Super-resolution with SRGAN and ESRGAN
SRGAN SRGAN is an algorithm that uses a neural network to increase the resolution of images, and this time I implemented it. reference https://qiita.com/pacifinapacific/items/ec338a500015ae8c33fe https://buildersbox.corp-sansan.com/entry/2019/04/29/110000
Implemented for the time being
github https://github.com/AokiMasataka/Super-resolution The dataset uses the same SRResNet that I created a long time ago. SResNet article https://qiita.com/AokiMasataka/items/3d382310d8a78f711c71 The network will be implemented in PyTorch as well as practicing PyTorch. SRGAN's Generator network consists of ResNet + Pixcelshuffer.
If you write it in code, it will look like this.
class ResidualBlock(nn.Module):
def __init__(self, nf=64):
super(ResidualBlock, self).__init__()
self.Block = nn.Sequential(
nn.Conv2d(nf, nf, kernel_size=3, padding=1),
nn.BatchNorm2d(nf),
nn.PReLU(),
nn.Conv2d(nf, nf, kernel_size=3, padding=1),
nn.BatchNorm2d(nf),
)
def forward(self, x):
out = self.Block(x)
return x + out
class Generator(nn.Module):
def __init__(self):
super(Generator, self).__init__()
self.conv1 = nn.Conv2d(3, 64, kernel_size=9, padding=4)
self.relu = nn.PReLU()
self.residualLayer = nn.Sequential(
ResidualBlock(),
ResidualBlock(),
ResidualBlock(),
ResidualBlock(),
ResidualBlock()
)
self.pixelShuffle = nn.Sequential(
nn.Conv2d(64, 64*4, kernel_size=3, padding=1),
nn.PReLU(),
nn.PixelShuffle(2),
nn.Conv2d(64, 3, kernel_size=9, padding=4),
nn.Tanh()
)
def forward(self, x):
x = self.conv1(x)
skip = self.relu(x)
x = self.residualLayer(skip)
x = self.pixelShuffle(x + skip)
return x
Discriminator uses an unconventional convolutional network. The size of the argument is the vertical and horizontal size of the image, this time the size of the input image is 64x64.
class Discriminator(nn.Module):
def __init__(self, size=64):
super(Discriminator, self).__init__()
size = int(size / 8) ** 2
self.net = nn.Sequential(
nn.Conv2d(3, 32, kernel_size=3, padding=1),
nn.LeakyReLU(0.2),
nn.Conv2d(32, 32, kernel_size=3, stride=2, padding=1),
nn.BatchNorm2d(32),
nn.LeakyReLU(0.2),
nn.Conv2d(32, 64, kernel_size=3, padding=1),
nn.BatchNorm2d(64),
nn.LeakyReLU(0.2),
nn.Conv2d(64, 64, kernel_size=3, stride=2, padding=1),
nn.BatchNorm2d(64),
nn.LeakyReLU(0.2),
nn.Conv2d(64, 128, kernel_size=3, padding=1),
nn.BatchNorm2d(128),
nn.LeakyReLU(0.2),
nn.Conv2d(128, 128, kernel_size=3, stride=2, padding=1),
nn.BatchNorm2d(128),
nn.LeakyReLU(0.2),
Flatten(),
nn.Linear(128 * size, 1024),
nn.LeakyReLU(0.2),
nn.Linear(1024, 1)
)
def forward(self, x):
return self.net(x)
class Flatten(nn.Module):
def forward(self, x):
return x.view(x.shape[0], -1)
Generator loss Vggloss is used for the loss of the Generator, vggloss is clearer by making the average of the features through the layers of the trained vgg model, whereas mseloss makes the average of the pixels of the image loss Generate an image.
class VGGLoss(nn.Module):
def __init__(self):
super(VGGLoss, self).__init__()
vgg = models.vgg16(pretrained=True)
self.contentLayers = nn.Sequential(*list(vgg.features)[:31]).cuda().eval()
for param in self.contentLayers.parameters():
param.requires_grad = False
def forward(self, fakeFrame, frameY):
MSELoss = nn.MSELoss()
content_loss = MSELoss(self.contentLayers(fakeFrame), self.contentLayers(frameY))
return content_loss
The loss of the Generator is the sum of this content_loss and the BCE Loss output from the Discriminator. Based on these, we will create a train function
def train(loader):
tensor_x, tensor_y = torch.tensor(x, dtype=torch.float), torch.tensor(y, dtype=torch.float)
DS = TensorDataset(tensor_x, tensor_y)
loader = DataLoader(DS, batch_size=BATCH_SIZE, shuffle=True)
D.train()
G.train()
D_optimizer = torch.optim.Adam(D.parameters(), lr=DiscriminatorLR, betas=(0.9, 0.999))
G_optimizer = torch.optim.Adam(G.parameters(), lr=GeneratorLR, betas=(0.9, 0.999))
realLabel = torch.ones(BATCH_SIZE, 1).cuda()
fakeLabel = torch.zeros(BATCH_SIZE, 1).cuda()
BCE = torch.nn.BCELoss()
VggLoss = VGGLoss()
for batch_idx, (X, Y) in enumerate(loader):
if X.shape[0] < BATCH_SIZE:
break
X = X.cuda()
Y = Y.cuda()
fakeFrame = G(X)
D.zero_grad()
DReal = D(Y)
DFake = D(fakeFrame)
D_loss = (BCE(DFake, fakeLabel) + BCE(DReal, realLabel)) / 2
D_loss.backward(retain_graph=True)
D_optimizer.step()
G.zero_grad()
G_label_loss= BCE(DFake, realLabel)
G_loss = VggLoss(fakeFrame, Y) + 1e-3 * G_label_loss
G_loss.backward()
G_optimizer.step()
print("G_loss :", G_loss, " D_loss :", D_loss)
The image below shows the result of 32epoch training. The top is a reduced image, the middle is the SRGAN output, and the bottom is the original image.
