Myosotis-Researches 0.0.13__py3-none-any.whl → 0.0.15__py3-none-any.whl

myosotis_researches/CcGAN/models_256/cGAN_SAGAN.py

@@ -0,0 +1,280 @@
+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.block4 = GenBlock(g_conv_dim*4, g_conv_dim*2, num_classes)
+        self.self_attn = Self_Attn(g_conv_dim*2)
+        self.block5 = GenBlock(g_conv_dim*2, g_conv_dim*2, num_classes)
+        self.block6 = 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
+        act4 = self.block4(act3, labels)    # n x g_conv_dim*2 x 64 x 64
+        act4 = self.self_attn(act4)         # n x g_conv_dim*2 x 64 x 64
+        act5 = self.block5(act4, labels)    # n x g_conv_dim  x 128 x 128
+        act6 = self.block6(act5, labels)    # n x g_conv_dim  x 256 x 256
+        act6 = self.bn(act6)                # n x g_conv_dim  x 256 x 256
+        act6 = self.relu(act6)              # n x g_conv_dim  x 256 x 256
+        act7 = self.snconv2d1(act6)         # n x 3 x 256 x 256
+        act7 = self.tanh(act7)              # n x 3 x 256 x 256
+        return act7
+
+
+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*6)
+        self.block4 = DiscBlock(d_conv_dim*6, d_conv_dim*12)
+        self.block5 = DiscBlock(d_conv_dim*12, d_conv_dim*12)
+        self.block6 = DiscBlock(d_conv_dim*12, 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 256 x 256
+        h0 = self.opt_block1(x) # n x d_conv_dim   x 128 x 128
+        h1 = self.block1(h0)    # n x d_conv_dim*2 x 64 x 64
+        h1 = self.self_attn(h1) # n x d_conv_dim*2 x 64 x 64
+        h2 = self.block2(h1)    # n x d_conv_dim*4 x 32 x 32
+        h3 = self.block3(h2)    # n x d_conv_dim*8 x 16 x 16
+        h4 = self.block4(h3)    # n x d_conv_dim*16 x 8 x  8
+        h5 = self.block5(h4)    # n x d_conv_dim*16 x 4 x 4
+        h6 = self.block6(h5, downsample=False)  # n x d_conv_dim*16 x 4 x 4
+        h6 = self.relu(h6)              # n x d_conv_dim*16 x 4 x 4
+        h6 = torch.sum(h6, 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())

