src/pixel3dmm/preprocessing/PIPNet/lib/networks.py

import torch
import torch.nn as nn
import torch.nn.functional as F
import torchvision.models as models
import numpy as np

# net_stride output_size
# 128        2x2
# 64         4x4
# 32         8x8
# pip regression, resnet101
class Pip_resnet101(nn.Module):
    def __init__(self, resnet, num_nb, num_lms=68, input_size=256, net_stride=32):
        super(Pip_resnet101, self).__init__()
        self.num_nb = num_nb
        self.num_lms = num_lms
        self.input_size = input_size
        self.net_stride = net_stride
        self.conv1 = resnet.conv1
        self.bn1 = resnet.bn1
        self.maxpool = resnet.maxpool
        self.sigmoid = nn.Sigmoid()
        self.layer1 = resnet.layer1
        self.layer2 = resnet.layer2
        self.layer3 = resnet.layer3
        self.layer4 = resnet.layer4
        if self.net_stride == 128:
            self.layer5 = nn.Conv2d(2048, 512, kernel_size=3, stride=2, padding=1)
            self.bn5 = nn.BatchNorm2d(512)
            self.layer6 = nn.Conv2d(512, 512, kernel_size=3, stride=2, padding=1)
            self.bn6 = nn.BatchNorm2d(512)
            # init
            nn.init.normal_(self.layer5.weight, std=0.001)
            if self.layer5.bias is not None:
                nn.init.constant_(self.layer5.bias, 0)
            nn.init.constant_(self.bn5.weight, 1)
            nn.init.constant_(self.bn5.bias, 0)

            nn.init.normal_(self.layer6.weight, std=0.001)
            if self.layer6.bias is not None:
                nn.init.constant_(self.layer6.bias, 0)
            nn.init.constant_(self.bn6.weight, 1)
            nn.init.constant_(self.bn6.bias, 0)
        elif self.net_stride == 64:
            self.layer5 = nn.Conv2d(2048, 512, kernel_size=3, stride=2, padding=1)
            self.bn5 = nn.BatchNorm2d(512)
            # init
            nn.init.normal_(self.layer5.weight, std=0.001)
            if self.layer5.bias is not None:
                nn.init.constant_(self.layer5.bias, 0)
            nn.init.constant_(self.bn5.weight, 1)
            nn.init.constant_(self.bn5.bias, 0)
        elif self.net_stride == 32:
            pass
        else:
            print('No such net_stride!')
            exit(0)

        self.cls_layer = nn.Conv2d(2048, num_lms, kernel_size=1, stride=1, padding=0)
        self.x_layer = nn.Conv2d(2048, num_lms, kernel_size=1, stride=1, padding=0)
        self.y_layer = nn.Conv2d(2048, num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_x_layer = nn.Conv2d(2048, num_nb*num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_y_layer = nn.Conv2d(2048, num_nb*num_lms, kernel_size=1, stride=1, padding=0)

        nn.init.normal_(self.cls_layer.weight, std=0.001)
        if self.cls_layer.bias is not None:
            nn.init.constant_(self.cls_layer.bias, 0)

        nn.init.normal_(self.x_layer.weight, std=0.001)
        if self.x_layer.bias is not None:
            nn.init.constant_(self.x_layer.bias, 0)

        nn.init.normal_(self.y_layer.weight, std=0.001)
        if self.y_layer.bias is not None:
            nn.init.constant_(self.y_layer.bias, 0)

        nn.init.normal_(self.nb_x_layer.weight, std=0.001)
        if self.nb_x_layer.bias is not None:
            nn.init.constant_(self.nb_x_layer.bias, 0)

        nn.init.normal_(self.nb_y_layer.weight, std=0.001)
        if self.nb_y_layer.bias is not None:
            nn.init.constant_(self.nb_y_layer.bias, 0)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = F.relu(x)
        x = self.maxpool(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        if self.net_stride == 128:
            x = F.relu(self.bn5(self.layer5(x)))
            x = F.relu(self.bn6(self.layer6(x)))
        elif self.net_stride == 64:
            x = F.relu(self.bn5(self.layer5(x)))
        else:
            pass
        x1 = self.cls_layer(x)
        x2 = self.x_layer(x)
        x3 = self.y_layer(x)
        x4 = self.nb_x_layer(x)
        x5 = self.nb_y_layer(x)
        return x1, x2, x3, x4, x5

