\n \uc0dd\uc131\uc801 \uc801\ub300 \uc2e0\uacbd\ub9dd(Generative Adversarial Networks, GAN)\uc740 Ian Goodfellow\uac00 2014\ub144\uc5d0 \uc81c\uc548\ud55c \ud601\uc2e0\uc801\uc778 \ub525\ub7ec\ub2dd \ubaa8\ub378\ub85c, \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc774 \uc11c\ub85c \ub300\ub9bd\ud558\uba70 \ud559\uc2b5\ud558\ub294 \ubc29\ubc95\ub860\uc785\ub2c8\ub2e4. GAN\uc740 \uc8fc\ub85c \uc774\ubbf8\uc9c0 \uc0dd\uc131, \ud14d\uc2a4\ud2b8 \uc0dd\uc131, \ube44\ub514\uc624 \uc0dd\uc131 \ub4f1 \uc5ec\ub7ec \ubd84\uc57c\uc5d0\uc11c \uad11\ubc94\uc704\ud558\uac8c \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec GAN\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \uad6c\ud604 \ubc29\ubc95\uc5d0 \ub300\ud574 \ub2e8\uacc4\uc801\uc73c\ub85c \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n GAN\uc740 \uc0dd\uc131\uc790(Generator)\uc640 \ud310\ubcc4\uc790(Discriminator)\ub77c\ub294 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc73c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4. \uc0dd\uc131\uc790\uc758 \uc5ed\ud560\uc740 \uc9c4\uc9dc\ucc98\ub7fc \ubcf4\uc774\ub294 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \uac83\uc774\uace0, \ud310\ubcc4\uc790\ub294 \uc8fc\uc5b4\uc9c4 \ub370\uc774\ud130\uac00 \uc9c4\uc9dc\uc778\uc9c0 \uc0dd\uc131\uc790\uac00 \ub9cc\ub4e0 \uc704\uc870 \ub370\uc774\ud130\uc778\uc9c0 \ud310\ubcc4\ud558\ub294 \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4. \uc774 \ub450 \uc2e0\uacbd\ub9dd\uc740 \ub3d9\uc2dc\uc5d0 \ud559\uc2b5\ud558\uba70, \uc0dd\uc131\uc790\ub294 \ud310\ubcc4\uc790\ub97c \uc18d\uc774\uae30 \uc704\ud574 \ub354\uc6b1 \uc815\uad50\ud55c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub3c4\ub85d \ubc1c\uc804\ud558\uace0, \ud310\ubcc4\uc790\ub294 \ub354\uc6b1 \uc815\uad50\ud558\uac8c \uc704\uc870 \ub370\uc774\ud130\ub97c \uc2dd\ubcc4\ud558\ub3c4\ub85d \ubc1c\uc804\ud569\ub2c8\ub2e4.\n <\/p>\n
\n GAN\uc758 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc774 \ub2e8\uc21c\ud788 \uc124\uba85\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4:\n <\/p>\n
\n GAN\uc758 \ud559\uc2b5 \uacfc\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc774 \uc9c4\ud589\ub429\ub2c8\ub2e4:\n <\/p>\n
\n \uc774\uc81c PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c GAN\uc744 \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc774\uc6a9\ud574 \uc218\uce58 \ud615\ud0dc\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 GAN\uc744 \ub9cc\ub4dc\ub294 \uc791\uc5c5\uc744 \uc9c4\ud589\ud560 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n
\n \uccab \ubc88\uc9f8\ub85c, \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uace0 \uc784\ud3ec\ud2b8\ud569\ub2c8\ub2e4. PyTorch \ubc0f torchvision\uc744 \uc0ac\uc6a9\ud558\uc5ec \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud558\uace0 \ubaa8\ub378\uc744 \uad6c\ucd95\ud569\ub2c8\ub2e4.\n <\/p>\n
