\ub525\ub7ec\ub2dd\uc758 \ubc1c\uc804\uacfc \ud568\uaed8 \uc774\ubbf8\uc9c0 \uc0dd\uc131, \uac15\ud654\ud559\uc2b5, \uc774\ubbf8\uc9c0 \ubcc0\ud658, \uc774\ubbf8\uc9c0 \uacb0\ud569 \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c Generative Adversarial Networks(GANs)\uc758 \uc0ac\uc6a9\uc774 \uc99d\uac00\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4. GAN\uc740 \ub450 \ub124\ud2b8\uc6cc\ud06c, \uc989 \uc0dd\uc131\uc790(Generator)\uc640 \ud310\ubcc4\uc790(Discriminator) \uac04\uc758 \uacbd\uc7c1\uc744 \ud1b5\ud574 \uace0\ud574\uc0c1\ub3c4\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294\ub370 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c\ub294 GAN\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \ud568\uaed8 WGAN(Wasserstein GAN)\uc758 \uad6c\uc870\uc640 \uc791\ub3d9 \ubc29\uc2dd, \uc774\ub97c \uad6c\ud604\ud558\uae30 \uc704\ud55c \ud30c\uc774\ud1a0\uce58 \uc608\uc81c \ucf54\ub4dc\ub97c \ub2e4\ub8e8\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
GAN\uc740 Ian Goodfellow\uac00 2014\ub144\uc5d0 \uc81c\uc548\ud55c \ubaa8\ub378\ub85c, \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub77c\ub294 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc73c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4. \uc0dd\uc131\uc790\ub294 \ub09c\uc218 \ubca1\ud130\ub97c \uc785\ub825\ubc1b\uc544 \uc2e4\uc81c\uc640 \uc720\uc0ac\ud55c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \ud310\ubcc4\uc790\ub294 \uc785\ub825 \ub370\uc774\ud130\uac00 \uc2e4\uc81c \ub370\uc774\ud130\uc778\uc9c0 \uc0dd\uc131\ub41c \ub370\uc774\ud130\uc778\uc9c0\ub97c \ud310\ub2e8\ud569\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c \ub450 \uc2e0\uacbd\ub9dd\uc740 \uc11c\ub85c \uacbd\uc7c1\ud558\uba74\uc11c \uc810\uc810 \uc644\ubcbd\ud55c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \ubc29\ud5a5\uc73c\ub85c \ud559\uc2b5\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
GAN\uc758 \uc190\uc2e4 \ud568\uc218\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4.<\/p>\n
\n L(D) = -E[log(D(x))] - E[log(1 - D(G(z)))],\n L(G) = -E[log(D(G(z)))]\n <\/pre>\n\uc5ec\uae30\uc11c
D(x)<\/code>\ub294 \uc2e4\uc81c \ub370\uc774\ud130\uc758 \uc9c4\uc9dc\ub85c \ud310\ub2e8\ud560 \ud655\ub960\uc774\uba70,G(z)<\/code>\ub294 \uc0dd\uc131\uc790\uac00 \uc0dd\uc131\ud55c \ub370\uc774\ud130\uc785\ub2c8\ub2e4.<\/p>\n2. WGAN – \uc640\uc11c\uc2a4\ud14c\uc778 GAN<\/h2>\n
\uae30\uc874\uc758 GAN\uc740 \ud310\ubcc4\uc790\uc758 \uc190\uc2e4 \ud568\uc218\uac00 \uc548\uc815\uc801\uc774\uc9c0 \uc54a\uace0 \ud559\uc2b5\uc774 \ubd88\uc548\uc815\ud558\ub2e4\ub294 \ubb38\uc81c\uc810\uc774 \uc788\uc5c8\uc2b5\ub2c8\ub2e4. WGAN\uc740 Wasserstein Distance\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc774\ub7ec\ud55c \ubb38\uc81c\uc810\uc744 \ud574\uacb0\ud569\ub2c8\ub2e4. Wasserstein \uac70\ub9ac(\ud639\uc740 Earth Mover’s Distance)\ub294 \ub450 \ud655\ub960 \ubd84\ud3ec \uac04\uc758 \ucd5c\uc801 \uc6b4\ubc18 \ube44\uc6a9\uc744 \uce21\uc815\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/p>\n
2.1 WGAN\uc758 \uac1c\uc120\uc810<\/h3>\n
\n
