\n Generative Adversarial Networks (GANs)\uc740 Ian Goodfellow\uc640 \uadf8\uc758 \ub3d9\ub8cc\ub4e4\uc5d0 \uc758\ud574 2014\ub144 \uc81c\uc548\ub41c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc785\ub2c8\ub2e4. GAN\uc740 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc778 \uc0dd\uc131\uc790(Generator)\uc640 \ud310\ubcc4\uc790(Discriminator)\uac00 \uc11c\ub85c \uacbd\uc7c1\ud558\uba74\uc11c \ud559\uc2b5\ud558\ub294 \uad6c\uc870\ub85c \ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uc0dd\uc131\uc790\ub294 \ub354\uc6b1 \ub354 \uc9c4\uc9dc \uac19\uc740 \ub370\uc774\ud130\ub97c \ub9cc\ub4e4\uace0, \ud310\ubcc4\uc790\ub294 \uc9c4\uc9dc \ub370\uc774\ud130\uc640 \uac00\uc9dc \ub370\uc774\ud130\ub97c \uad6c\ubcc4\ud558\ub294 \ub2a5\ub825\uc744 \ud0a4\uc6b0\uac8c \ub429\ub2c8\ub2e4.\n<\/p>\n
\n GAN\uc758 \uae30\ubcf8 \uc544\uc774\ub514\uc5b4\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4. \uc0dd\uc131\uc790\ub294 \ub79c\ub364 \ub178\uc774\uc988\ub97c \uc785\ub825\uc73c\ub85c \ubc1b\uc544 \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \ud310\ubcc4\uc790\ub294 \uc774 \ub370\uc774\ud130\uac00 \uc2e4\uc81c \ub370\uc774\ud130\uc778\uc9c0 \uc0dd\uc131\ub41c \ub370\uc774\ud130\uc778\uc9c0 \ud310\ubcc4\ud569\ub2c8\ub2e4. \uc774 \ub450 \ubaa8\ub378\uc740 \ubc18\ubcf5\uc801\uc73c\ub85c \ub300\uacb0\ud558\uba74\uc11c \uc11c\ub85c\uc758 \uc131\ub2a5\uc744 \uac1c\uc120\ud574 \ub098\uac11\ub2c8\ub2e4. \uc774\ub807\uac8c\ud574\uc11c \uc0dd\uc131\uc790\ub294 \uc810\uc810 \ub354 \uc9c4\uc9dc \uac19\uc740 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uac8c \ub418\uace0, \ud310\ubcc4\uc790\ub294 \ub354\uc6b1 \uc815\uad50\ud558\uac8c \uc9c4\uc9dc\uc640 \uac00\uc9dc\ub97c \uad6c\ubd84\ud558\uac8c \ub429\ub2c8\ub2e4.\n<\/p>\n
\n CycleGAN\uc740 GAN\uc758 \ubcc0\ud615\uc73c\ub85c, \uc11c\ub85c \ub2e4\ub978 \ub3c4\uba54\uc778 \uac04\uc758 \uc774\ubbf8\uc9c0 \ubcc0\ud658\uc744 \ud559\uc2b5\ud558\ub294 \ub370 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4 \ub9d0\uc758 \uc774\ubbf8\uc9c0\ub97c \uc5bc\ub8e9\ub9d0\uc758 \uc774\ubbf8\uc9c0\ub85c \ubcc0\ud658\ud558\uac70\ub098, \uc5ec\ub984 \ud48d\uacbd \uc0ac\uc9c4\uc744 \uaca8\uc6b8 \ud48d\uacbd \uc0ac\uc9c4\uc73c\ub85c \ubcc0\ud658\ud558\ub294 \uc791\uc5c5\uc774 \uac00\ub2a5\ud574\uc9d1\ub2c8\ub2e4. CycleGAN\uc740 \ub450 \uac1c\uc758 \uc0dd\uc131\uc790\uc640 \ub450 \uac1c\uc758 \ud310\ubcc4\uc790\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub450 \ub3c4\uba54\uc778 \uc0ac\uc774\uc758 \ubcc0\ud658\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.\n<\/p>\n
\n CycleGAN\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \ub2e8\uacc4\ub85c \uc791\ub3d9\ud569\ub2c8\ub2e4:\n<\/p>\n
\n \uc774\uc81c CycleGAN\uc744 \ud30c\uc774\ud1a0\uce58\ub85c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \ud30c\uc774\ud1a0\uce58\ub294 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \uc791\uc131\ud558\uae30\uc5d0 \ud6a8\uc728\uc801\uc778 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub85c, \uc0ac\uc6a9\uc790 \uce5c\ud654\uc801\uc778 API\uc640 \ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4. CycleGAN\uc744 \uad6c\ud604\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud569\ub2c8\ub2e4.\n<\/p>\n
