VToonify / vtoonify /model /vtoonify.py
PKUWilliamYang's picture
V1
983684c
raw
history blame
11.5 kB
import torch
import numpy as np
import math
from torch import nn
from model.stylegan.model import ConvLayer, EqualLinear, Generator, ResBlock
from model.dualstylegan import AdaptiveInstanceNorm, AdaResBlock, DualStyleGAN
import torch.nn.functional as F
# IC-GAN: stylegan discriminator
class ConditionalDiscriminator(nn.Module):
def __init__(self, size, channel_multiplier=2, blur_kernel=[1, 3, 3, 1], use_condition=False, style_num=None):
super().__init__()
channels = {
4: 512,
8: 512,
16: 512,
32: 512,
64: 256 * channel_multiplier,
128: 128 * channel_multiplier,
256: 64 * channel_multiplier,
512: 32 * channel_multiplier,
1024: 16 * channel_multiplier,
}
convs = [ConvLayer(3, channels[size], 1)]
log_size = int(math.log(size, 2))
in_channel = channels[size]
for i in range(log_size, 2, -1):
out_channel = channels[2 ** (i - 1)]
convs.append(ResBlock(in_channel, out_channel, blur_kernel))
in_channel = out_channel
self.convs = nn.Sequential(*convs)
self.stddev_group = 4
self.stddev_feat = 1
self.use_condition = use_condition
if self.use_condition:
self.condition_dim = 128
# map style degree to 64-dimensional vector
self.label_mapper = nn.Sequential(
nn.Linear(1, 64),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Linear(64, 64),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Linear(64, self.condition_dim//2),
)
# map style code index to 64-dimensional vector
self.style_mapper = nn.Embedding(style_num, self.condition_dim-self.condition_dim//2)
else:
self.condition_dim = 1
self.final_conv = ConvLayer(in_channel + 1, channels[4], 3)
self.final_linear = nn.Sequential(
EqualLinear(channels[4] * 4 * 4, channels[4], activation="fused_lrelu"),
EqualLinear(channels[4], self.condition_dim),
)
def forward(self, input, degree_label=None, style_ind=None):
out = self.convs(input)
batch, channel, height, width = out.shape
group = min(batch, self.stddev_group)
stddev = out.view(
group, -1, self.stddev_feat, channel // self.stddev_feat, height, width
)
stddev = torch.sqrt(stddev.var(0, unbiased=False) + 1e-8)
stddev = stddev.mean([2, 3, 4], keepdims=True).squeeze(2)
stddev = stddev.repeat(group, 1, height, width)
out = torch.cat([out, stddev], 1)
out = self.final_conv(out)
out = out.view(batch, -1)
if self.use_condition:
h = self.final_linear(out)
condition = torch.cat((self.label_mapper(degree_label), self.style_mapper(style_ind)), dim=1)
out = (h * condition).sum(dim=1, keepdim=True) * (1 / np.sqrt(self.condition_dim))
else:
out = self.final_linear(out)
return out
class VToonifyResBlock(nn.Module):
def __init__(self, fin):
super().__init__()
self.conv = nn.Conv2d(fin, fin, 3, 1, 1)
self.conv2 = nn.Conv2d(fin, fin, 3, 1, 1)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
def forward(self, x):
out = self.lrelu(self.conv(x))
out = self.lrelu(self.conv2(out))
out = (out + x) / math.sqrt(2)
return out
class Fusion(nn.Module):
def __init__(self, in_channels, skip_channels, out_channels):
super().__init__()
# create conv layers
self.conv = nn.Conv2d(in_channels + skip_channels, out_channels, 3, 1, 1, bias=True)
self.norm = AdaptiveInstanceNorm(in_channels + skip_channels, 128)
self.conv2 = nn.Conv2d(in_channels + skip_channels, 1, 3, 1, 1, bias=True)
#'''
self.linear = nn.Sequential(
nn.Linear(1, 64),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Linear(64, 128),
nn.LeakyReLU(negative_slope=0.2, inplace=True)
)
def forward(self, f_G, f_E, d_s=1):
# label of style degree
label = self.linear(torch.zeros(f_G.size(0),1).to(f_G.device) + d_s)
out = torch.cat([f_G, abs(f_G-f_E)], dim=1)
m_E = (F.relu(self.conv2(self.norm(out, label)))).tanh()
f_out = self.conv(torch.cat([f_G, f_E * m_E], dim=1))
return f_out, m_E
class VToonify(nn.Module):
def __init__(self,
in_size=256,
out_size=1024,
img_channels=3,
style_channels=512,
num_mlps=8,
channel_multiplier=2,
num_res_layers=6,
backbone = 'dualstylegan',
):
super().__init__()
self.backbone = backbone
if self.backbone == 'dualstylegan':
# DualStyleGAN, with weights being fixed
self.generator = DualStyleGAN(out_size, style_channels, num_mlps, channel_multiplier)
else:
# StyleGANv2, with weights being fixed
self.generator = Generator(out_size, style_channels, num_mlps, channel_multiplier)
self.in_size = in_size
self.style_channels = style_channels
channels = self.generator.channels
