\ucd5c\uadfc \uba87 \ub144 \uac04 \ub525\ub7ec\ub2dd\uc740 \uc774\ubbf8\uc9c0 \uc0dd\uc131, \ubcc0\ud658, \ubd84\ud560 \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud601\uc2e0\uc801\uc778 \ubc1c\uc804\uc744 \uc774\ub8e8\uc5c8\uc2b5\ub2c8\ub2e4. \uadf8\uc911\uc5d0\uc11c\ub3c4 GAN(Generative Adversarial Network)\uc740 \uc774\ubbf8\uc9c0 \uc0dd\uc131\uc758 \uc0c8\ub85c\uc6b4 \uac00\ub2a5\uc131\uc744 \uc5f4\uc5b4\uc8fc\uc5c8\uc2b5\ub2c8\ub2e4. GAN\uc740 \uc0dd\uc131\uc790(Generator)\uc640 \ud310\ubcc4\uc790(Discriminator)\ub85c \uad6c\uc131\ub41c \ub450 \uac1c\uc758 \ub124\ud2b8\uc6cc\ud06c\uac00 \uc11c\ub85c \uacbd\uc7c1\ud558\uba70 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \uad6c\uc870\uc785\ub2c8\ub2e4. \ubcf8 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 GAN\uc758 \uac1c\uc694 \ubc0f \ud30c\uc774\ud1a0\uce58(PyTorch) \ud504\ub808\uc784\uc6cc\ud06c\ub97c \uc774\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud558\uae30 \uc704\ud55c \ud658\uacbd \uc124\uc815 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
GAN\uc740 Ian Goodfellow\uac00 2014\ub144\uc5d0 \uc81c\uc548\ud55c \ubaa8\ub378\ub85c, \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc774 \uc0c1\ud638\uc791\uc6a9\ud558\uc5ec \ud6c8\ub828\ub418\ub294 \ubc29\uc2dd\uc785\ub2c8\ub2e4. \uc0dd\uc131\uc790\ub294 \uc2e4\uc81c\uc640 \uc720\uc0ac\ud55c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \ud310\ubcc4\uc790\ub294 \uc0dd\uc131\ub41c \ub370\uc774\ud130\uac00 \uc2e4\uc81c \ub370\uc774\ud130\uc778\uc9c0 \uc544\ub2cc\uc9c0\ub97c \ud310\ub2e8\ud569\ub2c8\ub2e4. \uc774 \ub450 \ub124\ud2b8\uc6cc\ud06c\ub294 \uc11c\ub85c \uacac\uc81c\ud558\uba70 \uc810\uc810 \ub354 \ubc1c\uc804\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
GAN\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \ub450 \uac00\uc9c0 \uad6c\uc131 \uc694\uc18c\ub85c \uc774\ub8e8\uc5b4\uc838 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
GAN\uc758 \ubaa9\ud45c\ub294 Minimax \uac8c\uc784\uc73c\ub85c \ud45c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc0dd\uc131\uc790\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ubaa9\ud45c\ub97c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
G^{*} = arg \\min_{G} \\max_{D} V(D, G) = E_{x \\sim pdata(x)}[\\log D(x)] + E_{z \\sim pz(z)}[\\log(1 - D(G(z)))]<\/code><\/pre>\n\uc5ec\uae30\uc11c G\ub294 \uc0dd\uc131\uc790, D\ub294 \ud310\ubcc4\uc790\ub97c \uc758\ubbf8\ud558\uba70, pdata<\/sub>(x)\ub294 \uc2e4\uc81c \ub370\uc774\ud130\uc758 \ubd84\ud3ec, pz<\/sub>(z)\ub294 \uc0dd\uc131\uc790\uac00 \uc0ac\uc6a9\ud558\ub294 \ub178\uc774\uc988 \ubd84\ud3ec\uc785\ub2c8\ub2e4.<\/p>\n2. \ud30c\uc774\ud1a0\uce58(PyTorch) \ud658\uacbd \uc124\uc815<\/h2>\n
\ud30c\uc774\ud1a0\uce58\ub294 \ud150\uc11c \uc5f0\uc0b0, \uc790\ub3d9 \ubbf8\ubd84, \uadf8\ub9ac\uace0 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \uc190\uc27d\uac8c \uad6c\ucd95\ud560 \uc218 \uc788\ub294 \uc5ec\ub7ec \ub3c4\uad6c\ub97c \uc81c\uacf5\ud558\ub294 \uc624\ud508\uc18c\uc2a4 \uba38\uc2e0\ub7ec\ub2dd \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. GAN \uad6c\ud604\uc744 \uc704\ud574 \ud30c\uc774\ud1a0\uce58\ub97c \uc124\uce58\ud558\uace0 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uad6c\uc131\ud558\ub294 \ubc29\ubc95\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4.<\/p>\n
2.1 \ud30c\uc774\ud1a0\uce58 \uc124\uce58\ud558\uae30<\/h3>\n
\ud30c\uc774\ud1a0\uce58\ub294 CUDA\ub97c \uc9c0\uc6d0\ud558\uc5ec NVIDIA GPU\uc5d0\uc11c \ud6a8\uc728\uc801\uc73c\ub85c \ub3d9\uc791\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\uc758 \uba85\ub839\uc5b4\ub97c \ud1b5\ud574 \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