pip install torch torchvision<\/code><\/pre>\n3.1 \ub77c\uc774\ube0c\ub7ec\ub9ac \uc784\ud3ec\ud2b8<\/h3>\n\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torchvision import datasets, transforms\n<\/code><\/pre>\n3.2 \ubaa8\ub378 \uc815\uc758<\/h3>\n
\n CycleGAN\uc758 \uc0dd\uc131\uc790\ub294 \uc77c\ubc18\uc801\uc73c\ub85c U-Net \uad6c\uc870\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\uc758 \uad6c\uc870\ub97c \uc544\ub798\uc640 \uac19\uc774 \uc815\uc758\ud558\uaca0\uc2b5\ub2c8\ub2e4.\n<\/p>\n
\nclass Generator(nn.Module):\n def __init__(self):\n super(Generator, self).__init__()\n self.model = nn.Sequential(\n nn.Conv2d(3, 64, kernel_size=7, stride=1, padding=3),\n nn.ReLU(inplace=True),\n # \ucd94\uac00\uc801\uc778 \ub808\uc774\uc5b4\ub97c \uc5ec\uae30\uc5d0 \ucd94\uac00\n nn.ConvTranspose2d(64, 3, kernel_size=7, stride=1, padding=3)\n )\n\n def forward(self, x):\n return self.model(x)\n\nclass Discriminator(nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.model = nn.Sequential(\n nn.Conv2d(3, 64, kernel_size=4, stride=2, padding=1),\n nn.LeakyReLU(0.2, inplace=True),\n # \ucd94\uac00\uc801\uc778 \ub808\uc774\uc5b4\ub97c \uc5ec\uae30\uc5d0 \ucd94\uac00\n nn.Conv2d(64, 1, kernel_size=4, stride=1, padding=1)\n )\n\n def forward(self, x):\n return self.model(x)\n<\/code><\/pre>\n3.3 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\n CycleGAN\uc744 \ud559\uc2b5\ud558\uae30 \uc704\ud574 \uc774\ubbf8\uc9c0 \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc5d0\uc11c\ub294 \u2018horse2zebra\u2019 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \ub370\uc774\ud130\uc14b\uc744 \ub2e4\uc6b4\ub85c\ub4dc\ud558\uace0 \ub370\uc774\ud130 \ub85c\ub354\ub97c \uc815\uc758\ud558\ub294 \ucf54\ub4dc\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4.\n<\/p>\n
\ntransform = transforms.Compose([\n transforms.Resize((256, 256)),\n transforms.ToTensor(),\n])\n\ntrain_dataset_x = datasets.ImageFolder('path_to_horse_dataset', transform=transform)\ntrain_loader_x = torch.utils.data.DataLoader(train_dataset_x, batch_size=1, shuffle=True)\n\ntrain_dataset_y = datasets.ImageFolder('path_to_zebra_dataset', transform=transform)\ntrain_loader_y = torch.utils.data.DataLoader(train_dataset_y, batch_size=1, shuffle=True)\n<\/code><\/pre>\n3.4 \uc190\uc2e4 \ud568\uc218 \ubc0f \ucd5c\uc801\ud654\uae30 \uc124\uc815<\/h3>\n
\n CycleGAN\uc5d0\uc11c\ub294 \ub450 \uac00\uc9c0 \uc190\uc2e4 \ud568\uc218, \uc989 \uc801\ub300\uc801 \uc190\uc2e4(Discriminator Loss)\uacfc \uc0ac\uc774\ud074 \uc77c\uad00\uc131 \uc190\uc2e4(Cycle Consistency Loss)\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc544\ub798\uc5d0 \uc774\ub4e4\uc744 \uc815\uc758\ud55c \uc608\uc81c\uac00 \uc788\uc2b5\ub2c8\ub2e4.\n<\/p>\n
\ndef discriminator_loss(real, fake):\n real_loss = criterion(real, torch.ones_like(real))\n fake_loss = criterion(fake, torch.zeros_like(fake))\n return (real_loss + fake_loss) \/ 2\n\ndef cycle_loss(real_image, cycled_image, lambda_cycle):\n return lambda_cycle * nn.L1Loss()(real_image, cycled_image)\n<\/code><\/pre>\n3.5 \ubaa8\ub378 \ud559\uc2b5<\/h3>\n
\n CycleGAN\uc758 \ud559\uc2b5 \uacfc\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4. \uac01 \uc5d0\ud3ed\ub9c8\ub2e4 \ub450 \ub3c4\uba54\uc778\uc5d0\uc11c \ubaa8\ub378\uc744 \uc5c5\ub370\uc774\ud2b8\ud558\uace0 \uc190\uc2e4\uc744 \uacc4\uc0b0\ud569\ub2c8\ub2e4.\n<\/p>\n