- WGAN\uc740 \ud310\ubcc4\uc790 \ub300\uc2e0 \ube44\uc120\ud615 \ud68c\uadc0\ubaa8\ub378\uc778 ‘Critic’\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/li>\n
- WGAN\uc758 \uc190\uc2e4 \ud568\uc218\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:\n
\n L(D) = E[D(x)] - E[D(G(z))],\n L(G) = -E[D(G(z))]\n <\/pre>\n<\/li>\n- WGAN\uc740 Weight Clipping\uc744 \ud1b5\ud574 Critic\uc758 Lipschitz \uc5f0\uc18d\uc131\uc744 \ubcf4\uc7a5\ud569\ub2c8\ub2e4.<\/li>\n
- Gradient Penalty \uae30\ubc95\uc744 \uc0ac\uc6a9\ud558\uc5ec Lipschitz \uc81c\uc57d \uc870\uac74\uc744 \uc644\ud654\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
2.2 WGAN\uc758 \uad6c\uc870<\/h3>\n
WGAN\uc740 \uae30\ubcf8 GAN\uc758 \uad6c\uc870\uc5d0\uc11c Critic\uc744 \ub3c4\uc785\ud574 \uc218\uc815\ub41c \ud615\ud0dc\uc785\ub2c8\ub2e4. \ub2e4\uc74c\uc740 WGAN\uc758 \ub124\ud2b8\uc6cc\ud06c \uad6c\uc870\uc785\ub2c8\ub2e4:<\/p>\n
\n
- \uc774\uc804\uc758 \ud310\ubcc4\uc790\ub294 \ud604\uc7ac\uc758 Critic\uc73c\ub85c \ub300\uccb4\ub429\ub2c8\ub2e4.<\/li>\n<\/ul>\n
3. \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c WGAN \uad6c\ud604<\/h2>\n
\uc774\uc81c \ud30c\uc774\ud1a0\uce58\ub85c WGAN\uc744 \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774\ubc88 \uc608\uc81c\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc190\uae00\uc528 \uc22b\uc790\ub97c \uc0dd\uc131\ud558\ub294\ubaa8\ub378\uc744 \uad6c\ucd95\ud560 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
3.1 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\uba3c\uc800 \ub370\uc774\ud130\uc14b\uc744 \ubd88\ub7ec\uc624\uace0 \uc804\ucc98\ub9ac\ud569\ub2c8\ub2e4.<\/p>\n
\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torchvision import datasets, transforms\nfrom torch.utils.data import DataLoader\n\n# \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud558\uace0 \uc804\ucc98\ub9ac\ud569\ub2c8\ub2e4.\ntransform = transforms.Compose([\n transforms.Resize(28),\n transforms.ToTensor(),\n transforms.Normalize((0.5,), (0.5,))\n])\n\ntrain_data = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = DataLoader(train_data, batch_size=64, shuffle=True)\n <\/code><\/pre>\n3.2 WGAN \ubaa8\ub378 \uc815\uc758<\/h3>\n
\uc774\uc81c Generator\uc640 Critic \ubaa8\ub378\uc744 \uc815\uc758\ud560 \ucc28\ub840\uc785\ub2c8\ub2e4.<\/p>\n
\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(),\n nn.Linear(256, 512),\n nn.ReLU(),\n nn.Linear(512, 1024),\n nn.ReLU(),\n nn.Linear(1024, 784),\n nn.Tanh()\n )\n \n def forward(self, z):\n return self.model(z).view(-1, 1, 28, 28) # 28x28 \uc774\ubbf8\uc9c0\ub85c \ubcc0\ud615\n\n# \ube44\ud3c9\uac00(critic) \ubaa8\ub378 \uc815\uc758\nclass Critic(nn.Module):\n def __init__(self):\n super(Critic, self).__init__()\n self.model = nn.Sequential(\n nn.Linear(784, 512),\n nn.LeakyReLU(0.2),\n nn.Linear(512, 256),\n nn.LeakyReLU(0.2),\n nn.Linear(256, 1)\n )\n \n def forward(self, img):\n return self.model(img.view(-1, 784)) # 784 \ucc28\uc6d0\uc73c\ub85c \ubcc0\ud615\n <\/code><\/pre>\n3.3 WGAN \ud559\uc2b5 \uacfc\uc815<\/h3>\n
\uc774\uc81c WGAN\uc758 \ud559\uc2b5 \uacfc\uc815\uc744 \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\n