# net_stride output_size
# 128        2x2
# 64         4x4
# 32         8x8
# pip regression, resnet50
class Pip_resnet50(nn.Module):
    def __init__(self, resnet, num_nb, num_lms=68, input_size=256, net_stride=32):
        super(Pip_resnet50, self).__init__()
        self.num_nb = num_nb
        self.num_lms = num_lms
        self.input_size = input_size
        self.net_stride = net_stride
        self.conv1 = resnet.conv1
        self.bn1 = resnet.bn1
        self.maxpool = resnet.maxpool
        self.sigmoid = nn.Sigmoid()
        self.layer1 = resnet.layer1
        self.layer2 = resnet.layer2
        self.layer3 = resnet.layer3
        self.layer4 = resnet.layer4
        if self.net_stride == 128:
            self.layer5 = nn.Conv2d(2048, 512, kernel_size=3, stride=2, padding=1)
            self.bn5 = nn.BatchNorm2d(512)
            self.layer6 = nn.Conv2d(512, 512, kernel_size=3, stride=2, padding=1)
            self.bn6 = nn.BatchNorm2d(512)
            # init
            nn.init.normal_(self.layer5.weight, std=0.001)
            if self.layer5.bias is not None:
                nn.init.constant_(self.layer5.bias, 0)
            nn.init.constant_(self.bn5.weight, 1)
            nn.init.constant_(self.bn5.bias, 0)

            nn.init.normal_(self.layer6.weight, std=0.001)
            if self.layer6.bias is not None:
                nn.init.constant_(self.layer6.bias, 0)
            nn.init.constant_(self.bn6.weight, 1)
            nn.init.constant_(self.bn6.bias, 0)
        elif self.net_stride == 64:
            self.layer5 = nn.Conv2d(2048, 512, kernel_size=3, stride=2, padding=1)
            self.bn5 = nn.BatchNorm2d(512)
            # init
            nn.init.normal_(self.layer5.weight, std=0.001)
            if self.layer5.bias is not None:
                nn.init.constant_(self.layer5.bias, 0)
            nn.init.constant_(self.bn5.weight, 1)
            nn.init.constant_(self.bn5.bias, 0)
        elif self.net_stride == 32:
            pass
        else:
            print('No such net_stride!')
            exit(0)

        self.cls_layer = nn.Conv2d(2048, num_lms, kernel_size=1, stride=1, padding=0)
        self.x_layer = nn.Conv2d(2048, num_lms, kernel_size=1, stride=1, padding=0)
        self.y_layer = nn.Conv2d(2048, num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_x_layer = nn.Conv2d(2048, num_nb*num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_y_layer = nn.Conv2d(2048, num_nb*num_lms, kernel_size=1, stride=1, padding=0)

        nn.init.normal_(self.cls_layer.weight, std=0.001)
        if self.cls_layer.bias is not None:
            nn.init.constant_(self.cls_layer.bias, 0)

        nn.init.normal_(self.x_layer.weight, std=0.001)
        if self.x_layer.bias is not None:
            nn.init.constant_(self.x_layer.bias, 0)

        nn.init.normal_(self.y_layer.weight, std=0.001)
        if self.y_layer.bias is not None:
            nn.init.constant_(self.y_layer.bias, 0)

        nn.init.normal_(self.nb_x_layer.weight, std=0.001)
        if self.nb_x_layer.bias is not None:
            nn.init.constant_(self.nb_x_layer.bias, 0)

        nn.init.normal_(self.nb_y_layer.weight, std=0.001)
        if self.nb_y_layer.bias is not None:
            nn.init.constant_(self.nb_y_layer.bias, 0)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = F.relu(x)
        x = self.maxpool(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        if self.net_stride == 128:
            x = F.relu(self.bn5(self.layer5(x)))
            x = F.relu(self.bn6(self.layer6(x)))
        elif self.net_stride == 64:
            x = F.relu(self.bn5(self.layer5(x)))
        else:
            pass
        x1 = self.cls_layer(x)
        x2 = self.x_layer(x)
        x3 = self.y_layer(x)
        x4 = self.nb_x_layer(x)
        x5 = self.nb_y_layer(x)
        return x1, x2, x3, x4, x5

