def __init__(
self,
hidden_size=64,
num_layer=2,
readout="avg",
layernorm: bool = False,
set2set_lstm_layer: int = 3,
set2set_iter: int = 6,
):
super(UnsupervisedGCN, self).__init__()
self.layers = nn.ModuleList([
GCNLayer(
in_feats=hidden_size,
out_feats=hidden_size,
activation=F.relu if i + 1 < num_layer else None,
residual=False,
batchnorm=False,
dropout=0.0,
) for i in range(num_layer)
])
if readout == "avg":
self.readout = AvgPooling()
elif readout == "set2set":
self.readout = Set2Set(hidden_size,
n_iters=set2set_iter,
n_layers=set2set_lstm_layer)
self.linear = nn.Linear(2 * hidden_size, hidden_size)
elif readout == "root":
# HACK: process outside the model part
self.readout = lambda _, x: x
else:
raise NotImplementedError
self.layernorm = layernorm
if layernorm:
self.ln = nn.LayerNorm(hidden_size, elementwise_affine=False)
def __init__(self,pretrained_weights_fn=None):
super().__init__()
self.alexnet = AlexNet()
self.lpips_weights = nn.ModuleList()
for channels in self.alexnet.channels:
self.lpips_weights.append(Conv1x1(channels, 1))
if pretrained_weights_fn is not None:
self._load_lpips_weights(pretrained_weights_fn)
# imagenet normalization for range [-1, 1]
self.mu = torch.tensor([-0.03, -0.088, -0.188]).view(1, 3, 1, 1)
self.sigma = torch.tensor([0.458, 0.448, 0.450]).view(1, 3, 1, 1)
def __init__(self, img_size=256, style_dim=64, max_conv_dim=512, w_hpf=1):
super().__init__()
dim_in = 2**14 // img_size
self.img_size = img_size
self.from_rgb = nn.Conv2d(3, dim_in, 3, 1, 1)
self.encode = nn.ModuleList()
self.decode = nn.ModuleList()
self.to_rgb = nn.Sequential(nn.InstanceNorm2d(dim_in, affine=True),
nn.LeakyReLU(0.2),
nn.Conv2d(dim_in, 3, 1, 1, 0))
# down/up-sampling blocks
repeat_num = int(np.log2(img_size)) - 4
if w_hpf > 0:
repeat_num += 1
for _ in range(repeat_num):
dim_out = min(dim_in * 2, max_conv_dim)
self.encode.append(
ResBlk(dim_in, dim_out, normalize=True, downsample=True))
self.decode.insert(0,
AdainResBlk(dim_out,
dim_in,
style_dim,
w_hpf=w_hpf,
upsample=True)) # stack-like
dim_in = dim_out
# bottleneck blocks
for _ in range(2):
self.encode.append(ResBlk(dim_out, dim_out, normalize=True))
self.decode.insert(
0, AdainResBlk(dim_out, dim_out, style_dim, w_hpf=w_hpf))
if w_hpf > 0:
device = porch.device(
'cuda' if porch.cuda.is_available() else 'cpu')
self.hpf = HighPass(w_hpf, device)
def __init__(self, latent_dim=16, style_dim=64, num_domains=2):
super().__init__()
layers = []
layers += [nn.Linear(latent_dim, 512)]
layers += [nn.ReLU()]
for _ in range(3):
layers += [nn.Linear(512, 512)]
layers += [nn.ReLU()]
self.shared = nn.Sequential(*layers)
self.unshared = nn.ModuleList()
for _ in range(num_domains):
self.unshared += [
nn.Sequential(nn.Linear(512, 512), nn.ReLU(),
nn.Linear(512, 512), nn.ReLU(),
nn.Linear(512, 512), nn.ReLU(),
nn.Linear(512, style_dim))
]
def __init__(self,
img_size=256,
style_dim=64,
num_domains=2,
max_conv_dim=512):
super().__init__()
dim_in = 2**14 // img_size
blocks = []
blocks += [nn.Conv2d(3, dim_in, 3, 1, 1)]
repeat_num = int(np.log2(img_size)) - 2
for _ in range(repeat_num):
dim_out = min(dim_in * 2, max_conv_dim)
blocks += [ResBlk(dim_in, dim_out, downsample=True)]
dim_in = dim_out
blocks += [nn.LeakyReLU(0.2)]
blocks += [nn.Conv2d(dim_out, dim_out, 4, 1, 0)]
blocks += [nn.LeakyReLU(0.2)]
self.shared = nn.Sequential(*blocks)
self.unshared = nn.ModuleList()
for _ in range(num_domains):
self.unshared += [nn.Linear(dim_out, style_dim)]
def __init__(self, G_ch=64, G_depth=2, dim_z=128, bottom_width=4, resolution=128,
G_kernel_size=3, G_attn='64', n_classes=1000,
num_G_SVs=1, num_G_SV_itrs=1,
G_shared=True, shared_dim=0, hier=False,
cross_replica=False, mybn=False,
G_activation=nn.ReLU(inplace=False),
G_lr=5e-5, G_B1=0.0, G_B2=0.999, adam_eps=1e-8,
BN_eps=1e-5, SN_eps=1e-12, G_mixed_precision=False, G_fp16=False,
G_init='ortho', skip_init=False, no_optim=False,
G_param='SN', norm_style='bn',
**kwargs):
super(Generator, self).__init__()
