\ub525\ub7ec\ub2dd\uc758 \uc138\uacc4\uc5d0\uc11c GAN(Generative Adversarial Networks)\uc740 \uac00\uc7a5 \ud601\uc2e0\uc801\uc774\uace0 \ub9e4\ub825\uc801\uc778 \uc5f0\uad6c \uc8fc\uc81c \uc911 \ud558\ub098\ub85c \uc790\ub9ac\uc7a1\uace0 \uc788\uc2b5\ub2c8\ub2e4. Ian Goodfellow\uac00 2014\ub144 \ucc98\uc74c \uc81c\uc2dc\ud55c GAN\uc740 \uc0dd\uc131\uc790\uc640 \uad6c\ubd84\uc790 \uac04\uc758 \uacbd\uc7c1\uc801 \uad00\uacc4\ub97c \ud1b5\ud574 \uac15\ub825\ud55c \uc774\ubbf8\uc9c0 \uc0dd\uc131 \ubaa8\ub378\uc744 \uac00\ub2a5\ud558\uac8c \ud588\uc2b5\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c\ub294 GAN\uc758 \uae30\ucd08 \uac1c\ub150\uc744 \uc124\uba85\ud558\uace0, \ucd5c\uadfc 5\ub144\uac04\uc758 \ubc1c\uc804\uc744 \uc0b4\ud3b4\ubcf4\uba70, \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c GAN \uad6c\ud604 \uc608\uc81c\ub97c \uc81c\uacf5\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
GAN\uc740 \uc0dd\uc131\uc790\uc640 \uad6c\ubd84\uc790\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4. \uc0dd\uc131\uc790\ub294 \uac00\uc9dc \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \uad6c\ubd84\uc790\ub294 \uc774 \ub370\uc774\ud130\uac00 \uc2e4\uc81c\uc778\uc9c0 \uac00\uc9dc\uc778\uc9c0 \ud310\ubcc4\ud569\ub2c8\ub2e4. \ub450 \ub124\ud2b8\uc6cc\ud06c\ub294 \uc11c\ub85c\ub97c \uacbd\uc7c1\uc801\uc73c\ub85c \ubc1c\uc804\uc2dc\ud0a4\uba70, \uc774\ub97c \ud1b5\ud574 \uc0dd\uc131\uc790\ub294 \uc810\uc810 \ub354 \ud604\uc2e4\uc801\uc778 \ub370\uc774\ud130\ub97c \ub9cc\ub4e4\uc5b4\ub0bc \uc218 \uc788\uc2b5\ub2c8\ub2e4. GAN\uc758 \ubaa9\ud45c\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
GAN\uc758 \ud559\uc2b5 \uacfc\uc815\uc740 \ub450 \uac1c\uc758 \ub124\ud2b8\uc6cc\ud06c\ub97c \ucd5c\uc801\ud654\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \ub2e4\uc74c\uc758 \uc190\uc2e4 \ud568\uc218\uac00 \uc0ac\uc6a9\ub429\ub2c8\ub2e4:<\/p>\n
\nL(D, G) = E[log D(x)] + E[log(1 – D(G(z)))]<\/p>\n<\/blockquote>\n
\uc5ec\uae30\uc11c D\ub294 \uad6c\ubd84\uc790, G\ub294 \uc0dd\uc131\uc790, x\ub294 \uc2e4\uc81c \ub370\uc774\ud130, z\ub294 \ubb34\uc791\uc704 \ub178\uc774\uc988 \ubca1\ud130\uc785\ub2c8\ub2e4. GAN\uc758 \ubaa9\ud45c\ub294 \ub450 \ub124\ud2b8\uc6cc\ud06c\uac00 \uc81c\ub85c\uc12c \uac8c\uc784\uc744 \ud1b5\ud574 \uc11c\ub85c\ub97c \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
2. GAN\uc758 \ucd5c\uadfc \ubc1c\uc804<\/h2>\n
\ucd5c\uadfc 5\ub144\uac04 GAN\uc740 \uc5ec\ub7ec \uac00\uc9c0 \ubcc0\ud615\uacfc \uac1c\uc120\uc744 \uac70\ucce4\uc2b5\ub2c8\ub2e4. \uc544\ub798\ub294 \uadf8 \uc911 \uc77c\ubd80\uc785\ub2c8\ub2e4:<\/p>\n
2.1 DCGAN (Deep Convolutional GAN)<\/h3>\n
DCGAN\uc740 CNN(Convolutional Neural Network)\uc744 \ud65c\uc6a9\ud558\uc5ec GAN\uc758 \uc131\ub2a5\uc744 \uac1c\uc120\ud588\uc2b5\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc778 GAN \uad6c\uc870\uc5d0 CNN\uc744 \ub3c4\uc785\ud568\uc73c\ub85c\uc368 \ub192\uc740 \ud488\uc9c8\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \ub370 \uc131\uacf5\ud588\uc2b5\ub2c8\ub2e4.<\/p>\n
2.2 WGAN (Wasserstein GAN)<\/h3>\n
WGAN\uc740 GAN\uc758 \ud6c8\ub828 \uc548\uc815\uc131\uc744 \uac1c\uc120\ud558\uae30 \uc704\ud574 Wasserstein \uac70\ub9ac \uac1c\ub150\uc744 \ub3c4\uc785\ud588\uc2b5\ub2c8\ub2e4. WGAN\uc740 \uae30\uc874 GAN\ubcf4\ub2e4 \ub354 \ube60\ub974\uace0 \uc548\uc815\uc801\uc73c\ub85c \uc218\ub834\ud558\uba70, \ub354 \ub098\uc740 \ud488\uc9c8\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
2.3 CycleGAN<\/h3>\n
