Myosotis-Researches 0.0.12__py3-none-any.whl → 0.0.14__py3-none-any.whl

myosotis_researches/CcGAN/models_128/ResNet_regre_eval.py

@@ -0,0 +1,175 @@
+'''
+codes are based on
+@article{
+zhang2018mixup,
+title={mixup: Beyond Empirical Risk Minimization},
+author={Hongyi Zhang, Moustapha Cisse, Yann N. Dauphin, David Lopez-Paz},
+journal={International Conference on Learning Representations},
+year={2018},
+url={https://openreview.net/forum?id=r1Ddp1-Rb},
+}
+'''
+
+
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+NC = 3
+IMG_SIZE = 128
+
+
+class BasicBlock(nn.Module):
+    expansion = 1
+
+    def __init__(self, in_planes, planes, stride=1):
+        super(BasicBlock, self).__init__()
+        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=1, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+
+        self.shortcut = nn.Sequential()
+        if stride != 1 or in_planes != self.expansion*planes:
+            self.shortcut = nn.Sequential(
+                nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(self.expansion*planes)
+            )
+
+    def forward(self, x):
+        out = F.relu(self.bn1(self.conv1(x)))
+        out = self.bn2(self.conv2(out))
+        out += self.shortcut(x)
+        out = F.relu(out)
+        return out
+
+
+class Bottleneck(nn.Module):
+    expansion = 4
+
+    def __init__(self, in_planes, planes, stride=1):
+        super(Bottleneck, self).__init__()
+        self.conv1 = nn.Conv2d(in_planes, planes, kernel_size=1, bias=False)
+        self.bn1 = nn.BatchNorm2d(planes)
+        self.conv2 = nn.Conv2d(planes, planes, kernel_size=3, stride=stride, padding=1, bias=False)
+        self.bn2 = nn.BatchNorm2d(planes)
+        self.conv3 = nn.Conv2d(planes, self.expansion*planes, kernel_size=1, bias=False)
+        self.bn3 = nn.BatchNorm2d(self.expansion*planes)
+
+        self.shortcut = nn.Sequential()
+        if stride != 1 or in_planes != self.expansion*planes:
+            self.shortcut = nn.Sequential(
+                nn.Conv2d(in_planes, self.expansion*planes, kernel_size=1, stride=stride, bias=False),
+                nn.BatchNorm2d(self.expansion*planes)
+            )
+
+    def forward(self, x):
+        out = F.relu(self.bn1(self.conv1(x)))
+        out = F.relu(self.bn2(self.conv2(out)))
+        out = self.bn3(self.conv3(out))
+        out += self.shortcut(x)
+        out = F.relu(out)
+        return out
+
+
+class ResNet_regre_eval(nn.Module):
+    def __init__(self, block, num_blocks, nc=NC, ngpu = 1, feature_layer='f3'):
+        super(ResNet_regre_eval, self).__init__()
+        self.in_planes = 64
+        self.ngpu = ngpu
+        self.feature_layer=feature_layer
+
+        self.block1 = nn.Sequential(
+            nn.Conv2d(nc, 64, kernel_size=3, stride=1, padding=1, bias=False),  # h=h
+            nn.BatchNorm2d(64),
+            nn.ReLU(),
+            nn.MaxPool2d(2,2), #h=h/2 64
+            self._make_layer(block, 64, num_blocks[0], stride=2),  # h=h/2 32
+        )
+        self.block2 = self._make_layer(block, 128, num_blocks[1], stride=2) # h=h/2 16
+        self.block3 = self._make_layer(block, 256, num_blocks[2], stride=2) # h=h/2 8
+        self.block4 = self._make_layer(block, 512, num_blocks[3], stride=2) # h=h/2 4
+
+        self.pool1 = nn.AvgPool2d(kernel_size=4)
+        if self.feature_layer == 'f2':
+            self.pool2 = nn.AdaptiveAvgPool2d((2,2))
+        elif self.feature_layer == 'f3':
+            self.pool2 = nn.AdaptiveAvgPool2d((2,2))
+        else:
+            self.pool2 = nn.AdaptiveAvgPool2d((1,1))
+
+        linear_layers = [
+                nn.Linear(512*block.expansion, 128),
+                nn.BatchNorm1d(128),
+                nn.ReLU(),
