Wav2Lip-HD / basicsr /archs /dfdnet_arch.py
commanderx's picture
Upload 439 files
908a1ab
raw
history blame
7.08 kB
import numpy as np
import torch
import torch.nn as nn
import torch.nn.functional as F
import torch.nn.utils.spectral_norm as SpectralNorm
from basicsr.utils.registry import ARCH_REGISTRY
from .dfdnet_util import AttentionBlock, Blur, MSDilationBlock, UpResBlock, adaptive_instance_normalization
from .vgg_arch import VGGFeatureExtractor
class SFTUpBlock(nn.Module):
"""Spatial feature transform (SFT) with upsampling block."""
def __init__(self, in_channel, out_channel, kernel_size=3, padding=1):
super(SFTUpBlock, self).__init__()
self.conv1 = nn.Sequential(
Blur(in_channel),
SpectralNorm(nn.Conv2d(in_channel, out_channel, kernel_size, padding=padding)),
nn.LeakyReLU(0.04, True),
# The official codes use two LeakyReLU here, so 0.04 for equivalent
)
self.convup = nn.Sequential(
nn.Upsample(scale_factor=2, mode='bilinear', align_corners=False),
SpectralNorm(nn.Conv2d(out_channel, out_channel, kernel_size, padding=padding)),
nn.LeakyReLU(0.2, True),
)
# for SFT scale and shift
self.scale_block = nn.Sequential(
SpectralNorm(nn.Conv2d(in_channel, out_channel, 3, 1, 1)), nn.LeakyReLU(0.2, True),
SpectralNorm(nn.Conv2d(out_channel, out_channel, 3, 1, 1)))
self.shift_block = nn.Sequential(
SpectralNorm(nn.Conv2d(in_channel, out_channel, 3, 1, 1)), nn.LeakyReLU(0.2, True),
SpectralNorm(nn.Conv2d(out_channel, out_channel, 3, 1, 1)), nn.Sigmoid())
# The official codes use sigmoid for shift block, do not know why
def forward(self, x, updated_feat):
out = self.conv1(x)
# SFT
scale = self.scale_block(updated_feat)
shift = self.shift_block(updated_feat)
out = out * scale + shift
# upsample
out = self.convup(out)
return out
@ARCH_REGISTRY.register()
class DFDNet(nn.Module):
"""DFDNet: Deep Face Dictionary Network.
It only processes faces with 512x512 size.
"""
def __init__(self, num_feat, dict_path):
super().__init__()
self.parts = ['left_eye', 'right_eye', 'nose', 'mouth']
# part_sizes: [80, 80, 50, 110]
channel_sizes = [128, 256, 512, 512]
self.feature_sizes = np.array([256, 128, 64, 32])
self.vgg_layers = ['relu2_2', 'relu3_4', 'relu4_4', 'conv5_4']
self.flag_dict_device = False
# dict
self.dict = torch.load(dict_path)
# vgg face extractor
self.vgg_extractor = VGGFeatureExtractor(
layer_name_list=self.vgg_layers,
vgg_type='vgg19',
use_input_norm=True,
range_norm=True,
requires_grad=False)
# attention block for fusing dictionary features and input features
self.attn_blocks = nn.ModuleDict()
for idx, feat_size in enumerate(self.feature_sizes):
for name in self.parts:
self.attn_blocks[f'{name}_{feat_size}'] = AttentionBlock(channel_sizes[idx])
# multi scale dilation block
self.multi_scale_dilation = MSDilationBlock(num_feat * 8, dilation=[4, 3, 2, 1])
# upsampling and reconstruction
self.upsample0 = SFTUpBlock(num_feat * 8, num_feat * 8)
self.upsample1 = SFTUpBlock(num_feat * 8, num_feat * 4)
self.upsample2 = SFTUpBlock(num_feat * 4, num_feat * 2)
self.upsample3 = SFTUpBlock(num_feat * 2, num_feat)
self.upsample4 = nn.Sequential(
SpectralNorm(nn.Conv2d(num_feat, num_feat, 3, 1, 1)), nn.LeakyReLU(0.2, True), UpResBlock(num_feat),
UpResBlock(num_feat), nn.Conv2d(num_feat, 3, kernel_size=3, stride=1, padding=1), nn.Tanh())
def swap_feat(self, vgg_feat, updated_feat, dict_feat, location, part_name, f_size):
"""swap the features from the dictionary."""
# get the original vgg features
part_feat = vgg_feat[:, :, location[1]:location[3], location[0]:location[2]].clone()
# resize original vgg features
part_resize_feat = F.interpolate(part_feat, dict_feat.size()[2:4], mode='bilinear', align_corners=False)
# use adaptive instance normalization to adjust color and illuminations
dict_feat = adaptive_instance_normalization(dict_feat, part_resize_feat)
# get similarity scores
similarity_score = F.conv2d(part_resize_feat, dict_feat)
similarity_score = F.softmax(similarity_score.view(-1), dim=0)
# select the most similar features in the dict (after norm)
select_idx = torch.argmax(similarity_score)
swap_feat = F.interpolate(dict_feat[select_idx:select_idx + 1], part_feat.size()[2:4])
# attention
attn = self.attn_blocks[f'{part_name}_' + str(f_size)](swap_feat - part_feat)
attn_feat = attn * swap_feat
# update features
updated_feat[:, :, location[1]:location[3], location[0]:location[2]] = attn_feat + part_feat
return updated_feat
def put_dict_to_device(self, x):
if self.flag_dict_device is False:
for k, v in self.dict.items():
for kk, vv in v.items():
self.dict[k][kk] = vv.to(x)
self.flag_dict_device = True
def forward(self, x, part_locations):
"""
Now only support testing with batch size = 0.
Args:
x (Tensor): Input faces with shape (b, c, 512, 512).
part_locations (list[Tensor]): Part locations.
"""
self.put_dict_to_device(x)
# extract vggface features
vgg_features = self.vgg_extractor(x)
# update vggface features using the dictionary for each part
updated_vgg_features = []
batch = 0 # only supports testing with batch size = 0
for vgg_layer, f_size in zip(self.vgg_layers, self.feature_sizes):
dict_features = self.dict[f'{f_size}']
vgg_feat = vgg_features[vgg_layer]
updated_feat = vgg_feat.clone()
# swap features from dictionary
for part_idx, part_name in enumerate(self.parts):
location = (part_locations[part_idx][batch] // (512 / f_size)).int()
updated_feat = self.swap_feat(vgg_feat, updated_feat, dict_features[part_name], location, part_name,
f_size)
updated_vgg_features.append(updated_feat)
vgg_feat_dilation = self.multi_scale_dilation(vgg_features['conv5_4'])
# use updated vgg features to modulate the upsampled features with
# SFT (Spatial Feature Transform) scaling and shifting manner.
upsampled_feat = self.upsample0(vgg_feat_dilation, updated_vgg_features[3])
upsampled_feat = self.upsample1(upsampled_feat, updated_vgg_features[2])
upsampled_feat = self.upsample2(upsampled_feat, updated_vgg_features[1])
upsampled_feat = self.upsample3(upsampled_feat, updated_vgg_features[0])
out = self.upsample4(upsampled_feat)
return out