\ndef train(cycle_gan, dataloader_x, dataloader_y, num_epochs):\n for epoch in range(num_epochs):\n for real_x, real_y in zip(dataloader_x, dataloader_y):\n #\uce74\uc6b4\ud130 \ucf54\ub4dc \uc0dd\uc131 \ubc0f \uc190\uc2e4 \uacc4\uc0b0 \uacfc\uc815\n #\ubaa8\ub378 \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8\n #\uc190\uc2e4 \ucd9c\ub825\n<\/code><\/pre>\n3.6 \uacb0\uacfc \uc2dc\uac01\ud654<\/h3>\n
\n \ubaa8\ub378 \ud559\uc2b5\uc774 \uc644\ub8cc\ub418\uba74 \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \uc2dc\uac01\ud654\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc740 \ud559\uc2b5 \uacfc\uc815\uc5d0\uc11c \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \ud655\uc778\ud558\uace0, \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\ub294 \ub370 \uc720\uc6a9\ud569\ub2c8\ub2e4.\n<\/p>\n
\nimport matplotlib.pyplot as plt\n\ndef visualize_results(real_x, fake_y, cycled_x):\n plt.figure(figsize=(12, 12))\n plt.subplot(1, 3, 1)\n plt.title(\"Real X\")\n plt.imshow(real_x.permute(1, 2, 0).detach().numpy())\n \n plt.subplot(1, 3, 2)\n plt.title(\"Fake Y\")\n plt.imshow(fake_y.permute(1, 2, 0).detach().numpy())\n\n plt.subplot(1, 3, 3)\n plt.title(\"Cycled X\")\n plt.imshow(cycled_x.permute(1, 2, 0).detach().numpy())\n plt.show()\n<\/code><\/pre>\n4. CycleGAN\uc758 \uc751\uc6a9 \uc0ac\ub840<\/h2>\n
\n CycleGAN\uc740 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \uc751\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uba87 \uac00\uc9c0 \uc608\uc2dc\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:\n<\/p>\n
\n- \uc2a4\ud0c0\uc77c \uc804\uc774:<\/strong> \uc0ac\uc9c4\uc758 \uc2a4\ud0c0\uc77c\uc744 \ubcc0\uacbd\ud558\uc5ec \uc608\uc220 \uc791\ud488\uc73c\ub85c \ubcc0\ud658\ud558\ub294 \uc791\uc5c5\uc5d0 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/li>\n
- \uc774\ubbf8\uc9c0 \ubcf5\uc6d0:<\/strong> \uc800\ud574\uc0c1\ub3c4 \uc774\ubbf8\uc9c0\ub97c \uace0\ud574\uc0c1\ub3c4\ub85c \ubcc0\ud658\ud558\ub294 \uc791\uc5c5\uc774 \uac00\ub2a5\ud569\ub2c8\ub2e4.<\/li>\n
- \ube44\uac00\uc5ed\uc801 \ubcc0\ud658:<\/strong> \uc608\ub97c \ub4e4\uc5b4, \uc5ec\ub984 \uc774\ubbf8\uc9c0\ub97c \uaca8\uc6b8 \uc774\ubbf8\uc9c0\ub85c \ubcc0\ud658\ud558\ub294 \uac83\uacfc \uac19\uc740 \uc791\uc5c5\uc744 \uc9c0\uc6d0\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
5. \uacb0\ub860<\/h2>\n
\n CycleGAN\uc740 \uc774\ubbf8\uc9c0 \ubcc0\ud658 \ubd84\uc57c\uc5d0\uc11c \ub9e4\uc6b0 \uc720\uc6a9\ud55c \ub3c4\uad6c\ub85c, \ub450 \ub3c4\uba54\uc778 \uac04\uc758 \ube44\uc9c0\ub3c4 \ud559\uc2b5\uc744 \ud1b5\ud574 \ub6f0\uc5b4\ub09c \uc131\ub2a5\uc744 \ubcf4\uc785\ub2c8\ub2e4. \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud558\uba74 CycleGAN\uc744 \uac04\ud3b8\ud558\uac8c \uad6c\ud604\ud560 \uc218 \uc788\uc73c\uba70, \ub2e4\uc591\ud55c \uc774\ubbf8\uc9c0 \ubcc0\ud658 \uc791\uc5c5\uc5d0 \uc751\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uac15\uc88c\ub97c \ud1b5\ud574 CycleGAN\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c \uad6c\ud604 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub354 \ub9ce\uc740 \ud504\ub85c\uc81d\ud2b8\uc640 \uc2e4\ud5d8\uc744 \ud1b5\ud574 CycleGAN\uc758 \uc131\ub2a5\uc744 \uadf9\ub300\ud654\ud560 \uc218 \uc788\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4.\n<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"