# Channel width mulitplier
self.ch = G_ch
# Number of resblocks per stage
self.G_depth = G_depth
# Dimensionality of the latent space
self.dim_z = dim_z
# The initial spatial dimensions
self.bottom_width = bottom_width
# Resolution of the output
self.resolution = resolution
# Kernel size?
self.kernel_size = G_kernel_size
# Attention?
self.attention = G_attn
# number of classes, for use in categorical conditional generation
self.n_classes = n_classes
# Use shared embeddings?
self.G_shared = G_shared
# Dimensionality of the shared embedding? Unused if not using G_shared
self.shared_dim = shared_dim if shared_dim > 0 else dim_z
# Hierarchical latent space?
self.hier = hier
# Cross replica batchnorm?
self.cross_replica = cross_replica
# Use my batchnorm?
self.mybn = mybn
# nonlinearity for residual blocks
self.activation = G_activation
# Initialization style
self.init = G_init
# Parameterization style
self.G_param = G_param
# Normalization style
self.norm_style = norm_style
# Epsilon for BatchNorm?
self.BN_eps = BN_eps
# Epsilon for Spectral Norm?
self.SN_eps = SN_eps
# fp16?
self.fp16 = G_fp16
# Architecture dict
self.arch = G_arch(self.ch, self.attention)[resolution]
# Which convs, batchnorms, and linear layers to use
if self.G_param == 'SN':
self.which_conv = functools.partial(layers.SNConv2d,
kernel_size=3, padding=1,
num_svs=num_G_SVs, num_itrs=num_G_SV_itrs,
eps=self.SN_eps)
self.which_linear = functools.partial(layers.SNLinear,
num_svs=num_G_SVs, num_itrs=num_G_SV_itrs,
eps=self.SN_eps)
else:
self.which_conv = functools.partial(nn.Conv2d, kernel_size=3, padding=1)
self.which_linear = nn.Linear
# We use a non-spectral-normed embedding here regardless;
# For some reason applying SN to G's embedding seems to randomly cripple G
self.which_embedding = nn.Embedding
bn_linear = (functools.partial(self.which_linear, bias=False) if self.G_shared
else self.which_embedding)
self.which_bn = functools.partial(layers.ccbn,
which_linear=bn_linear,
cross_replica=self.cross_replica,
mybn=self.mybn,
input_size=(self.shared_dim + self.dim_z if self.G_shared
else self.n_classes),
norm_style=self.norm_style,
eps=self.BN_eps)
# Prepare model
# If not using shared embeddings, self.shared is just a passthrough
self.shared = (self.which_embedding(n_classes, self.shared_dim) if G_shared
else layers.identity())
# First linear layer
self.linear = self.which_linear(self.dim_z + self.shared_dim, self.arch['in_channels'][0] * (self.bottom_width **2))
# self.blocks is a doubly-nested list of modules, the outer loop intended
# to be over blocks at a given resolution (resblocks and/or self-attention)
# while the inner loop is over a given block
self.blocks = []
for index in range(len(self.arch['out_channels'])):
self.blocks += [[GBlock(in_channels=self.arch['in_channels'][index],
out_channels=self.arch['in_channels'][index] if g_index==0 else self.arch['out_channels'][index],
which_conv=self.which_conv,
which_bn=self.which_bn,
activation=self.activation,
upsample=(functools.partial(F.interpolate, scale_factor=2)
if self.arch['upsample'][index] and g_index == (self.G_depth-1) else None))]
for g_index in range(self.G_depth)]
# If attention on this block, attach it to the end
if self.arch['attention'][self.arch['resolution'][index]]:
print('Adding attention layer in G at resolution %d' % self.arch['resolution'][index])
self.blocks[-1] += [layers.Attention(self.arch['out_channels'][index], self.which_conv)]
# Turn self.blocks into a ModuleList so that it's all properly registered.
self.blocks = nn.ModuleList([nn.ModuleList(block) for block in self.blocks])
# output layer: batchnorm-relu-conv.
# Consider using a non-spectral conv here
self.output_layer = nn.Sequential(layers.bn(self.arch['out_channels'][-1],
cross_replica=self.cross_replica,
mybn=self.mybn),
self.activation,
self.which_conv(self.arch['out_channels'][-1], 3))
# Initialize weights. Optionally skip init for testing.
if not skip_init:
self.init_weights()
# Set up optimizer
# If this is an EMA copy, no need for an optim, so just return now
if no_optim:
return
self.lr, self.B1, self.B2, self.adam_eps = G_lr, G_B1, G_B2, adam_eps
if G_mixed_precision:
print('Using fp16 adam in G...')