+                nn.Linear(128, 128),
+                nn.BatchNorm1d(128),
+                nn.ReLU(),
+                nn.Linear(128, 1),
+                # nn.Sigmoid()
+                nn.ReLU(),
+            ]
+        self.linear = nn.Sequential(*linear_layers)
+
+
+    def _make_layer(self, block, planes, num_blocks, stride):
+        strides = [stride] + [1]*(num_blocks-1)
+        layers = []
+        for stride in strides:
+            layers.append(block(self.in_planes, planes, stride))
+            self.in_planes = planes * block.expansion
+        return nn.Sequential(*layers)
+
+    def forward(self, x):
+
+        if x.is_cuda and self.ngpu > 1:
+            ft1 = nn.parallel.data_parallel(self.block1, x, range(self.ngpu))
+            ft2 = nn.parallel.data_parallel(self.block2, ft1, range(self.ngpu))
+            ft3 = nn.parallel.data_parallel(self.block3, ft2, range(self.ngpu))
+            ft4 = nn.parallel.data_parallel(self.block4, ft3, range(self.ngpu))
+            out = nn.parallel.data_parallel(self.pool1, ft4, range(self.ngpu))
+            out = out.view(out.size(0), -1)
+            out = nn.parallel.data_parallel(self.linear, out, range(self.ngpu))
+        else:
+            ft1 = self.block1(x)
+            ft2 = self.block2(ft1)
+            ft3 = self.block3(ft2)
+            ft4 = self.block4(ft3)
+            out = self.pool1(ft4)
+            out = out.view(out.size(0), -1)
+            out = self.linear(out)
+
+        if self.feature_layer == 'f2':
+            ext_features = self.pool2(ft2)
+        elif self.feature_layer == 'f3':
+            ext_features = self.pool2(ft3)
+        else:
+            ext_features = self.pool2(ft4)
+
+        ext_features = ext_features.view(ext_features.size(0), -1)
+
+        return out, ext_features
+
+
+def ResNet18_regre_eval(ngpu = 1):
+    return ResNet_regre_eval(BasicBlock, [2,2,2,2], ngpu = ngpu)
+
+def ResNet34_regre_eval(ngpu = 1):
+    return ResNet_regre_eval(BasicBlock, [3,4,6,3], ngpu = ngpu)
+
+def ResNet50_regre_eval(ngpu = 1):
+    return ResNet_regre_eval(Bottleneck, [3,4,6,3], ngpu = ngpu)
+
+def ResNet101_regre_eval(ngpu = 1):
+    return ResNet_regre_eval(Bottleneck, [3,4,23,3], ngpu = ngpu)
+
+def ResNet152_regre_eval(ngpu = 1):
+    return ResNet_regre_eval(Bottleneck, [3,8,36,3], ngpu = ngpu)
+
+
+if __name__ == "__main__":
+    net = ResNet34_regre_eval(ngpu = 1).cuda()
+    x = torch.randn(4,NC,IMG_SIZE,IMG_SIZE).cuda()
+    out, features = net(x)
+    print(out.size())
+    print(features.size())

myosotis_researches/CcGAN/models_128/__init__.py

@@ -0,0 +1,8 @@
+from .autoencoder import *
+from .cGAN_SAGAN import cGAN_SAGAN_Generator, cGAN_SAGAN_Discriminator
+from .cGAN_concat_SAGAN import cGAN_concat_SAGAN_Generator, cGAN_concat_SAGAN_Discriminator
+from .CcGAN_SAGAN import CcGAN_SAGAN_Generator, CcGAN_SAGAN_Discriminator
+from .ResNet_embed import ResNet18_embed, ResNet34_embed, ResNet50_embed, model_y2h
+from .ResNet_regre_eval import ResNet34_regre_eval
+from .ResNet_class_eval import ResNet34_class_eval
+

myosotis_researches/CcGAN/models_128/autoencoder.py

@@ -0,0 +1,119 @@
+import torch
+from torch import nn
+
+
+
+class encoder(nn.Module):
+    def __init__(self, dim_bottleneck=512, ch=64):
+        super(encoder, self).__init__()
+        self.ch = ch
+        self.dim_bottleneck = dim_bottleneck
+
+        self.conv = nn.Sequential(
+            nn.Conv2d(3, ch, kernel_size=4, stride=2, padding=1), #h=h/2; 64
+            nn.BatchNorm2d(ch),
+            nn.ReLU(),
+            nn.MaxPool2d(2,2), #h=h/2; 32
+            nn.Conv2d(ch, ch, kernel_size=3, stride=1, padding=1), #h=h
+            nn.BatchNorm2d(ch),
+            nn.ReLU(),
+
