zyainfal
commented
3 years ago Hi, as I haven't re-implement my code, the test code will not be released in near future. But here is a snippet of the HieRFE code for your reference (We use resnet50 as backbone, and the omega_only indicates if the constant input C predicted, discussed in ablation study).
model_urls = {
'resnet18': 'https://s3.amazonaws.com/pytorch/models/resnet18-5c106cde.pth',
'resnet34': 'https://s3.amazonaws.com/pytorch/models/resnet34-333f7ec4.pth',
'resnet50': 'https://s3.amazonaws.com/pytorch/models/resnet50-19c8e357.pth',
'resnet101': 'https://s3.amazonaws.com/pytorch/models/resnet101-5d3b4d8f.pth',
'resnet152': 'https://s3.amazonaws.com/pytorch/models/resnet152-b121ed2d.pth',
}
def conv3x3(in_planes, out_planes, stride=1, groups=1, dilation=1):
"""3x3 convolution with padding"""
return nn.Conv2d(in_planes, out_planes, kernel_size=3, stride=stride,
padding=dilation, groups=groups, bias=False, dilation=dilation)
def conv1x1(in_planes, out_planes, stride=1):
"""1x1 convolution"""
return nn.Conv2d(in_planes, out_planes, kernel_size=1, stride=stride, bias=False)
class BasicBlock(nn.Module):
expansion = 1
def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
base_width=64, dilation=1, norm_layer=None):
super(BasicBlock, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
if groups != 1 or base_width != 64:
raise ValueError('BasicBlock only supports groups=1 and base_width=64')
if dilation > 1:
raise NotImplementedError("Dilation > 1 not supported in BasicBlock")
# Both self.conv1 and self.downsample layers downsample the input when stride != 1
self.conv1 = conv3x3(inplanes, planes, stride)
self.bn1 = norm_layer(planes)
self.relu = nn.ReLU(inplace=True)
self.conv2 = conv3x3(planes, planes)
self.bn2 = norm_layer(planes)
self.downsample = downsample
self.stride = stride
def forward(self, x):
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
class Bottleneck(nn.Module):
# Bottleneck in torchvision places the stride for downsampling at 3x3 convolution(self.conv2)
# while original implementation places the stride at the first 1x1 convolution(self.conv1)
# according to "Deep residual learning for image recognition"https://arxiv.org/abs/1512.03385.
# This variant is also known as ResNet V1.5 and improves accuracy according to
# https://ngc.nvidia.com/catalog/model-scripts/nvidia:resnet_50_v1_5_for_pytorch.
expansion = 4
def __init__(self, inplanes, planes, stride=1, downsample=None, groups=1,
base_width=64, dilation=1, norm_layer=None):
super(Bottleneck, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
width = int(planes * (base_width / 64.)) * groups
# Both self.conv2 and self.downsample layers downsample the input when stride != 1
self.conv1 = conv1x1(inplanes, width)
self.bn1 = norm_layer(width)
self.conv2 = conv3x3(width, width, stride, groups, dilation)
self.bn2 = norm_layer(width)
self.conv3 = conv1x1(width, planes * self.expansion)
self.bn3 = norm_layer(planes * self.expansion)
self.relu = nn.ReLU(inplace=True)
self.downsample = downsample
self.stride = stride
def forward(self, x):
identity = x
out = self.conv1(x)
out = self.bn1(out)
out = self.relu(out)
out = self.conv2(out)
out = self.bn2(out)
out = self.relu(out)
out = self.conv3(out)
out = self.bn3(out)
if self.downsample is not None:
identity = self.downsample(x)
out += identity
out = self.relu(out)
return out
class ResNet(nn.Module):
def __init__(self, block, layers, num_classes=1000, zero_init_residual=False,
groups=1, width_per_group=64, replace_stride_with_dilation=None,
norm_layer=None, omega_only=False):
super(ResNet, self).__init__()
if norm_layer is None:
norm_layer = nn.BatchNorm2d
self._norm_layer = norm_layer
self.omega_only = omega_only
self.inplanes = 64
self.dilation = 1
if replace_stride_with_dilation is None:
# each element in the tuple indicates if we should replace
# the 2x2 stride with a dilated convolution instead
replace_stride_with_dilation = [False, False, False, False]
if len(replace_stride_with_dilation) != 4:
raise ValueError("replace_stride_with_dilation should be None "
"or a 3-element tuple, got {}".format(replace_stride_with_dilation))
self.groups = groups
self.base_width = width_per_group
self.conv1 = nn.Conv2d(3, self.inplanes, kernel_size=7, stride=2, padding=3,
bias=False)
self.bn1 = norm_layer(self.inplanes)
self.relu = nn.ReLU(inplace=True)
self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)
self.layer1 = self._make_layer(block, 64, layers[0])
self.layer2 = self._make_layer(block, 128, layers[1], stride=2,
dilate=replace_stride_with_dilation[0])
self.layer3 = self._make_layer(block, 256, layers[2], stride=2,
dilate=replace_stride_with_dilation[1])
self.layer4 = self._make_layer(block, 512, layers[3], stride=2,
dilate=replace_stride_with_dilation[2])
if not omega_only:
self.layer5 = self._make_layer(block, 512, layers[4], stride=2,
dilate=replace_stride_with_dilation[2])
for m in self.modules():
if isinstance(m, nn.Conv2d):
nn.init.kaiming_normal_(m.weight, mode='fan_out', nonlinearity='relu')
elif isinstance(m, (nn.BatchNorm2d, nn.GroupNorm)):
nn.init.constant_(m.weight, 1)
nn.init.constant_(m.bias, 0)