# encoder
num_styles = int(np.log2(out_size)) * 2 - 2
encoder_res = [2**i for i in range(int(np.log2(in_size)), 4, -1)]
self.encoder = nn.ModuleList()
self.encoder.append(
nn.Sequential(
nn.Conv2d(img_channels+19, 32, 3, 1, 1, bias=True),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(32, channels[in_size], 3, 1, 1, bias=True),
nn.LeakyReLU(negative_slope=0.2, inplace=True)))
for res in encoder_res:
in_channels = channels[res]
if res > 32:
out_channels = channels[res // 2]
block = nn.Sequential(
nn.Conv2d(in_channels, out_channels, 3, 2, 1, bias=True),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.Conv2d(out_channels, out_channels, 3, 1, 1, bias=True),
nn.LeakyReLU(negative_slope=0.2, inplace=True))
self.encoder.append(block)
else:
layers = []
for _ in range(num_res_layers):
layers.append(VToonifyResBlock(in_channels))
self.encoder.append(nn.Sequential(*layers))
block = nn.Conv2d(in_channels, img_channels, 1, 1, 0, bias=True)
self.encoder.append(block)
# trainable fusion module
self.fusion_out = nn.ModuleList()
self.fusion_skip = nn.ModuleList()
for res in encoder_res[::-1]:
num_channels = channels[res]
if self.backbone == 'dualstylegan':
self.fusion_out.append(
Fusion(num_channels, num_channels, num_channels))
else:
self.fusion_out.append(
nn.Conv2d(num_channels * 2, num_channels, 3, 1, 1, bias=True))
self.fusion_skip.append(
nn.Conv2d(num_channels + 3, 3, 3, 1, 1, bias=True))
# Modified ModRes blocks in DualStyleGAN, with weights being fixed
if self.backbone == 'dualstylegan':
self.res = nn.ModuleList()
self.res.append(AdaResBlock(self.generator.channels[2 ** 2])) # for conv1, no use in this model
for i in range(3, 6):
out_channel = self.generator.channels[2 ** i]
self.res.append(AdaResBlock(out_channel, dilation=2**(5-i)))
self.res.append(AdaResBlock(out_channel, dilation=2**(5-i)))
def forward(self, x, style, d_s=None, return_mask=False, return_feat=False):
# map style to W+ space
if style is not None and style.ndim < 3:
if self.backbone == 'dualstylegan':
resstyles = self.generator.style(style).unsqueeze(1).repeat(1, self.generator.n_latent, 1)
adastyles = style.unsqueeze(1).repeat(1, self.generator.n_latent, 1)
elif style is not None:
nB, nL, nD = style.shape
if self.backbone == 'dualstylegan':
resstyles = self.generator.style(style.reshape(nB*nL, nD)).reshape(nB, nL, nD)
adastyles = style
if self.backbone == 'dualstylegan':
adastyles = adastyles.clone()
for i in range(7, self.generator.n_latent):
adastyles[:, i] = self.generator.res[i](adastyles[:, i])
# obtain multi-scale content features
feat = x
encoder_features = []
# downsampling conv parts of E
for block in self.encoder[:-2]:
feat = block(feat)
encoder_features.append(feat)
encoder_features = encoder_features[::-1]
# Resblocks in E
for ii, block in enumerate(self.encoder[-2]):
feat = block(feat)
# adjust Resblocks with ModRes blocks
if self.backbone == 'dualstylegan':
feat = self.res[ii+1](feat, resstyles[:, ii+1], d_s)
# the last-layer feature of E (inputs of backbone)
out = feat
skip = self.encoder[-1](feat)
if return_feat:
return out, skip
# 32x32 ---> higher res
_index = 1
m_Es = []
for conv1, conv2, to_rgb in zip(
self.stylegan().convs[6::2], self.stylegan().convs[7::2], self.stylegan().to_rgbs[3:]):
# pass the mid-layer features of E to the corresponding resolution layers of G
if 2 ** (5+((_index-1)//2)) <= self.in_size:
fusion_index = (_index - 1) // 2
f_E = encoder_features[fusion_index]
if self.backbone == 'dualstylegan':
out, m_E = self.fusion_out[fusion_index](out, f_E, d_s)
skip = self.fusion_skip[fusion_index](torch.cat([skip, f_E*m_E], dim=1))
m_Es += [m_E]
else:
out = self.fusion_out[fusion_index](torch.cat([out, f_E], dim=1))
skip = self.fusion_skip[fusion_index](torch.cat([skip, f_E], dim=1))
# remove the noise input
batch, _, height, width = out.shape
noise = x.new_empty(batch, 1, height * 2, width * 2).normal_().detach() * 0.0
out = conv1(out, adastyles[:, _index+6], noise=noise)
out = conv2(out, adastyles[:, _index+7], noise=noise)
skip = to_rgb(out, adastyles[:, _index+8], skip)
_index += 2
image = skip
if return_mask and self.backbone == 'dualstylegan':
return image, m_Es
return image
def stylegan(self):
if self.backbone == 'dualstylegan':
return self.generator.generator
else:
return self.generator
def zplus2wplus(self, zplus):
return self.stylegan().style(zplus.reshape(zplus.shape[0]*zplus.shape[1], zplus.shape[2])).reshape(zplus.shape)