+            nn.Conv2d(ch, ch*2, kernel_size=4, stride=2, padding=1), #h=h/2; 16
+            nn.BatchNorm2d(ch*2),
+            nn.ReLU(),
+            nn.Conv2d(ch*2, ch*2, kernel_size=3, stride=1, padding=1), #h=h
+            nn.BatchNorm2d(ch*2),
+            nn.ReLU(),
+
+            nn.Conv2d(ch*2, ch*4, kernel_size=4, stride=2, padding=1), #h=h/2; 8
+            nn.BatchNorm2d(ch*4),
+            nn.ReLU(),
+            nn.Conv2d(ch*4, ch*4, kernel_size=3, stride=1, padding=1), #h=h
+            nn.BatchNorm2d(ch*4),
+            nn.ReLU(),
+
+            nn.Conv2d(ch*4, ch*8, kernel_size=4, stride=2, padding=1), #h=h/2; 4
+            nn.BatchNorm2d(ch*8),
+            nn.ReLU(),
+            nn.Conv2d(ch*8, ch*8, kernel_size=3, stride=1, padding=1), #h=h
+            nn.BatchNorm2d(ch*8),
+            nn.ReLU(),
+        )
+
+        self.linear = nn.Sequential(
+            nn.Linear(ch*8*4*4, dim_bottleneck),
+            # nn.ReLU()
+        )
+
+    def forward(self, x):
+        feature = self.conv(x)
+        feature = feature.view(-1, self.ch*8*4*4)
+        feature = self.linear(feature)
+        return feature
+
+
+
+class decoder(nn.Module):
+    def __init__(self, dim_bottleneck=512, ch=64):
+        super(decoder, self).__init__()
+        self.ch = ch
+        self.dim_bottleneck = dim_bottleneck
+
+        self.linear = nn.Sequential(
+            nn.Linear(dim_bottleneck, ch*16*4*4)
+        )
+
+        self.deconv = nn.Sequential(
+            nn.ConvTranspose2d(ch*16, ch*8, kernel_size=4, stride=2, padding=1), #h=2h; 8
+            nn.BatchNorm2d(ch*8),
+            nn.ReLU(True),
+            nn.Conv2d(ch*8, ch*8, kernel_size=3, stride=1, padding=1), #h=h
+            nn.BatchNorm2d(ch*8),
+            nn.ReLU(),
+
+            nn.ConvTranspose2d(ch*8, ch*4, kernel_size=4, stride=2, padding=1), #h=2h; 16
+            nn.BatchNorm2d(ch*4),
+            nn.ReLU(True),
+            nn.Conv2d(ch*4, ch*4, kernel_size=3, stride=1, padding=1), #h=h
+            nn.BatchNorm2d(ch*4),
+            nn.ReLU(),
+
+            nn.ConvTranspose2d(ch*4, ch*2, kernel_size=4, stride=2, padding=1), #h=2h; 32
+            nn.BatchNorm2d(ch*2),
+            nn.ReLU(True),
+            nn.Conv2d(ch*2, ch*2, kernel_size=3, stride=1, padding=1), #h=h
+            nn.BatchNorm2d(ch*2),
+            nn.ReLU(),
+
+            nn.ConvTranspose2d(ch*2, ch, kernel_size=4, stride=2, padding=1), #h=2h; 64
+            nn.BatchNorm2d(ch),
+            nn.ReLU(True),
+            nn.Conv2d(ch, ch, kernel_size=3, stride=1, padding=1), #h=h
+            nn.BatchNorm2d(ch),
+            nn.ReLU(),
+
+            nn.ConvTranspose2d(ch, ch, kernel_size=4, stride=2, padding=1), #h=2h; 128
+            nn.BatchNorm2d(ch),
+            nn.ReLU(True),
+            nn.Conv2d(ch, 3, kernel_size=3, stride=1, padding=1), #h=h
+            nn.Tanh()
+        )
+
+    def forward(self, feature):
+        out = self.linear(feature)
+        out = out.view(-1, self.ch*16, 4, 4)
+        out = self.deconv(out)
+        return out
+
+
+if __name__=="__main__":
+    #test
+
+    net_encoder = encoder(dim_bottleneck=512, ch=64).cuda()
+    net_decoder = decoder(dim_bottleneck=512, ch=64).cuda()
+
+    x = torch.randn(10, 3, 128, 128).cuda()
+    f = net_encoder(x)
+    xh = net_decoder(f)
+    print(f.size())
+    print(xh.size())

myosotis_researches/CcGAN/models_128/cGAN_SAGAN.py

@@ -0,0 +1,276 @@
+import numpy as np
+import torch
+import torch.nn as nn
+import torch.nn.functional as F
+
+from torch.nn.utils import spectral_norm
+from torch.nn.init import xavier_uniform_
+
+
+def init_weights(m):
+    if type(m) == nn.Linear or type(m) == nn.Conv2d:
+        xavier_uniform_(m.weight)
+        m.bias.data.fill_(0.)