myosotis_researches/CcGAN/models_256/cGAN_concat_SAGAN.py

@@ -0,0 +1,249 @@
+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 GenBlock(nn.Module):
+    def __init__(self, in_channels, out_channels):
+        super(GenBlock, self).__init__()
+        self.cond_bn1 = nn.BatchNorm2d(in_channels)
+        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 = nn.BatchNorm2d(out_channels)
+        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):
+        x0 = x
+
+        x = self.cond_bn1(x)
+        x = self.relu(x)
+        x = F.interpolate(x, scale_factor=2, mode='nearest') # upsample
+        x = self.snconv2d1(x)
+        x = self.cond_bn2(x)
+        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_concat_SAGAN_Generator(nn.Module):
+    """Generator."""
+
+    def __init__(self, z_dim, dim_c=1, g_conv_dim=64):
+        super(cGAN_concat_SAGAN_Generator, self).__init__()
+
+        self.z_dim = z_dim
+        self.dim_c = dim_c
+        self.g_conv_dim = g_conv_dim
+        self.snlinear0 = snlinear(in_features=z_dim+dim_c, out_features=g_conv_dim*16*4*4)
+        self.block1 = GenBlock(g_conv_dim*16, g_conv_dim*16)
+        self.block2 = GenBlock(g_conv_dim*16, g_conv_dim*8)
+        self.block3 = GenBlock(g_conv_dim*8, g_conv_dim*4)
+        self.block4 = GenBlock(g_conv_dim*4, g_conv_dim*2)
+        self.self_attn = Self_Attn(g_conv_dim*2)
+        self.block5 = GenBlock(g_conv_dim*2, g_conv_dim*2)
+        self.block6 = GenBlock(g_conv_dim*2, g_conv_dim)
+        self.bn = nn.BatchNorm2d(g_conv_dim, eps=1e-5, momentum=0.0001, affine=True)
+        self.relu = nn.ReLU()
+        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(torch.cat((z, labels.view(-1,1)),dim=1))            # 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)    # n x g_conv_dim*16 x 8 x 8
+        act2 = self.block2(act1)    # n x g_conv_dim*8 x 16 x 16
+        act3 = self.block3(act2)    # n x g_conv_dim*4 x 32 x 32
+        act4 = self.block4(act3)    # n x g_conv_dim*2 x 64 x 64
+        act4 = self.self_attn(act4)         # n x g_conv_dim*2 x 64 x 64
+        act5 = self.block5(act4)    # n x g_conv_dim  x 128 x 128
+        act6 = self.block6(act5)    # n x g_conv_dim  x 256 x 256
+        act6 = self.bn(act6)                # n x g_conv_dim  x 256 x 256
+        act6 = self.relu(act6)              # n x g_conv_dim  x 256 x 256
+        act7 = self.snconv2d1(act6)         # n x 3 x 256 x 256
+        act7 = self.tanh(act7)              # n x 3 x 256 x 256
+        return act7
+
+
+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_concat_SAGAN_Discriminator(nn.Module):
+    """Discriminator."""
+
+    def __init__(self, dim_c=1, d_conv_dim=64):
+        super(cGAN_concat_SAGAN_Discriminator, self).__init__()
+        self.d_conv_dim = d_conv_dim
+        self.dim_c = dim_c
+        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*6)
+        self.block4 = DiscBlock(d_conv_dim*6, d_conv_dim*12)
+        self.block5 = DiscBlock(d_conv_dim*12, d_conv_dim*12)
+        self.block6 = DiscBlock(d_conv_dim*12, d_conv_dim*16)
+        self.relu = nn.ReLU()
+        self.snlinear1 = snlinear(in_features=d_conv_dim*16*4*4+dim_c, out_features=1)
+
+    def forward(self, x, labels):
+        labels = labels.view(-1,1)
+        # n x 3 x 256 x 256
+        h0 = self.opt_block1(x) # n x d_conv_dim   x 128 x 128
+        h1 = self.block1(h0)    # n x d_conv_dim*2 x 64 x 64
+        h1 = self.self_attn(h1) # n x d_conv_dim*2 x 64 x 64
+        h2 = self.block2(h1)    # n x d_conv_dim*4 x 32 x 32
+        h3 = self.block3(h2)    # n x d_conv_dim*8 x 16 x 16
+        h4 = self.block4(h3)    # n x d_conv_dim*16 x 8 x  8
+        h5 = self.block5(h4)    # n x d_conv_dim*16 x 4 x 4
+        h6 = self.block6(h5, downsample=False)  # n x d_conv_dim*16 x 4 x 4
+        out = self.relu(h6)              # n x d_conv_dim*16 x 4 x 4
+        out = out.view(-1,self.d_conv_dim*16*4*4)
+        out = torch.cat((out, labels),dim=1)
+        out = self.snlinear1(out)
+        return out
+
+
+
+if __name__ == "__main__":
+    
+    netG = cGAN_concat_SAGAN_Generator(z_dim=128).cuda()
+    netD = cGAN_concat_SAGAN_Discriminator().cuda()
+
+    n = 4
+    y = torch.randn(n, 1).cuda()
+    z = torch.randn(n, 128).cuda()
+    x = netG(z,y)
+    print(x.size())
+    o = netD(x,y)
+    print(o.size())

