 256, Tmel)
)
self.net = nn.ModuleList([
# (N, n_mags, Tmel*4)
ll.CustomConv1d(args.n_mags,
64,
kernel_size=3,
stride=1,
padding='same',
lrelu=True),
# (N, 16, Tmel*4)
ll.CustomConv1d(64,
128,
kernel_size=5,
stride=2,
padding='same',
lrelu=True),
# (N, 64, Tmel*2)
ll.CustomConv1d(128,
256,
kernel_size=5,
stride=2,
padding='same',
lrelu=True),
# (N, 256, Tmel)
])
self.postnet = nn.ModuleList([
ll.CustomConv1d(256,
256,
kernel_size=3,
stride=1,
padding='same',
lrelu=True),
ll.CustomConv1d(256,
128,
kernel_size=3,
stride=1,
padding='same',
lrelu=True),
ll.CustomConv1d(128,
1,
kernel_size=3,
stride=1,
padding='same',
lrelu=False),
])
def __init__(self, in_planes, planes, ksize=3):
super(ResidualBlock1d, self).__init__()
self.conv1 = ll.CustomConv1d(in_planes, planes, ksize, lrelu=True)
self.conv2 = ll.CustomConv1d(planes, planes, ksize, lrelu=True)
self.proj = nn.Conv1d(in_planes, planes,
1) if in_planes != planes else None
def __init__(self, d_hidden):
super(DotProductAttention, self).__init__()
self.d_k = d_hidden
self.linear_q = ll.CustomConv1d(d_hidden, d_hidden, 1)
self.linear_k = ll.CustomConv1d(d_hidden, d_hidden, 1)
self.linear_v = ll.CustomConv1d(d_hidden, d_hidden, 1)