+
+
+def snconv2d(in_channels, out_channels, kernel_size, stride=1, padding=0, dilation=1, groups=1, bias=True):
+    return spectral_norm(nn.Conv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=kernel_size,
+                                   stride=stride, padding=padding, dilation=dilation, groups=groups, bias=bias))
+
+
+def snlinear(in_features, out_features):
+    return spectral_norm(nn.Linear(in_features=in_features, out_features=out_features))
+
+
+def sn_embedding(num_embeddings, embedding_dim):
+    return spectral_norm(nn.Embedding(num_embeddings=num_embeddings, embedding_dim=embedding_dim))
+
+
+class Self_Attn(nn.Module):
+    """ Self attention Layer"""
+
+    def __init__(self, in_channels):
+        super(Self_Attn, self).__init__()
+        self.in_channels = in_channels
+        self.snconv1x1_theta = snconv2d(in_channels=in_channels, out_channels=in_channels//8, kernel_size=1, stride=1, padding=0)
+        self.snconv1x1_phi = snconv2d(in_channels=in_channels, out_channels=in_channels//8, kernel_size=1, stride=1, padding=0)
+        self.snconv1x1_g = snconv2d(in_channels=in_channels, out_channels=in_channels//2, kernel_size=1, stride=1, padding=0)
+        self.snconv1x1_attn = snconv2d(in_channels=in_channels//2, out_channels=in_channels, kernel_size=1, stride=1, padding=0)
+        self.maxpool = nn.MaxPool2d(2, stride=2, padding=0)
+        self.softmax  = nn.Softmax(dim=-1)
+        self.sigma = nn.Parameter(torch.zeros(1))
+
+    def forward(self, x):
+        """
+            inputs :
+                x : input feature maps(B X C X W X H)
+            returns :
+                out : self attention value + input feature 
+                attention: B X N X N (N is Width*Height)
+        """
+        _, ch, h, w = x.size()
+        # Theta path
+        theta = self.snconv1x1_theta(x)
+        theta = theta.view(-1, ch//8, h*w)
+        # Phi path
+        phi = self.snconv1x1_phi(x)
+        phi = self.maxpool(phi)
+        phi = phi.view(-1, ch//8, h*w//4)
+        # Attn map
+        attn = torch.bmm(theta.permute(0, 2, 1), phi)
+        attn = self.softmax(attn)
+        # g path
+        g = self.snconv1x1_g(x)
+        g = self.maxpool(g)
+        g = g.view(-1, ch//2, h*w//4)
+        # Attn_g
+        attn_g = torch.bmm(g, attn.permute(0, 2, 1))
+        attn_g = attn_g.view(-1, ch//2, h, w)
+        attn_g = self.snconv1x1_attn(attn_g)
+        # Out
+        out = x + self.sigma*attn_g
+        return out
+
+
+class ConditionalBatchNorm2d(nn.Module):
+    # https://github.com/pytorch/pytorch/issues/8985#issuecomment-405080775
+    def __init__(self, num_features, num_classes):
+        super().__init__()
+        self.num_features = num_features
+        self.bn = nn.BatchNorm2d(num_features, momentum=0.001, affine=False)
+        self.embed = nn.Embedding(num_classes, num_features * 2)
+        # self.embed.weight.data[:, :num_features].normal_(1, 0.02)  # Initialise scale at N(1, 0.02)
+        self.embed.weight.data[:, :num_features].fill_(1.)  # Initialize scale to 1
+        self.embed.weight.data[:, num_features:].zero_()  # Initialize bias at 0
+
+    def forward(self, x, y):
+        out = self.bn(x)
+        gamma, beta = self.embed(y).chunk(2, 1)
+        out = gamma.view(-1, self.num_features, 1, 1) * out + beta.view(-1, self.num_features, 1, 1)
+        return out
+
+
+class GenBlock(nn.Module):
+    def __init__(self, in_channels, out_channels, num_classes):
+        super(GenBlock, self).__init__()
+        self.cond_bn1 = ConditionalBatchNorm2d(in_channels, num_classes)
+        self.relu = nn.ReLU(inplace=True)
+        self.snconv2d1 = snconv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1)
+        self.cond_bn2 = ConditionalBatchNorm2d(out_channels, num_classes)
+        self.snconv2d2 = snconv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1)
+        self.snconv2d0 = snconv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x, labels):
+        x0 = x
+
+        x = self.cond_bn1(x, labels)
+        x = self.relu(x)
+        x = F.interpolate(x, scale_factor=2, mode='nearest') # upsample
+        x = self.snconv2d1(x)
+        x = self.cond_bn2(x, labels)
+        x = self.relu(x)
+        x = self.snconv2d2(x)
+
+        x0 = F.interpolate(x0, scale_factor=2, mode='nearest') # upsample
+        x0 = self.snconv2d0(x0)
+
+        out = x + x0
+        return out
+
+
+class cGAN_SAGAN_Generator(nn.Module):
+    """Generator."""