CycleGAN\uc740 \uc774\ubbf8\uc9c0 \ubcc0\ud658 \ubb38\uc81c\ub97c \ud574\uacb0\ud558\ub294 \ub370 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uc0ac\uc9c4 \uc774\ubbf8\uc9c0\ub97c \ud654\ud48d\uc73c\ub85c \ubcc0\ud658\ud558\ub294 \ub4f1\uc758 \uc791\uc5c5\uc5d0 \ud65c\uc6a9\ub429\ub2c8\ub2e4. CycleGAN\uc740 \uc8fc\uc5b4\uc9c4 \uc774\ubbf8\uc9c0 \uc30d \uc5c6\uc774\ub3c4 \ud559\uc2b5\ud560 \uc218 \uc788\ub294 \ub2a5\ub825\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
2.4 StyleGAN<\/h3>\n
StyleGAN\uc740 \uace0\ud488\uc9c8 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \ucd5c\uc2e0 GAN \uc544\ud0a4\ud14d\ucc98\uc785\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \uc0dd\uc131 \uacfc\uc815\uc5d0\uc11c \uc2a4\ud0c0\uc77c\uc744 \uc870\uc815\ud560 \uc218 \uc788\uc5b4 \ub2e4\uc591\ud55c \uc2a4\ud0c0\uc77c\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
3. \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c GAN \uad6c\ud604<\/h2>\n
\uc774\uc81c \ud30c\uc774\ud1a0\uce58(PyTorch)\ub97c \uc0ac\uc6a9\ud558\uc5ec \uae30\ubcf8 GAN\uc744 \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc544\ub798 \ucf54\ub4dc\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc22b\uc790 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \uac04\ub2e8\ud55c GAN\uc758 \uad6c\ud604 \uc608\uc81c\uc785\ub2c8\ub2e4.<\/p>\n
3.1 \ub77c\uc774\ube0c\ub7ec\ub9ac \uc784\ud3ec\ud2b8<\/h3>\n
import torch\nimport torch.nn as nn\nimport torch.optim as optim\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets\nimport matplotlib.pyplot as plt\nimport numpy as np\n <\/code><\/pre>\n3.2 \ub370\uc774\ud130\uc14b \ub85c\ub4dc<\/h3>\n
MNIST \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud558\uace0 \ubcc0\ud658\uc744 \uc218\ud589\ud569\ub2c8\ub2e4.<\/p>\n
# MNIST \ub370\uc774\ud130\uc14b \ub85c\ub4dc \ubc0f \ubcc0\ud658\ntransform = transforms.Compose([\n transforms.ToTensor(),\n transforms.Normalize((0.5,), (0.5,))\n])\n\nmnist = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ndataloader = torch.utils.data.DataLoader(mnist, batch_size=64, shuffle=True)\n <\/code><\/pre>\n3.3 \uc0dd\uc131\uc790 \ubc0f \uad6c\ubd84\uc790 \ubaa8\ub378 \uc815\uc758<\/h3>\n
\uc0dd\uc131\uc790\uc640 \uad6c\ubd84\uc790 \ubaa8\ub378\uc744 \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\n
# \uc0dd\uc131\uc790 \ubaa8\ub378 \uc815\uc758\nclass Generator(nn.Module):\n def __init__(self):\n super(Generator, self).__init__()\n self.model = nn.Sequential(\n nn.Linear(100, 256),\n nn.ReLU(True),\n nn.Linear(256, 512),\n nn.ReLU(True),\n nn.Linear(512, 1024),\n nn.ReLU(True),\n nn.Linear(1024, 784),\n nn.Tanh(),\n )\n\n def forward(self, z):\n return self.model(z)\n\n# \uad6c\ubd84\uc790 \ubaa8\ub378 \uc815\uc758\nclass Discriminator(nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.model = nn.Sequential(\n nn.Linear(784, 1024),\n nn.LeakyReLU(0.2, inplace=True),\n nn.Linear(1024, 512),\n nn.LeakyReLU(0.2, inplace=True),\n nn.Linear(512, 256),\n nn.LeakyReLU(0.2, inplace=True),\n nn.Linear(256, 1),\n nn.Sigmoid(),\n )\n\n def forward(self, img):\n return self.model(img.view(img.size(0), -1))\n <\/code><\/pre>\n3.4 \uc190\uc2e4 \ud568\uc218 \ubc0f \uc635\ud2f0\ub9c8\uc774\uc800 \uc815\uc758<\/h3>\n
GAN\uc758 \ud559\uc2b5\uc744 \uc704\ud55c \uc190\uc2e4 \ud568\uc218\uc640 \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\n