pip install torch torchvision torchaudio<\/code><\/pre>\n\ub9cc\uc57d CUDA\ub97c \uc0ac\uc6a9\ud558\uace0 \uc788\ub2e4\uba74, \ud30c\uc774\ud1a0\uce58 \uacf5\uc2dd \ud648\ud398\uc774\uc9c0\uc5d0\uc11c \ud658\uacbd\uc5d0 \ub9de\ub294 \uc124\uce58 \uba85\ub839\uc5b4\ub97c \ud655\uc778\ud558\uc5ec \uc124\uce58\ud558\uae30 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n
2.2 \ucd94\uac00 \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58\ud558\uae30<\/h3>\n
\uc774\ubbf8\uc9c0 \ucc98\ub9ac\ub97c \uc704\ud574 \ud544\uc694\ud55c \ucd94\uac00 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub3c4 \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. \uc544\ub798 \uba85\ub839\uc5b4\ub97c \ud1b5\ud574 \uc124\uce58\ud569\ub2c8\ub2e4:<\/p>\n
pip install matplotlib numpy<\/code><\/pre>\n2.3 \uae30\ubcf8\uc801\uc778 \ub514\ub809\ud1a0\ub9ac \uad6c\uc870 \uc124\uc815\ud558\uae30<\/h3>\n
\ud504\ub85c\uc81d\ud2b8 \ub514\ub809\ud1a0\ub9ac\ub97c \ub2e4\uc74c\uacfc \uac19\uc740 \uad6c\uc870\ub85c \ub9cc\ub4e4\uc5b4 \uc90d\ub2c8\ub2e4:<\/p>\n
\n gan_project\/\n \u251c\u2500\u2500 dataset\/\n \u251c\u2500\u2500 models\/\n \u251c\u2500\u2500 results\/\n \u2514\u2500\u2500 train.py\n <\/code><\/pre>\n\uac01\uac01\uc758 \ub514\ub809\ud1a0\ub9ac\ub294 \ub370\uc774\ud130\uc14b, \ubaa8\ub378, \uacb0\uacfc\ubb3c\uc744 \uc800\uc7a5\ud558\ub294 \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4. train.py<\/code> \ud30c\uc77c\uc740 GAN\uc744 \ud559\uc2b5\uc2dc\ud0a4\uace0 \ud3c9\uac00\ud558\ub294 \uc2a4\ud06c\ub9bd\ud2b8\ub97c \ub2f4\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n3. GAN \uad6c\ud604\uc5d0 \ud544\uc694\ud55c \ucf54\ub4dc \uc608\uc81c<\/h2>\n
\uc774\uc81c GAN\uc744 \uad6c\ud604\ud558\uae30 \uc704\ud55c \uae30\ubcf8\uc801\uc778 \ucf54\ub4dc\ub97c \uc791\uc131\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \ucf54\ub4dc\ub294 \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub97c \uc815\uc758\ud558\uace0, \uc774\ub97c \uc774\uc6a9\ud558\uc5ec GAN\uc744 \ud6c8\ub828\ud558\ub294 \uacfc\uc815\uc744 \ud3ec\ud568\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
3.1 \ubaa8\ub378 \uc815\uc758\ud558\uae30<\/h3>\n
\uba3c\uc800 \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790 \ub124\ud2b8\uc6cc\ud06c\ub97c \uc815\uc758\ud569\ub2c8\ub2e4. \uc544\ub798 \ucf54\ub4dc\ub294 \uac04\ub2e8\ud55c CNN(\ucee8\ubcfc\ub8e8\uc158 \uc2e0\uacbd\ub9dd)\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub97c \uad6c\ucd95\ud558\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4:<\/p>\n
import torch\nimport torch.nn as nn\n\n# \uc0dd\uc131\uc790 \ubaa8\ub378\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, 28 * 28), # MNIST \uc774\ubbf8\uc9c0 \ud06c\uae30\n nn.Tanh() # [-1, 1]\ub85c \uc815\uaddc\ud654\n )\n \n def forward(self, z):\n return self.model(z).view(-1, 1, 28, 28)\n\n# \ud310\ubcc4\uc790 \ubaa8\ub378\nclass Discriminator(nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.model = nn.Sequential(\n nn.Flatten(),\n nn.Linear(28 * 28, 512),\n nn.LeakyReLU(0.2),\n nn.Linear(512, 256),\n nn.LeakyReLU(0.2),\n nn.Linear(256, 1),\n nn.Sigmoid() # [0, 1]\ub85c \uc815\uaddc\ud654\n )\n \n def forward(self, img):\n return self.model(img)\n <\/code><\/pre>\n3.2 \ub370\uc774\ud130\uc14b \uc900\ube44\ud558\uae30<\/h3>\n
MNIST \ub370\uc774\ud130\uc14b\uc744 \uac00\uc838\uc624\uace0 \uc804\ucc98\ub9ac\ud569\ub2c8\ub2e4. torchvision \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc774\uc6a9\ud558\uba74 \uc27d\uac8c \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