# Zero-initialize the last BN in each residual branch,
# so that the residual branch starts with zeros, and each residual block behaves like an identity.
# This improves the model by 0.2~0.3% according to https://arxiv.org/abs/1706.02677
if zero_init_residual:
for m in self.modules():
if isinstance(m, Bottleneck):
nn.init.constant_(m.bn3.weight, 0)
elif isinstance(m, BasicBlock):
nn.init.constant_(m.bn2.weight, 0)
def _make_layer(self, block, planes, blocks, stride=1, dilate=False):
norm_layer = self._norm_layer
downsample = None
previous_dilation = self.dilation
if dilate:
self.dilation *= stride
stride = 1
if stride != 1 or self.inplanes != planes * block.expansion:
downsample = nn.Sequential(
conv1x1(self.inplanes, planes * block.expansion, stride),
norm_layer(planes * block.expansion),
)
layers = []
layers.append(block(self.inplanes, planes, stride, downsample, self.groups,
self.base_width, previous_dilation, norm_layer))
self.inplanes = planes * block.expansion
for _ in range(1, blocks):
layers.append(block(self.inplanes, planes, groups=self.groups,
base_width=self.base_width, dilation=self.dilation,
norm_layer=norm_layer))
return nn.Sequential(*layers)
def _forward_impl(self, x):
# See note [TorchScript super()]
x = self.conv1(x)
x = self.bn1(x)
x = self.relu(x)
x = self.maxpool(x)
x = self.layer1(x)
c1 = self.layer2(x)
c2 = self.layer3(c1)
c3 = self.layer4(c2)
if not self.omega_only:
c4 = self.layer5(c3)
return c1, c2, c3, c4
return c1, c2, c3
def forward(self, x):
return self._forward_impl(x)
def _resnet(arch, block, layers, **kwargs):
model = ResNet(block, layers, **kwargs)
return model
def resnet18(pretrained=False):
"""Constructs a ResNet-18 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNet(BasicBlock, [2, 2, 2, 2, 2])
if pretrained:
model.load_state_dict(model_zoo.load_url(model_urls['resnet18']), strict=False)
return model
def resnet34(pretrained=False):
"""Constructs a ResNet-34 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNet(BasicBlock, [3, 4, 6, 3, 2])
if pretrained:
model.load_state_dict(model_zoo.load_url(model_urls['resnet34']), strict=False)
return model
def resnet50(pretrained=False, omega_only=False):
"""Constructs a ResNet-50 model.
Args:
pretrained (bool): If True, returns a model pre-trained on ImageNet
"""
model = ResNet(Bottleneck, [3, 4, 6, 3, 2], omega_only=omega_only)
if pretrained:
model.load_state_dict(model_zoo.load_url(model_urls['resnet50']), strict=False)
return model
def conv_bn(inp, oup, stride = 1):
return nn.Sequential(
nn.Conv2d(inp, oup, 3, stride, 1, bias=True),
nn.BatchNorm2d(oup),
nn.ReLU()
)
def conv_bn1X1(inp, oup, stride):
return nn.Sequential(
nn.Conv2d(inp, oup, 1, stride, padding=0, bias=True),
nn.BatchNorm2d(oup),
nn.ReLU()
)
class FPN(nn.Module):
def __init__(self, backbone, depth, omega_only):
super(FPN, self).__init__()
scale = 4 if depth==50 else 1
self.omega_only = omega_only
self.backbone = backbone
self.output1 = conv_bn(128*scale, 512, stride=1)
self.output2 = conv_bn(256*scale, 512, stride=1)
self.output3 = conv_bn(512*scale, 512, stride=1)
if not omega_only:
self.output4 = nn.Sequential(
nn.Conv2d(512*scale, 512, 3, 1, padding=1, bias=True),
nn.BatchNorm2d(512)
)
def forward(self, x):
# names = list(input.keys())
if not self.omega_only:
c1, c2, c3, c4 = self.backbone(x)
p4 = self.output4(c4) # N, 512, 4, 4
else:
c1, c2, c3 = self.backbone(x)
p4 = None
p3 = self.output3(c3) # N, 512, 8, 8
p2 = self.output2(c2) + F.upsample(p3, scale_factor=2, mode='bilinear', align_corners=True) # N, 512, 16, 16
p1 = self.output1(c1) + F.upsample(p2, scale_factor=2, mode='bilinear', align_corners=True) # N, 512, 32, 32
return p4, p3, p2, p1
class styleMapping(nn.Module):