criterion = nn.BCELoss()\ngenerator = Generator()\ndiscriminator = Discriminator()\n\noptimizer_G = optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))\noptimizer_D = optim.Adam(discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999))\n <\/code><\/pre>\n3.5 GAN \ud559\uc2b5 \ud504\ub85c\uc138\uc2a4<\/h3>\n
\uc774\uc81c GAN\uc744 \ud559\uc2b5\uc2dc\ud0a4\ub294 \ud568\uc218\ub97c \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\n
def train_gan(num_epochs=50):\n G_losses = []\n D_losses = []\n for epoch in range(num_epochs):\n for i, (imgs, _) in enumerate(dataloader):\n # \uc9c4\uc9dc \uc774\ubbf8\uc9c0 \ub808\uc774\ube14\uc740 1, \uac00\uc9dc \uc774\ubbf8\uc9c0 \ub808\uc774\ube14\uc740 0\n real_labels = torch.ones(imgs.size(0), 1)\n fake_labels = torch.zeros(imgs.size(0), 1)\n\n # \uad6c\ubd84\uc790 \ud559\uc2b5\n optimizer_D.zero_grad()\n outputs = discriminator(imgs)\n D_loss_real = criterion(outputs, real_labels)\n D_loss_real.backward()\n\n z = torch.randn(imgs.size(0), 100)\n fake_imgs = generator(z)\n outputs = discriminator(fake_imgs.detach())\n D_loss_fake = criterion(outputs, fake_labels)\n D_loss_fake.backward()\n optimizer_D.step()\n\n # \uc0dd\uc131\uc790 \ud559\uc2b5\n optimizer_G.zero_grad()\n outputs = discriminator(fake_imgs)\n G_loss = criterion(outputs, real_labels)\n G_loss.backward()\n optimizer_G.step()\n\n G_losses.append(G_loss.item())\n D_losses.append(D_loss_real.item() + D_loss_fake.item())\n\n print(f'Epoch [{epoch}\/{num_epochs}], D_loss: {D_loss_fake.item() + D_loss_real.item()}, G_loss: {G_loss.item()}')\n \n return G_losses, D_losses\n <\/code><\/pre>\n3.6 \ud559\uc2b5 \uc2e4\ud589 \ubc0f \uacb0\uacfc \uc2dc\uac01\ud654<\/h3>\n
\ud559\uc2b5\uc744 \uc2e4\ud589\ud558\uace0 \uc190\uc2e4 \uac12\uc744 \uc2dc\uac01\ud654\ud569\ub2c8\ub2e4.<\/p>\n
G_losses, D_losses = train_gan(num_epochs=50)\n\nplt.plot(G_losses, label='Generator Loss')\nplt.plot(D_losses, label='Discriminator Loss')\nplt.title('Losses during Training')\nplt.xlabel('Iterations')\nplt.ylabel('Loss')\nplt.legend()\nplt.show()\n <\/code><\/pre>\n4. \uacb0\ub860<\/h2>\n
\uc774 \uae00\uc5d0\uc11c\ub294 GAN\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \ucd5c\uadfc 5\ub144\uac04\uc758 \ubc1c\uc804\uc744 \uc0b4\ud3b4\ubcf4\uace0, \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud558\ub294 \uc608\uc81c\ub97c \ubcf4\uc5ec\uc8fc\uc5c8\uc2b5\ub2c8\ub2e4. GAN\uc740 \uacc4\uc18d\ud574\uc11c \ubc1c\uc804\ud558\uace0 \uc788\uc73c\uba70, \ub525\ub7ec\ub2dd \ubd84\uc57c\uc5d0\uc11c \uc911\uc694\ud55c \uc704\uce58\ub97c \ucc28\uc9c0\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c\uc758 \uc5f0\uad6c \ubc29\ud5a5\uc73c\ub85c\ub294 \ub354\uc6b1 \uc548\uc815\uc801\uc778 \ud559\uc2b5 \ubc29\ubc95\uacfc \uace0\ud574\uc0c1\ub3c4 \uc774\ubbf8\uc9c0 \uc0dd\uc131\uc744 \uc704\ud55c \uc5f0\uad6c\uac00 \uae30\ub300\ub429\ub2c8\ub2e4.<\/p>\n
5. \ucc38\uace0 \ubb38\ud5cc<\/h2>\n
\n
- Goodfellow, I., et al. (2014). Generative Adversarial Nets. Advances in Neural Information Processing Systems<\/em>.<\/li>\n
- Brock, A., Donahue, J., & Simonyan, K. (2019). Large Scale GAN Training for High Fidelity Natural Image Synthesis. International Conference on Learning Representations<\/em>.<\/li>\n
- Karras, T., Laine, S., & Aila, T. (2019). A Style-Based Generator Architecture for Generative Adversarial Networks. IEEE\/CVF Conference on Computer Vision and Pattern Recognition<\/em>.<\/li>\n
- CycleGAN: Unpaired Image-to-Image Translation using Cycle Consistent Adversarial Networks. IEEE International Conference on Computer Vision (ICCV)<\/em>.<\/li>\n<\/ul>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"