from torchvision import datasets, transforms\n\n# \ub370\uc774\ud130 \uc804\ucc98\ub9ac\ntransform = transforms.Compose([\n transforms.ToTensor(),\n transforms.Normalize((0.5,), (0.5,))\n])\n\n# MNIST \ub370\uc774\ud130\uc14b \ub85c\ub4dc\ndataloader = torch.utils.data.DataLoader(\n datasets.MNIST('dataset\/', download=True, transform=transform),\n batch_size=64,\n shuffle=True\n) \n <\/code><\/pre>\n3.3 GAN \ud6c8\ub828 \ucf54\ub4dc<\/h3>\n
\uc774\uc81c \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub97c \ud6c8\ub828\ud560 \uc218 \uc788\ub294 \ub8e8\ud504\ub97c \ub9cc\ub4e4\uc5b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
import torch.optim as optim\n\n# \ubaa8\ub378 \ucd08\uae30\ud654\ngenerator = Generator()\ndiscriminator = Discriminator()\n\n# \uc190\uc2e4 \ud568\uc218 \ubc0f \uc635\ud2f0\ub9c8\uc774\uc800\ncriterion = nn.BCELoss()\noptimizer_g = optim.Adam(generator.parameters(), lr=0.0002)\noptimizer_d = optim.Adam(discriminator.parameters(), lr=0.0002)\n\nnum_epochs = 50\nfor epoch in range(num_epochs):\n for i, (imgs, _) in enumerate(dataloader):\n batch_size = imgs.size(0)\n imgs = imgs.view(batch_size, -1)\n\n # \uc9c4\uc9dc\uc640 \uac00\uc9dc \ub808\uc774\ube14 \uc0dd\uc131\n real_labels = torch.ones(batch_size, 1)\n fake_labels = torch.zeros(batch_size, 1)\n\n # \ud310\ubcc4\uc790 \ud6c8\ub828\n optimizer_d.zero_grad()\n \n outputs = discriminator(imgs)\n d_loss_real = criterion(outputs, real_labels)\n d_loss_real.backward()\n \n z = torch.randn(batch_size, 100)\n fake_images = generator(z)\n outputs = discriminator(fake_images)\n d_loss_fake = criterion(outputs, fake_labels)\n d_loss_fake.backward()\n \n optimizer_d.step()\n\n # \uc0dd\uc131\uc790 \ud6c8\ub828\n optimizer_g.zero_grad()\n z = torch.randn(batch_size, 100)\n fake_images = generator(z)\n outputs = discriminator(fake_images)\n g_loss = criterion(outputs, real_labels)\n g_loss.backward()\n \n optimizer_g.step()\n\n if i % 100 == 0:\n print(f\"[Epoch {epoch}\/{num_epochs}] [Batch {i}\/{len(dataloader)}] \"\n f\"[D loss: {d_loss_real.item() + d_loss_fake.item()}] \"\n f\"[G loss: {g_loss.item()}]\")\n <\/code><\/pre>\n4. \uacb0\uacfc \uc2dc\uac01\ud654<\/h2>\n
\ud6c8\ub828\uc774 \uc644\ub8cc\ub41c \ud6c4\uc5d0 \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \uc2dc\uac01\ud654\ud558\ub294 \uac83\ub3c4 \uc911\uc694\ud569\ub2c8\ub2e4. matplotlib\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \ucd9c\ub825\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
import matplotlib.pyplot as plt\n\ndef generate_and_plot_images(generator, n_samples=25):\n z = torch.randn(n_samples, 100)\n generated_images = generator(z).detach().numpy()\n\n plt.figure(figsize=(5, 5))\n for i in range(n_samples):\n plt.subplot(5, 5, i + 1)\n plt.imshow(generated_images[i][0], cmap='gray')\n plt.axis('off')\n plt.show()\n\ngenerate_and_plot_images(generator)\n <\/code><\/pre>\n5. \ub9c8\uce58\uba70<\/h2>\n
\uc774\ubc88 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 GAN\uc758 \uc6d0\ub9ac\uc640 \uae30\ubcf8 \uad6c\uc870\ub97c \uc124\uba85\ud558\uace0, \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud558\uae30 \uc704\ud55c \ud658\uacbd \uc124\uc815\uacfc \ucf54\ub4dc \uc608\uc81c\ub97c \uc81c\uacf5\ud588\uc2b5\ub2c8\ub2e4. GAN\uc740 \ub9e4\uc6b0 \uac15\ub825\ud55c \uc0dd\uc131 \ubaa8\ub378\uc774\uba70, \ub2e4\uc591\ud55c \uc751\uc6a9 \ubd84\uc57c\ub97c \uac16\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c GAN\uc744 \ud65c\uc6a9\ud55c \ub2e4\uc591\ud55c \ud504\ub85c\uc81d\ud2b8\ub97c \uc2dc\ub3c4\ud574\ubcf4\uae38 \uad8c\uc7a5\ud569\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0 \uc790\ub8cc<\/h2>\n