\ndef train_wgan(num_epochs):\n generator = Generator()\n critic = Critic()\n \n # \uc635\ud2f0\ub9c8\uc774\uc800 \uc124\uc815\n optimizer_G = optim.RMSprop(generator.parameters(), lr=0.00005)\n optimizer_C = optim.RMSprop(critic.parameters(), lr=0.00005)\n\n for epoch in range(num_epochs):\n for i, (imgs, _) in enumerate(train_loader):\n imgs = imgs.to(device)\n\n # Critic \uc794\uc5ec \ubc29\uc815\uc2dd\n optimizer_C.zero_grad()\n z = torch.randn(imgs.size(0), 100).to(device)\n fake_imgs = generator(z)\n c_real = critic(imgs)\n c_fake = critic(fake_imgs.detach())\n c_loss = c_fake.mean() - c_real.mean()\n c_loss.backward()\n optimizer_C.step()\n\n # Weight Clipping\n for p in critic.parameters():\n p.data.clamp_(-0.01, 0.01)\n\n # Generator \uc5c5\ub370\uc774\ud2b8\n if i % 5 == 0:\n optimizer_G.zero_grad()\n g_loss = -critic(fake_imgs).mean()\n g_loss.backward()\n optimizer_G.step()\n \n print(f'Epoch [{epoch}\/{num_epochs}], Loss C: {c_loss.item()}, Loss G: {g_loss.item()}')\n\n# GPU \uc0ac\uc6a9 \uc124\uc815\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\ntrain_wgan(num_epochs=50)\n <\/code><\/pre>\n3.4 \uacb0\uacfc \uc2dc\uac01\ud654<\/h3>\n
\ud559\uc2b5\uc774 \ub05d\ub098\uba74 \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \uc2dc\uac01\ud654\ud558\uc5ec \uacb0\uacfc\ub97c \ud655\uc778\ud569\ub2c8\ub2e4.<\/p>\n
\nimport matplotlib.pyplot as plt\n\ndef show_generated_images(num_images):\n z = torch.randn(num_images, 100).to(device)\n generated_imgs = generator(z).cpu().detach()\n \n fig, axes = plt.subplots(1, num_images, figsize=(15, 15))\n for i in range(num_images):\n axes[i].imshow(generated_imgs[i][0], cmap='gray')\n axes[i].axis('off')\n plt.show()\n\n# \uacb0\uacfc \uc2dc\uac01\ud654\nshow_generated_images(5)\n <\/code><\/pre>\n4. \uacb0\ub860<\/h2>\n
WGAN\uc740 \uae30\uc874 GAN\uc758 \ubb38\uc81c\uc810\uc744 \uadf9\ubcf5\ud558\uae30 \uc704\ud574 Wasserstein Distance\ub97c \uc774\uc6a9\ud558\uc5ec \ub354 \uc548\uc815\uc801\uc778 \ud559\uc2b5 \uacfc\uc815\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. \ubcf8 \uae00\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud558\uc5ec WGAN\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \uc18c\uac1c\ud558\uc600\uc73c\uba70, \uc774\ub97c \ud1b5\ud574 \uc0dd\uc131\uc801 \uc801\ub300 \uc2e0\uacbd\ub9dd\uc758 \uc774\ud574\ub97c \ub192\uc77c \uc218 \uc788\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4. GAN\uacfc \uadf8 \ubcc0\ud615 \ubaa8\ub378\ub4e4\uc740 \uc774\ubbf8\uc9c0 \uc0dd\uc131\ubfd0\ub9cc \uc544\ub2c8\ub77c \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud601\uc2e0\uc801\uc778 \uacb0\uacfc\ub97c \uac00\uc838\uc62c \uc218 \uc788\ub294 \uac15\ub825\ud55c \ub3c4\uad6c\uc785\ub2c8\ub2e4.<\/p>\n
5. \ucc38\uace0\ubb38\ud5cc<\/h2>\n
\n
- Ian J. Goodfellow et al., “Generative Adversarial Nets”, 2014.<\/li>\n
- Martin Arjovsky et al., “Wasserstein Generative Adversarial Networks”, 2017.<\/li>\n
- PyTorch Documentation: https:\/\/pytorch.org\/docs\/stable\/index.html<\/a><\/li>\n<\/ul>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"