Feeling not bad for accuracy,
ESRGAN
Difference from SRGAN
RRDN(Residual in Residual Dense Network)
・ It seems that the generation capacity will increase by removing the batch normalization.
DenseBlock adds layer output to all layer inputs
・ Furthermore, connect three Dense Blocks in the same way as ResNet.
When implemented, it looks like this
class ResidualDenseBlock(nn.Module):
def __init__(self, nf=64, gc=32, bias=True):
super(ResidualDenseBlock, self).__init__()
self.conv1 = nn.Conv2d(nf, gc, 3, padding=1, bias=bias)
self.conv2 = nn.Conv2d(nf + gc, gc, 3, padding=1, bias=bias)
self.conv3 = nn.Conv2d(nf + 2 * gc, gc, 3, padding=1, bias=bias)
self.conv4 = nn.Conv2d(nf + 3 * gc, gc, 3, padding=1, bias=bias)
self.conv5 = nn.Conv2d(nf + 4 * gc, nf, 3, padding=1, bias=bias)
self.lrelu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
def forward(self, x):
x1 = self.lrelu(self.conv1(x))
x2 = self.lrelu(self.conv2(torch.cat((x, x1), dim=1)))
x3 = self.lrelu(self.conv3(torch.cat((x, x1, x2), dim=1)))
x4 = self.lrelu(self.conv4(torch.cat((x, x1, x2, x3), dim=1)))
x5 = self.conv5(torch.cat((x, x1, x2, x3, x4), dim=1))
return x5 * 0.2 + x
class Generator(nn.Module):
def __init__(self, nf=64):
super(Generator, self).__init__()
self.conv1 = nn.Conv2d(3, nf, kernel_size=3, padding=1)
self.relu = nn.LeakyReLU(negative_slope=0.2, inplace=True)
self.blockLayer = nn.Sequential(
ResidualDenseBlock(),
ResidualDenseBlock(),
ResidualDenseBlock(),
)
self.pixelShuffle = nn.Sequential(
nn.Conv2d(nf, nf * 4, kernel_size=3, padding=1),
nn.LeakyReLU(negative_slope=0.2, inplace=True),
nn.PixelShuffle(2),
nn.Conv2d(nf, nf, kernel_size=3, padding=1),
nn.Conv2d(nf, 3, kernel_size=3, padding=1),
nn.Tanh()
)
Relativistic GAN The SRGAN discriminator trains the real one to output 1 fake as 0, but the Relativistic GAN compares the real image with the fake image and sets the difference and label as BC Eloss. reference https://github.com/Yagami360/MachineLearning-Papers_Survey/issues/51 VGG Perceptual Loss In SRGAN, features were extracted using VGG16, but in Perceptual Loss, the structure is such that L1_loss for each pooling layer of VGG16 is added. It looks like this when I write it roughly
class VGGPerceptualLoss(torch.nn.Module):
def __init__(self):
super(VGGPerceptualLoss, self).__init__()
blocks = []
blocks.append(models.vgg16(pretrained=True).features[:4].eval())
blocks.append(models.vgg16(pretrained=True).features[4:9].eval())
blocks.append(models.vgg16(pretrained=True).features[9:16].eval())
blocks.append(models.vgg16(pretrained=True).features[16:23].eval())
blocks.append(models.vgg16(pretrained=True).features[23:30].eval())
for bl in blocks:
for p in bl:
p.requires_grad = False
self.blocks = torch.nn.ModuleList(blocks).cuda()
self.mean = torch.nn.Parameter(torch.tensor([0.485, 0.456, 0.406]).view(1, 3, 1, 1), requires_grad=False).cuda()
self.std = torch.nn.Parameter(torch.tensor([0.229, 0.224, 0.225]).view(1, 3, 1, 1), requires_grad=False).cuda()
def forward(self, fakeFrame, frameY):
fakeFrame = (fakeFrame - self.mean) / self.std
frameY = (frameY - self.mean) / self.std
loss = 0.0
x = fakeFrame
y = frameY
for block in self.blocks:
x = block(x)
y = block(y)
loss += torch.nn.functional.l1_loss(x, y)
return loss
Learning results
The top is a reduced image, the middle is the output with ESRGAN, and the bottom is the original image
Like SRGAN, 32epoch is upscaled from 32px to 64px.
Let's compare the generated images side by side, the top is SRGAN and the bottom is ESRGAN. Noise is noticeable in SRGAN, but there is less noise in ESRGAN, and the overall outline is clearer than in SRGAN.
Recommended Posts