\n \n !pip install torch torchvision matplotlib\n <\/code>\n <\/pre>\n2.2 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\n MNIST \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud558\uace0 \ub370\uc774\ud130 \uc804\ucc98\ub9ac\ub97c \uc218\ud589\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uc774\ubbf8\uc9c0 \ub370\uc774\ud130\ub97c 0~1 \uc0ac\uc774\ub85c \uc2a4\ucf00\uc77c\ub9c1\ud558\uace0, \ubc30\uce58 \ub2e8\uc704\ub85c \ub098\ub204\ub3c4\ub85d \ud558\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \n import torch\n from torchvision import datasets, transforms\n from torch.utils.data import DataLoader\n\n # \ub370\uc774\ud130 \ub85c\ub529 \ubc0f \uc804\ucc98\ub9ac\n transform = transforms.Compose([\n transforms.ToTensor(),\n transforms.Normalize((0.5,), (0.5,))\n ])\n\n dataset = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\n dataloader = DataLoader(dataset, batch_size=64, shuffle=True)\n <\/code>\n <\/pre>\n2.3 \uc0dd\uc131\uc790(Generator) \ubc0f \ud310\ubcc4\uc790(Discriminator) \uc815\uc758<\/h3>\n
\n \ub2e4\uc74c\uc73c\ub85c, GAN\uc758 \ub450 \ud575\uc2ec \uad6c\uc131 \uc694\uc18c\uc778 \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub97c \uc815\uc758\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c \uc0dd\uc131\uc790\ub294 \ub79c\ub364 \ub178\uc774\uc988\ub97c \uc785\ub825\ubc1b\uc544 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\uace0, \ud310\ubcc4\uc790\ub294 \uc774\ubbf8\uc9c0\ub97c \uc785\ub825\ubc1b\uc544 \uadf8 \uc774\ubbf8\uc9c0\uac00 \uc9c4\uc9dc\uc778\uc9c0 \uc704\uc870\uc778\uc9c0 \ud310\ub2e8\ud569\ub2c8\ub2e4.\n <\/p>\n
\n \n import torch.nn as nn\n\n class 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),\n nn.Tanh()\n )\n\n def forward(self, z):\n return self.model(z).view(-1, 1, 28, 28)\n\n class 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()\n )\n\n def forward(self, img):\n return self.model(img)\n <\/code>\n <\/pre>\n2.4 \ubaa8\ub378 \ucd08\uae30\ud654 \ubc0f \uc190\uc2e4 \ud568\uc218, \uc635\ud2f0\ub9c8\uc774\uc800 \uc124\uc815<\/h3>\n
\n \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub97c \ucd08\uae30\ud654\ud558\uace0, \uc190\uc2e4 \ud568\uc218\uc640 \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \uc9c0\uc815\ud558\ub3c4\ub85d \ud558\uaca0\uc2b5\ub2c8\ub2e4. CrossEntropyLoss \ubc0f Adam \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n
\n \n generator = Generator()\n discriminator = Discriminator()\n\n ad = torch.optim.Adam(discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999))\n ag = torch.optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))\n\n criterion = nn.BCELoss()\n <\/code>\n <\/pre>\n2.5 GAN \ud559\uc2b5<\/h3>\n
\n \uc774\uc81c GAN\uc744 \ud559\uc2b5\uc2dc\ucf1c\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uac01 \uc5d0\ud3ed\ub9c8\ub2e4 \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub97c \ud559\uc2b5\uc2dc\ud0a4\uace0, \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \n import matplotlib.pyplot as plt\n import numpy as np\n\n def train_gan(generator, discriminator, criterion, ag, ad, dataloader, epochs=50):\n for epoch in range(epochs):\n for real_imgs, _ in dataloader:\n batch_size = real_imgs.size(0)\n\n # \uc9c4\uc9dc \uc774\ubbf8\uc9c0\uc640 \ub808\uc774\ube14 \uc0dd\uc131\n real_labels = torch.ones(batch_size, 1)\n noise = torch.randn(batch_size, 100)\n fake_imgs = generator(noise)\n fake_labels = torch.zeros(batch_size, 1)\n\n # \ud310\ubcc4\uc790 \ud559\uc2b5\n discriminator.zero_grad()\n real_loss = criterion(discriminator(real_imgs), real_labels)\n fake_loss = criterion(discriminator(fake_imgs.detach()), fake_labels)\n d_loss = real_loss + fake_loss\n d_loss.backward()\n ad.step()\n\n # \uc0dd\uc131\uc790 \ud559\uc2b5\n generator.zero_grad()\n g_loss = criterion(discriminator(fake_imgs), real_labels)\n g_loss.backward()\n ag.step()\n\n print(f'Epoch [{epoch + 1}\/{epochs}], D Loss: {d_loss.item():.4f}, G Loss: {g_loss.item():.4f}')\n\n # \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0 \uc800\uc7a5\n if (epoch + 1) % 10 == 0:\n save_generated_images(generator, epoch + 1)\n\n def save_generated_images(generator, epoch):\n noise = torch.randn(64, 100)\n generated_imgs = generator(noise)\n generated_imgs = generated_imgs.detach().numpy()\n generated_imgs = (generated_imgs + 1) \/ 2 # Rescale to [0, 1]\n\n fig, axs = plt.subplots(8, 8, figsize=(8, 8))\n for i, ax in enumerate(axs.flat):\n ax.imshow(generated_imgs[i][0], cmap='gray')\n ax.axis('off')\n plt.savefig(f'generated_images_epoch_{epoch}.png')\n plt.close()\n\n train_gan(generator, discriminator, criterion, ag, ad, dataloader, epochs=50)\n <\/code>\n <\/pre>\n2.6 \uacb0\uacfc \ud655\uc778<\/h3>\n
\n \ud559\uc2b5\uc774 \uc644\ub8cc\ub41c \ud6c4, \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \ud655\uc778\ud574\ubcf4\uc138\uc694. GAN\uc740 \ubc18\ubcf5 \ud559\uc2b5\uc744 \ud1b5\ud574 \uc810\uc810 \ub354 \uc2e4\uc81c\uc640 \uc720\uc0ac\ud55c \ub370\uc774\ud130 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\uac8c \ub429\ub2c8\ub2e4. \uacb0\uacfc\uc801\uc73c\ub85c, GAN\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\ub294 \uac83\uc740 \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\uc758 \ud488\uc9c8\uc785\ub2c8\ub2e4. \ud6c8\ub828\uc774 \uc798\ub418\uc5c8\uc744 \uacbd\uc6b0, \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub4e4\uc774 \uc0dd\uc18c\ud558\uba74\uc11c\ub3c4 \uc544\ub984\ub2e4\uc6b4 \ud615\ud0dc\ub97c \uac00\uc9c8 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n
3. \uacb0\ub860<\/h2>\n
\n \uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 PyTorch\ub97c \ud65c\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc124\uba85\ud588\uc2b5\ub2c8\ub2e4. GAN\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \ud568\uaed8 \uc2e4\uc81c \ucf54\ub4dc\ub97c \ud1b5\ud574 \uc790\uc2e0\ub9cc\uc758 GAN\uc744 \ub9cc\ub4e4\uc5b4\ubcf4\ub294 \uacbd\ud5d8\uc744 \ud560 \uc218 \uc788\uc5c8\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4. GAN\uc740 \uac15\ub825\ud55c \ub3c4\uad6c\uc774\uc9c0\ub9cc, \uacac\uace0\ud55c \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\ub294 \ub370\uc5d0\ub294 \ub2e4\uc591\ud558\uace0 \uc2ec\ub3c4 \uc788\ub294 \uc5f0\uad6c\uac00 \ud544\uc694\ud569\ub2c8\ub2e4. \uc544\ub984\ub2f5\uace0 \ucc3d\uc758\uc801\uc778 \uacb0\uacfc\ub97c \ub9cc\ub4e4\uc5b4\ub0b4\ub294 GAN\uc758 \uc138\uc0c1\uc73c\ub85c \uc5ec\ub7ec\ubd84\uc744 \ucd08\ub300\ud569\ub2c8\ub2e4!\n <\/p>\n
4. \ucc38\uace0 \uc790\ub8cc<\/h2>\n\n- Ian Goodfellow et al. “Generative Adversarial Networks”. NIPS 2014.<\/li>\n
- PyTorch Documentation: https:\/\/pytorch.org\/docs\/stable\/index.html<\/a><\/li>\n
- GANs in Action book<\/li>\n<\/ul>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"