# net_stride output_size
# 128        2x2
# 64         4x4
# 32         8x8
# pip regression, resnet18
class Pip_resnet18(nn.Module):
    def __init__(self, resnet, num_nb, num_lms=68, input_size=256, net_stride=32):
        super(Pip_resnet18, self).__init__()
        self.num_nb = num_nb
        self.num_lms = num_lms
        self.input_size = input_size
        self.net_stride = net_stride
        self.conv1 = resnet.conv1
        self.bn1 = resnet.bn1
        self.maxpool = resnet.maxpool
        self.sigmoid = nn.Sigmoid()
        self.layer1 = resnet.layer1
        self.layer2 = resnet.layer2
        self.layer3 = resnet.layer3
        self.layer4 = resnet.layer4
        if self.net_stride == 128:
            self.layer5 = nn.Conv2d(512, 512, kernel_size=3, stride=2, padding=1)
            self.bn5 = nn.BatchNorm2d(512)
            self.layer6 = nn.Conv2d(512, 512, kernel_size=3, stride=2, padding=1)
            self.bn6 = nn.BatchNorm2d(512)
            # init
            nn.init.normal_(self.layer5.weight, std=0.001)
            if self.layer5.bias is not None:
                nn.init.constant_(self.layer5.bias, 0)
            nn.init.constant_(self.bn5.weight, 1)
            nn.init.constant_(self.bn5.bias, 0)

            nn.init.normal_(self.layer6.weight, std=0.001)
            if self.layer6.bias is not None:
                nn.init.constant_(self.layer6.bias, 0)
            nn.init.constant_(self.bn6.weight, 1)
            nn.init.constant_(self.bn6.bias, 0)
        elif self.net_stride == 64:
            self.layer5 = nn.Conv2d(512, 512, kernel_size=3, stride=2, padding=1)
            self.bn5 = nn.BatchNorm2d(512)
            # init
            nn.init.normal_(self.layer5.weight, std=0.001)
            if self.layer5.bias is not None:
                nn.init.constant_(self.layer5.bias, 0)
            nn.init.constant_(self.bn5.weight, 1)
            nn.init.constant_(self.bn5.bias, 0)
        elif self.net_stride == 32:
            pass
        elif self.net_stride == 16:
            self.deconv1 = nn.ConvTranspose2d(512, 512, kernel_size=4, stride=2, padding=1, bias=False)
            self.bn_deconv1 = nn.BatchNorm2d(512)
            nn.init.normal_(self.deconv1.weight, std=0.001)
            if self.deconv1.bias is not None:
                nn.init.constant_(self.deconv1.bias, 0)
            nn.init.constant_(self.bn_deconv1.weight, 1)
            nn.init.constant_(self.bn_deconv1.bias, 0)
        else:
            print('No such net_stride!')
            exit(0)

        self.cls_layer = nn.Conv2d(512, num_lms, kernel_size=1, stride=1, padding=0)
        self.x_layer = nn.Conv2d(512, num_lms, kernel_size=1, stride=1, padding=0)
        self.y_layer = nn.Conv2d(512, num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_x_layer = nn.Conv2d(512, num_nb*num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_y_layer = nn.Conv2d(512, num_nb*num_lms, kernel_size=1, stride=1, padding=0)

        nn.init.normal_(self.cls_layer.weight, std=0.001)
        if self.cls_layer.bias is not None:
            nn.init.constant_(self.cls_layer.bias, 0)

        nn.init.normal_(self.x_layer.weight, std=0.001)
        if self.x_layer.bias is not None:
            nn.init.constant_(self.x_layer.bias, 0)

        nn.init.normal_(self.y_layer.weight, std=0.001)
        if self.y_layer.bias is not None:
            nn.init.constant_(self.y_layer.bias, 0)

        nn.init.normal_(self.nb_x_layer.weight, std=0.001)
        if self.nb_x_layer.bias is not None:
            nn.init.constant_(self.nb_x_layer.bias, 0)

        nn.init.normal_(self.nb_y_layer.weight, std=0.001)
        if self.nb_y_layer.bias is not None:
            nn.init.constant_(self.nb_y_layer.bias, 0)

    def forward(self, x):
        x = self.conv1(x)
        x = self.bn1(x)
        x = F.relu(x)
        x = self.maxpool(x)
        x = self.layer1(x)
        x = self.layer2(x)
        x = self.layer3(x)
        x = self.layer4(x)
        if self.net_stride == 128:
            x = F.relu(self.bn5(self.layer5(x)))
            x = F.relu(self.bn6(self.layer6(x)))
        elif self.net_stride == 64:
            x = F.relu(self.bn5(self.layer5(x)))
        elif self.net_stride == 16:
            x = F.relu(self.bn_deconv1(self.deconv1(x)))
        else:
            pass
        x1 = self.cls_layer(x)
        x2 = self.x_layer(x)
        x3 = self.y_layer(x)
        x4 = self.nb_x_layer(x)
        x5 = self.nb_y_layer(x)
        return x1, x2, x3, x4, x5