def __init__(self, size):
super(styleMapping, self).__init__()
num_layers = int(math.log(size, 2))
convs = []
self.convs = nn.ModuleList()
for i in range(num_layers):
convs.append(nn.Conv2d(512, 512, 3, 2, padding=1, bias=True))
convs.append(nn.BatchNorm2d(512))
convs.append(nn.LeakyReLU(inplace=True))
self.convs = nn.Sequential(*convs)
def forward(self, x):
x = self.convs(x).squeeze(2).squeeze(2).unsqueeze(1)
return x
class HieRFE(nn.Module):
def __init__(self, backbone, num_latents=[4, 6, 8], depth=50, omega_only=False):
super(HieRFE, self).__init__()
self.fpn = FPN(backbone, depth, omega_only)
self.act = nn.Tanh()
self.mapping1 = nn.ModuleList()
for i in range(num_latents[0]):
self.mapping1.append(styleMapping(8))
self.mapping2 = nn.ModuleList()
for i in range(num_latents[1]):
self.mapping2.append(styleMapping(16))
self.mapping3 = nn.ModuleList()
for i in range(num_latents[2]):
self.mapping3.append(styleMapping(32))
def forward(self, x):
latents = []
f4, f8, f16, f32 = self.fpn(x)
for maps in self.mapping1:
latents.append(maps(f8))
for maps in self.mapping2:
latents.append(maps(f16))
for maps in self.mapping3:
latents.append(maps(f32))
latents = torch.cat(latents, 1)
return self.act(latents), f4And there is the snippet for FTM:
class TransferBlock(nn.Module):
def __init__(self, num_blocks):
super(TransferBlock, self).__init__()
self.num_blocks = num_blocks
self.idd_selectors = nn.ModuleList()
self.idd_shifters = nn.ModuleList()
self.att_selectors = nn.ModuleList()
self.att_shifters = nn.ModuleList()
self.act = nn.LeakyReLU(True)
for i in range(self.num_blocks):
self.idd_selectors.append(nn.Sequential(nn.Linear(1024, 256), nn.ReLU(True), nn.Linear(256, 512), nn.Sigmoid()))
self.idd_shifters.append(nn.Sequential(nn.Linear(1024, 256), nn.ReLU(True), nn.Linear(256, 512), nn.Tanh()))
self.att_selectors.append(nn.Sequential(nn.Linear(1024, 256), nn.ReLU(True), nn.Linear(256, 512), nn.Sigmoid()))
self.att_shifters.append(nn.Sequential(nn.Linear(1024, 256), nn.ReLU(True), nn.Linear(256, 512), nn.Tanh()))
def forward(self, idd, att):
for i in range(self.num_blocks):
fuse = torch.cat([idd, att], dim=1)
idd = self.act(idd * self.idd_selectors[i](fuse) + self.idd_shifters[i](fuse))
att = self.act(att * self.att_selectors[i](fuse) + self.att_shifters[i](fuse))
return idd.unsqueeze(1), att.unsqueeze(1)
class FaceTransferModule(nn.Module):
def __init__(self, num_blocks=3, swap_indice=4, num_latents=18):
super(FaceTransferModule, self).__init__()
self.swap_indice = swap_indice
self.num_latents = num_latents - swap_indice
self.blocks = nn.ModuleList()
for i in range(self.num_latents):
self.blocks.append(pspTransferBlock(num_blocks))
self.weight = nn.Parameter(torch.randn(1, self.num_latents, 512))
def forward(self, idd, att):
att_low = att[:, :self.swap_indice]
idd_high = idd[:, self.swap_indice:]
att_high = att[:, self.swap_indice:]
N = idd.size(0)
idds = []
atts = []
for i in range(self.num_latents):
new_idd, new_att = self.blocks[i](idd_high[:, i], att_high[:, i])
idds.append(new_idd)
atts.append(new_att)
idds = torch.cat(idds, 1)
atts = torch.cat(atts, 1)
scale = torch.sigmoid(self.weight).expand(N, -1, -1)
latents = scale * idds + (1-scale) * atts
return torch.cat([att_low, latents], 1)
Anyway, bugs may exist because I re-write this code in rush without testing. Hope this could be helpful.
HoiM
Thank you for your great work.
The paper does not explain the detailed structure of HieRFE. Therefore I would like to know if you could provide more details on the structure of the network. Specifically, may I know how the FPN is applied into the structure, details on the number of channels and feature map sizes, Non-linear mapping structure, etc. ?
Thank you in advance.