import utils
self.optim = utils.Adam16(params=self.parameters(), lr=self.lr,
betas=(self.B1, self.B2), weight_decay=0,
eps=self.adam_eps)
else:
self.optim = optim.Adam(params=self.parameters(), lr=self.lr,
betas=(self.B1, self.B2), weight_decay=0,
eps=self.adam_eps)
def __init__(self, D_ch=64, D_wide=True, D_depth=2, resolution=128,
D_kernel_size=3, D_attn='64', n_classes=1000,
num_D_SVs=1, num_D_SV_itrs=1, D_activation=nn.ReLU(inplace=False),
D_lr=2e-4, D_B1=0.0, D_B2=0.999, adam_eps=1e-8,
SN_eps=1e-12, output_dim=1, D_mixed_precision=False, D_fp16=False,
D_init='ortho', skip_init=False, D_param='SN', **kwargs):
super(Discriminator, self).__init__()
# Width multiplier
self.ch = D_ch
# Use Wide D as in BigGAN and SA-GAN or skinny D as in SN-GAN?
self.D_wide = D_wide
# How many resblocks per stage?
self.D_depth = D_depth
# Resolution
self.resolution = resolution
# Kernel size
self.kernel_size = D_kernel_size
# Attention?
self.attention = D_attn
# Number of classes
self.n_classes = n_classes
# Activation
self.activation = D_activation
# Initialization style
self.init = D_init
# Parameterization style
self.D_param = D_param
# Epsilon for Spectral Norm?
self.SN_eps = SN_eps
# Fp16?
self.fp16 = D_fp16
# Architecture
self.arch = D_arch(self.ch, self.attention)[resolution]
# Which convs, batchnorms, and linear layers to use
# No option to turn off SN in D right now
if self.D_param == 'SN':
self.which_conv = functools.partial(layers.SNConv2d,
kernel_size=3, padding=1,
num_svs=num_D_SVs, num_itrs=num_D_SV_itrs,
eps=self.SN_eps)
self.which_linear = functools.partial(layers.SNLinear,
num_svs=num_D_SVs, num_itrs=num_D_SV_itrs,
eps=self.SN_eps)
self.which_embedding = functools.partial(layers.SNEmbedding,
num_svs=num_D_SVs, num_itrs=num_D_SV_itrs,
eps=self.SN_eps)
# Prepare model
# Stem convolution
self.input_conv = self.which_conv(3, self.arch['in_channels'][0])
# self.blocks is a doubly-nested list of modules, the outer loop intended
# to be over blocks at a given resolution (resblocks and/or self-attention)
self.blocks = []
for index in range(len(self.arch['out_channels'])):
self.blocks += [[DBlock(in_channels=self.arch['in_channels'][index] if d_index==0 else self.arch['out_channels'][index],
out_channels=self.arch['out_channels'][index],
which_conv=self.which_conv,
wide=self.D_wide,
activation=self.activation,
preactivation=True,
downsample=(nn.AvgPool2d(2) if self.arch['downsample'][index] and d_index==0 else None))
for d_index in range(self.D_depth)]]
# If attention on this block, attach it to the end
if self.arch['attention'][self.arch['resolution'][index]]:
print('Adding attention layer in D at resolution %d' % self.arch['resolution'][index])
self.blocks[-1] += [layers.Attention(self.arch['out_channels'][index],
self.which_conv)]
# Turn self.blocks into a ModuleList so that it's all properly registered.
self.blocks = nn.ModuleList([nn.ModuleList(block) for block in self.blocks])
# Linear output layer. The output dimension is typically 1, but may be
# larger if we're e.g. turning this into a VAE with an inference output
self.linear = self.which_linear(self.arch['out_channels'][-1], output_dim)
# Embedding for projection discrimination
self.embed = self.which_embedding(self.n_classes, self.arch['out_channels'][-1])
# Initialize weights
if not skip_init:
self.init_weights()
# Set up optimizer
self.lr, self.B1, self.B2, self.adam_eps = D_lr, D_B1, D_B2, adam_eps
if D_mixed_precision:
print('Using fp16 adam in D...')
import utils
self.optim = utils.Adam16(params=self.parameters(), lr=self.lr,
betas=(self.B1, self.B2), weight_decay=0, eps=self.adam_eps)
else:
self.optim = optim.Adam(params=self.parameters(), lr=self.lr,
betas=(self.B1, self.B2), weight_decay=0, eps=self.adam_eps)