metadata
license: mit
Model weight for Fast Style Transfer
class TransformerNetwork(nn.Module):
"""Feedforward Transformation Network without Tanh
reference: https://arxiv.org/abs/1603.08155
exact architecture: https://cs.stanford.edu/people/jcjohns/papers/fast-style/fast-style-supp.pdf
"""
def __init__(self, tanh_multiplier=None):
super(TransformerNetwork, self).__init__()
self.ConvBlock = nn.Sequential(
ConvLayer(3, 32, 9, 1),
nn.ReLU(),
ConvLayer(32, 64, 3, 2),
nn.ReLU(),
ConvLayer(64, 128, 3, 2),
nn.ReLU()
)
self.ResidualBlock = nn.Sequential(
ResidualLayer(128, 3),
ResidualLayer(128, 3),
ResidualLayer(128, 3),
ResidualLayer(128, 3),
ResidualLayer(128, 3)
)
self.DeconvBlock = nn.Sequential(
DeconvLayer(128, 64, 3, 2, 1),
nn.ReLU(),
DeconvLayer(64, 32, 3, 2, 1),
nn.ReLU(),
ConvLayer(32, 3, 9, 1, norm="None")
)
self.tanh_multiplier = tanh_multiplier
def forward(self, x):
x = self.ConvBlock(x)
x = self.ResidualBlock(x)
x = self.DeconvBlock(x)
if isinstance(self.tanh_multiplier, int):
x = self.tanh_multiplier * F.tanh(x)
return x
class ConvLayer(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, stride, norm="instance"):
super(ConvLayer, self).__init__()
# Padding Layers
padding_size = kernel_size // 2
self.pad = nn.ReflectionPad2d(padding_size)
# Convolution Layer
self.conv = nn.Conv2d(in_channels, out_channels, kernel_size, stride)
# Normalization Layers
if norm == "instance":
self.norm = nn.InstanceNorm2d(out_channels, affine=True)
elif norm == "batch":
self.norm = nn.BatchNorm2d(out_channels, affine=True)
else:
self.norm = nn.Identity()
def forward(self, x):
x = self.pad(x)
x = self.conv(x)
x = self.norm(x)
return x
class ResidualLayer(nn.Module):
"""
Deep Residual Learning for Image Recognition
https://arxiv.org/abs/1512.03385
"""
def __init__(self, channels=128, kernel_size=3):
super(ResidualLayer, self).__init__()
self.conv1 = ConvLayer(channels, channels, kernel_size, stride=1)
self.relu = nn.ReLU()
self.conv2 = ConvLayer(channels, channels, kernel_size, stride=1)
def forward(self, x):
identity = x # preserve residual
out = self.relu(self.conv1(x)) # 1st conv layer + activation
out = self.conv2(out) # 2nd conv layer
out = out + identity # add residual
return out
class DeconvLayer(nn.Module):
def __init__(self, in_channels, out_channels, kernel_size, stride, output_padding, norm="instance"):
super(DeconvLayer, self).__init__()
# Transposed Convolution
padding_size = kernel_size // 2
self.conv_transpose = nn.ConvTranspose2d(in_channels, out_channels, kernel_size, stride, padding_size, output_padding)
# Normalization Layers
if norm == "instance":
self.norm = nn.InstanceNorm2d(out_channels, affine=True)
elif norm == "batch":
self.norm = nn.BatchNorm2d(out_channels, affine=True)
else:
self.norm = nn.Identity()
def forward(self, x):
x = self.conv_transpose(x)
out = self.norm(x)
return out