+
+    def __init__(self, z_dim, num_classes, g_conv_dim=64):
+        super(cGAN_SAGAN_Generator, self).__init__()
+
+        self.z_dim = z_dim
+        self.g_conv_dim = g_conv_dim
+        self.snlinear0 = snlinear(in_features=z_dim, out_features=g_conv_dim*16*4*4)
+        self.block1 = GenBlock(g_conv_dim*16, g_conv_dim*16, num_classes)
+        self.block2 = GenBlock(g_conv_dim*16, g_conv_dim*8, num_classes)
+        self.block3 = GenBlock(g_conv_dim*8, g_conv_dim*4, num_classes)
+        self.self_attn = Self_Attn(g_conv_dim*4)
+        self.block4 = GenBlock(g_conv_dim*4, g_conv_dim*2, num_classes)
+        self.block5 = GenBlock(g_conv_dim*2, g_conv_dim, num_classes)
+        self.bn = nn.BatchNorm2d(g_conv_dim, eps=1e-5, momentum=0.0001, affine=True)
+        self.relu = nn.ReLU(inplace=True)
+        self.snconv2d1 = snconv2d(in_channels=g_conv_dim, out_channels=3, kernel_size=3, stride=1, padding=1)
+        self.tanh = nn.Tanh()
+
+        # Weight init
+        self.apply(init_weights)
+
+    def forward(self, z, labels):
+        # n x z_dim
+        act0 = self.snlinear0(z)            # n x g_conv_dim*16*4*4
+        act0 = act0.view(-1, self.g_conv_dim*16, 4, 4) # n x g_conv_dim*16 x 4 x 4
+        act1 = self.block1(act0, labels)    # n x g_conv_dim*16 x 8 x 8
+        act2 = self.block2(act1, labels)    # n x g_conv_dim*8 x 16 x 16
+        act3 = self.block3(act2, labels)    # n x g_conv_dim*4 x 32 x 32
+        act3 = self.self_attn(act3)         # n x g_conv_dim*4 x 32 x 32
+        act4 = self.block4(act3, labels)    # n x g_conv_dim*2 x 64 x 64
+        act5 = self.block5(act4, labels)    # n x g_conv_dim  x 128 x 128
+        act5 = self.bn(act5)                # n x g_conv_dim  x 128 x 128
+        act5 = self.relu(act5)              # n x g_conv_dim  x 128 x 128
+        act6 = self.snconv2d1(act5)         # n x 3 x 128 x 128
+        act6 = self.tanh(act6)              # n x 3 x 128 x 128
+        return act6
+
+
+class DiscOptBlock(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(DiscOptBlock, self).__init__()
+        self.snconv2d1 = snconv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1)
+        self.relu = nn.ReLU(inplace=True)
+        self.snconv2d2 = snconv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1)
+        self.downsample = nn.AvgPool2d(2)
+        self.snconv2d0 = snconv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x):
+        x0 = x
+
+        x = self.snconv2d1(x)
+        x = self.relu(x)
+        x = self.snconv2d2(x)
+        x = self.downsample(x)
+
+        x0 = self.downsample(x0)
+        x0 = self.snconv2d0(x0)
+
+        out = x + x0
+        return out
+
+
+class DiscBlock(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(DiscBlock, self).__init__()
+        self.relu = nn.ReLU(inplace=True)
+        self.snconv2d1 = snconv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1)
+        self.snconv2d2 = snconv2d(in_channels=out_channels, out_channels=out_channels, kernel_size=3, stride=1, padding=1)
+        self.downsample = nn.AvgPool2d(2)
+        self.ch_mismatch = False
+        if in_channels != out_channels:
+            self.ch_mismatch = True
+        self.snconv2d0 = snconv2d(in_channels=in_channels, out_channels=out_channels, kernel_size=1, stride=1, padding=0)
+
+    def forward(self, x, downsample=True):
+        x0 = x
+
+        x = self.relu(x)
+        x = self.snconv2d1(x)
+        x = self.relu(x)
+        x = self.snconv2d2(x)
+        if downsample:
+            x = self.downsample(x)
+
+        if downsample or self.ch_mismatch:
+            x0 = self.snconv2d0(x0)
+            if downsample:
+                x0 = self.downsample(x0)
+
+        out = x + x0
+        return out
+
+
+class cGAN_SAGAN_Discriminator(nn.Module):
+    """Discriminator."""