myosotis_researches/CcGAN/train_128/DiffAugment_pytorch.py

@@ -0,0 +1,76 @@
+# Differentiable Augmentation for Data-Efficient GAN Training
+# Shengyu Zhao, Zhijian Liu, Ji Lin, Jun-Yan Zhu, and Song Han
+# https://arxiv.org/pdf/2006.10738
+
+import torch
+import torch.nn.functional as F
+
+
+def DiffAugment(x, policy='', channels_first=True):
+    if policy:
+        if not channels_first:
+            x = x.permute(0, 3, 1, 2)
+        for p in policy.split(','):
+            for f in AUGMENT_FNS[p]:
+                x = f(x)
+        if not channels_first:
+            x = x.permute(0, 2, 3, 1)
+        x = x.contiguous()
+    return x
+
+
+def rand_brightness(x):
+    x = x + (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) - 0.5)
+    return x
+
+
+def rand_saturation(x):
+    x_mean = x.mean(dim=1, keepdim=True)
+    x = (x - x_mean) * (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) * 2) + x_mean
+    return x
+
+
+def rand_contrast(x):
+    x_mean = x.mean(dim=[1, 2, 3], keepdim=True)
+    x = (x - x_mean) * (torch.rand(x.size(0), 1, 1, 1, dtype=x.dtype, device=x.device) + 0.5) + x_mean
+    return x
+
+
+def rand_translation(x, ratio=0.125):
+    shift_x, shift_y = int(x.size(2) * ratio + 0.5), int(x.size(3) * ratio + 0.5)
+    translation_x = torch.randint(-shift_x, shift_x + 1, size=[x.size(0), 1, 1], device=x.device)
+    translation_y = torch.randint(-shift_y, shift_y + 1, size=[x.size(0), 1, 1], device=x.device)
+    grid_batch, grid_x, grid_y = torch.meshgrid(
+        torch.arange(x.size(0), dtype=torch.long, device=x.device),
+        torch.arange(x.size(2), dtype=torch.long, device=x.device),
+        torch.arange(x.size(3), dtype=torch.long, device=x.device),
+    )
+    grid_x = torch.clamp(grid_x + translation_x + 1, 0, x.size(2) + 1)
+    grid_y = torch.clamp(grid_y + translation_y + 1, 0, x.size(3) + 1)
+    x_pad = F.pad(x, [1, 1, 1, 1, 0, 0, 0, 0])
+    x = x_pad.permute(0, 2, 3, 1).contiguous()[grid_batch, grid_x, grid_y].permute(0, 3, 1, 2)
+    return x
+
+
+def rand_cutout(x, ratio=0.5):
+    cutout_size = int(x.size(2) * ratio + 0.5), int(x.size(3) * ratio + 0.5)
+    offset_x = torch.randint(0, x.size(2) + (1 - cutout_size[0] % 2), size=[x.size(0), 1, 1], device=x.device)
+    offset_y = torch.randint(0, x.size(3) + (1 - cutout_size[1] % 2), size=[x.size(0), 1, 1], device=x.device)
+    grid_batch, grid_x, grid_y = torch.meshgrid(
+        torch.arange(x.size(0), dtype=torch.long, device=x.device),
+        torch.arange(cutout_size[0], dtype=torch.long, device=x.device),
+        torch.arange(cutout_size[1], dtype=torch.long, device=x.device),
+    )
+    grid_x = torch.clamp(grid_x + offset_x - cutout_size[0] // 2, min=0, max=x.size(2) - 1)
+    grid_y = torch.clamp(grid_y + offset_y - cutout_size[1] // 2, min=0, max=x.size(3) - 1)
+    mask = torch.ones(x.size(0), x.size(2), x.size(3), dtype=x.dtype, device=x.device)
+    mask[grid_batch, grid_x, grid_y] = 0
+    x = x * mask.unsqueeze(1)
+    return x
+
+
+AUGMENT_FNS = {
+    'color': [rand_brightness, rand_saturation, rand_contrast],
+    'translation': [rand_translation],
+    'cutout': [rand_cutout],
+}