def __init__(self, in_channels, mid_channels, out_channels, num_blocks,
kernel_size, padding, double_after_norm):
super(ResNet, self).__init__()
self.in_norm = nn.BatchNorm2d(in_channels)
self.double_after_norm = double_after_norm
self.in_conv = WNConv2d(2 * in_channels,
mid_channels,
kernel_size,
padding,
bias=True)
self.in_skip = WNConv2d(mid_channels,
mid_channels,
kernel_size=1,
padding=0,
bias=True)
self.blocks = nn.ModuleList([
ResidualBlock(mid_channels, mid_channels)
for _ in range(num_blocks)
])
self.skips = nn.ModuleList([
WNConv2d(mid_channels,
mid_channels,
kernel_size=1,
padding=0,
bias=True) for _ in range(num_blocks)
])
self.out_norm = nn.BatchNorm2d(mid_channels)
self.out_conv = WNConv2d(mid_channels,
out_channels,
kernel_size=1,
padding=0,
bias=True)
def __init__(self, in_channels, mid_channels, out_channels, num_blocks,
kernel_size, padding):
super(ResNet, self).__init__()
self.in_pad = nn.ReflectionPad2d(padding)
self.in_conv = WNConv2d(in_channels,
mid_channels,
kernel_size,
padding=0,
bias=True)
self.in_skip = WNConv2d(mid_channels,
mid_channels,
kernel_size=1,
padding=0,
bias=True)
self.blocks = nn.ModuleList([
ResidualBlock(mid_channels, mid_channels)
for _ in range(num_blocks)
])
self.skips = nn.ModuleList([
WNConv2d(mid_channels,
mid_channels,
kernel_size=1,
padding=0,
bias=True) for _ in range(num_blocks)
])
self.out_norm = nn.BatchNorm2d(mid_channels)
self.out_conv = WNConv2d(mid_channels,
out_channels,
kernel_size=1,
padding=0,
bias=True)
def __init__(self,
in_channels,
num_channels,
num_blocks,
num_components,
drop_prob,
use_attn=True,
aux_channels=None,
condition_embd_size=0):
super(NN, self).__init__()
self.k = num_components # k = number of mixture components
self.in_conv = WNConv2d(in_channels,
num_channels,
kernel_size=3,
padding=1)
self.mid_convs = nn.ModuleList([
ConvAttnBlock(num_channels, drop_prob, use_attn, aux_channels)
for _ in range(num_blocks)
])
self.out_conv = WNConv2d(num_channels,
in_channels * (2 + 3 * self.k),
kernel_size=3,
padding=1)
self.rescale = weight_norm(Rescale(in_channels))
if condition_embd_size > 0:
self.conditioning_projection = nn.Linear(condition_embd_size,
2 * in_channels *
(2 + 3 * self.k),
bias=False)
def __init__(self, in_channels, num_channels, num_blocks, num_components, drop_prob, use_attn=True, aux_channels=None):
super(NN, self).__init__()
self.k = num_components # k = number of mixture components
self.in_conv = WNConv2d(in_channels, num_channels, kernel_size=3, padding=1)
self.mid_convs = nn.ModuleList([ConvAttnBlock(num_channels, drop_prob, use_attn, aux_channels)
for _ in range(num_blocks)])
self.out_conv = WNConv2d(num_channels, in_channels * (2 + 3 * self.k),
kernel_size=3, padding=1)
self.rescale = weight_norm(Rescale(in_channels))
def __init__(self, num_channels, drop_prob=0., aux_channels=None):
super(GatedConv, self).__init__()
self.nlin = concat_elu
self.conv = WNConv2d(2 * num_channels, num_channels, kernel_size=3, padding=1)
self.drop = nn.Dropout2d(drop_prob)
self.gate = WNConv2d(2 * num_channels, 2 * num_channels, kernel_size=1, padding=0)
if aux_channels is not None:
self.aux_conv = WNConv2d(2 * aux_channels, num_channels, kernel_size=1, padding=0)
else:
self.aux_conv = None
def __init__(self, in_channels, mid_channels, out_channels, num_blocks,
drop_prob):
super(NN, self).__init__()
conv_blocks = [
WNConv2d(in_channels, mid_channels, kernel_size=3, padding=1)
]
conv_blocks += [
ConvBlock(mid_channels, drop_prob) for _ in range(num_blocks)
]
conv_blocks += [
WNConv2d(mid_channels, out_channels, kernel_size=3, padding=1)
]
self.conv_blocks = nn.Sequential(*conv_blocks)
def __init__(self, in_channels, out_channels):
super(ResidualBlock, self).__init__()
self.in_norm = nn.BatchNorm2d(in_channels)
self.in_conv = WNConv2d(in_channels,
out_channels,
kernel_size=3,
padding=1,
bias=False)
self.out_norm = nn.BatchNorm2d(out_channels)
self.out_conv = WNConv2d(out_channels,
out_channels,
kernel_size=3,
padding=1,
bias=True)
def __init__(self, in_channels, out_channels, drop_prob=0.):
super(GatedConv, self).__init__()
self.nlin = concat_elu
self.conv = WNConv2d(2 * in_channels,
out_channels,
kernel_size=3,
padding=1)
self.drop = nn.Dropout2d(drop_prob)
self.gate = Gate(2 * out_channels, 2 * out_channels)