+
+    def __init__(self, num_classes, d_conv_dim=64):
+        super(cGAN_SAGAN_Discriminator, self).__init__()
+        self.d_conv_dim = d_conv_dim
+        self.opt_block1 = DiscOptBlock(3, d_conv_dim)
+        self.block1 = DiscBlock(d_conv_dim, d_conv_dim*2)
+        self.self_attn = Self_Attn(d_conv_dim*2)
+        self.block2 = DiscBlock(d_conv_dim*2, d_conv_dim*4)
+        self.block3 = DiscBlock(d_conv_dim*4, d_conv_dim*8)
+        self.block4 = DiscBlock(d_conv_dim*8, d_conv_dim*16)
+        self.block5 = DiscBlock(d_conv_dim*16, d_conv_dim*16)
+        self.relu = nn.ReLU(inplace=True)
+        self.snlinear1 = snlinear(in_features=d_conv_dim*16, out_features=1)
+        self.sn_embedding1 = sn_embedding(num_classes, d_conv_dim*16)
+
+        # Weight init
+        self.apply(init_weights)
+        xavier_uniform_(self.sn_embedding1.weight)
+
+    def forward(self, x, labels):
+        # n x 3 x 128 x 128
+        h0 = self.opt_block1(x) # n x d_conv_dim   x 64 x 64
+        h1 = self.block1(h0)    # n x d_conv_dim*2 x 32 x 32
+        h1 = self.self_attn(h1) # n x d_conv_dim*2 x 32 x 32
+        h2 = self.block2(h1)    # n x d_conv_dim*4 x 16 x 16
+        h3 = self.block3(h2)    # n x d_conv_dim*8 x  8 x  8
+        h4 = self.block4(h3)    # n x d_conv_dim*16 x 4 x  4
+        h5 = self.block5(h4, downsample=False)  # n x d_conv_dim*16 x 4 x 4
+        h5 = self.relu(h5)              # n x d_conv_dim*16 x 4 x 4
+        h6 = torch.sum(h5, dim=[2,3])   # n x d_conv_dim*16
+        output1 = torch.squeeze(self.snlinear1(h6)) # n
+        # Projection
+        h_labels = self.sn_embedding1(labels)   # n x d_conv_dim*16
+        proj = torch.mul(h6, h_labels)          # n x d_conv_dim*16
+        output2 = torch.sum(proj, dim=[1])      # n
+        # Out
+        output = output1 + output2              # n
+        return output
+
+
+
+if __name__ == "__main__":
+    
+    num_classes = 10
+
+    netG = cGAN_SAGAN_Generator(z_dim=128, num_classes=num_classes, g_conv_dim=128).cuda()
+    netD = cGAN_SAGAN_Discriminator(num_classes=num_classes, d_conv_dim=128).cuda()
+
+    n = 4
+    # target = torch.randint(high=num_classes, size=(1,n)) # set size (2,10) for MHE
+    # y = torch.zeros(n, num_classes)
+    # y[range(y.shape[0]), target]=1
+    # y = y.type(torch.long).cuda()
+    y = torch.randint(high=num_classes, size=(n,)).type(torch.long).cuda()
+    z = torch.randn(n, 128).cuda()
+    x = netG(z,y)
+    o = netD(x,y)
+    print(x.size())
+    print(o.size())