class Pip_mbnetv2(nn.Module):
    def __init__(self, mbnet, num_nb, num_lms=68, input_size=256, net_stride=32):
        super(Pip_mbnetv2, self).__init__()
        self.num_nb = num_nb
        self.num_lms = num_lms
        self.input_size = input_size
        self.net_stride = net_stride
        self.features = mbnet.features
        self.sigmoid = nn.Sigmoid()

        self.cls_layer = nn.Conv2d(1280, num_lms, kernel_size=1, stride=1, padding=0)
        self.x_layer = nn.Conv2d(1280, num_lms, kernel_size=1, stride=1, padding=0)
        self.y_layer = nn.Conv2d(1280, num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_x_layer = nn.Conv2d(1280, num_nb*num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_y_layer = nn.Conv2d(1280, num_nb*num_lms, kernel_size=1, stride=1, padding=0)

        nn.init.normal_(self.cls_layer.weight, std=0.001)
        if self.cls_layer.bias is not None:
            nn.init.constant_(self.cls_layer.bias, 0)

        nn.init.normal_(self.x_layer.weight, std=0.001)
        if self.x_layer.bias is not None:
            nn.init.constant_(self.x_layer.bias, 0)

        nn.init.normal_(self.y_layer.weight, std=0.001)
        if self.y_layer.bias is not None:
            nn.init.constant_(self.y_layer.bias, 0)

        nn.init.normal_(self.nb_x_layer.weight, std=0.001)
        if self.nb_x_layer.bias is not None:
            nn.init.constant_(self.nb_x_layer.bias, 0)

        nn.init.normal_(self.nb_y_layer.weight, std=0.001)
        if self.nb_y_layer.bias is not None:
            nn.init.constant_(self.nb_y_layer.bias, 0)

    def forward(self, x):
        x = self.features(x)
        x1 = self.cls_layer(x)
        x2 = self.x_layer(x)
        x3 = self.y_layer(x)
        x4 = self.nb_x_layer(x)
        x5 = self.nb_y_layer(x)
        return x1, x2, x3, x4, x5

class Pip_mbnetv3(nn.Module):
    def __init__(self, mbnet, num_nb, num_lms=68, input_size=256, net_stride=32):
        super(Pip_mbnetv3, self).__init__()
        self.num_nb = num_nb
        self.num_lms = num_lms
        self.input_size = input_size
        self.net_stride = net_stride
        self.features = mbnet.features
        self.conv = mbnet.conv
        self.sigmoid = nn.Sigmoid()

        self.cls_layer = nn.Conv2d(960, num_lms, kernel_size=1, stride=1, padding=0)
        self.x_layer = nn.Conv2d(960, num_lms, kernel_size=1, stride=1, padding=0)
        self.y_layer = nn.Conv2d(960, num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_x_layer = nn.Conv2d(960, num_nb*num_lms, kernel_size=1, stride=1, padding=0)
        self.nb_y_layer = nn.Conv2d(960, num_nb*num_lms, kernel_size=1, stride=1, padding=0)

        nn.init.normal_(self.cls_layer.weight, std=0.001)
        if self.cls_layer.bias is not None:
            nn.init.constant_(self.cls_layer.bias, 0)

        nn.init.normal_(self.x_layer.weight, std=0.001)
        if self.x_layer.bias is not None:
            nn.init.constant_(self.x_layer.bias, 0)

        nn.init.normal_(self.y_layer.weight, std=0.001)
        if self.y_layer.bias is not None:
            nn.init.constant_(self.y_layer.bias, 0)

        nn.init.normal_(self.nb_x_layer.weight, std=0.001)
        if self.nb_x_layer.bias is not None:
            nn.init.constant_(self.nb_x_layer.bias, 0)

        nn.init.normal_(self.nb_y_layer.weight, std=0.001)
        if self.nb_y_layer.bias is not None:
            nn.init.constant_(self.nb_y_layer.bias, 0)

    def forward(self, x):
        x = self.features(x)
        x = self.conv(x)
        x1 = self.cls_layer(x)
        x2 = self.x_layer(x)
        x3 = self.y_layer(x)
        x4 = self.nb_x_layer(x)
        x5 = self.nb_y_layer(x)
        return x1, x2, x3, x4, x5
    
    
if __name__ == '__main__':
    pass