\n \ub525\ub7ec\ub2dd\uc740 \ucef4\ud4e8\ud130 \ube44\uc804, \uc790\uc5f0\uc5b4 \ucc98\ub9ac, \uc74c\uc131 \uc778\uc2dd \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud601\uc2e0\uc801\uc778 \ubc1c\uc804\uc744 \uc774\ub8e8\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\uc911\uc5d0\uc11c\ub3c4 \uc0dd\uc131\uc801 \uc801\ub300 \uc2e0\uacbd\ub9dd(Generative Adversarial Network, GAN)\uc740 \ud2b9\ubcc4\ud55c \uc8fc\ubaa9\uc744 \ubc1b\uace0 \uc788\ub294 \uae30\uc220\uc785\ub2c8\ub2e4. GAN\uc740 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd, \uc989 \uc0dd\uc131\uc790(Generator)\uc640 \ud310\ubcc4\uc790(Discriminator)\uac00 \uc11c\ub85c \uacbd\uc7c1\ud558\ub294 \uad6c\uc870\ub97c \uac00\uc9c0\uace0 \uc788\uc73c\uba70, \uc774\ub97c \ud1b5\ud574 \uc0ac\uc2e4\uc801\uc778 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \ub2a5\ub825\uc744 \uac16\ucd94\uac8c \ub429\ub2c8\ub2e4.\n <\/p>\n
\n \uc774\ubc88 \uae00\uc5d0\uc11c\ub294 GAN\uc758 \ubcc0\ud615 \uc911 \ud558\ub098\uc778 \uc870\uac74\ubd80 \uc0dd\uc131\uc801 \uc801\ub300 \uc2e0\uacbd\ub9dd(Conditional GAN, cGAN)\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. cGAN\uc740 \uc0dd\uc131 \uacfc\uc815\uc5d0 \uc870\uac74\uc744 \uc8fc\uc5b4 \ud2b9\uc815 \ud074\ub798\uc2a4\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\ub3c4\ub85d \ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uc22b\uc790 \uc774\ubbf8\uc9c0 \ub370\uc774\ud130\uc14b\uc778 MNIST\ub97c \ud65c\uc6a9\ud574 \ud2b9\uc815 \uc22b\uc790\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n GAN\uc740 \uae30\ubcf8\uc801\uc73c\ub85c \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc73c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4. \uc0dd\uc131\uc790\ub294 \ubb34\uc791\uc704 \ub178\uc774\uc988 \ubca1\ud130\ub97c \uc785\ub825\ubc1b\uc544 \uac00\uc9dc \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\uace0, \ud310\ubcc4\uc790\ub294 \uc785\ub825\ub41c \uc774\ubbf8\uc9c0\uac00 \uc9c4\uc9dc\uc778\uc9c0 \uac00\uc9dc\uc778\uc9c0 \ud310\ubcc4\ud558\ub294 \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4. \uc774 \ub458\uc740 \ub2e4\uc74c\uacfc \uac19\uc774 \uc0c1\ud638\uc791\uc6a9\ud569\ub2c8\ub2e4:\n <\/p>\n
\n cGAN\uc740 GAN\uc758 \uac1c\ub150\uc744 \ud655\uc7a5\ud558\uc5ec, \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790 \ubaa8\ub450\uc5d0 \uc870\uac74 \uc815\ubcf4\ub97c \ucd94\uac00\ud569\ub2c8\ub2e4. \uc774\ub294 \ud2b9\uc815 \ud074\ub798\uc2a4\uc5d0 \ub300\ud55c \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\ub3c4\ub85d \ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uc22b\uc790 \uc774\ubbf8\uc9c0 \uc0dd\uc131\uc5d0\uc11c \uc22b\uc790 ‘3’\uc744 \uc870\uac74\uc73c\ub85c \uc124\uc815\ud558\uba74, \uc0dd\uc131\uc790\ub294 ‘3’\uc5d0 \ud574\ub2f9\ud558\ub294 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\uac8c \ub429\ub2c8\ub2e4. cGAN\uc758 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:\n <\/p>\n
\n cGAN\uc744 \uad6c\ud604\ud558\uae30 \uc704\ud574 \ud544\uc694\ud55c \ud30c\uc774\uc36c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud569\ub2c8\ub2e4. \uae30\ubcf8\uc801\uc73c\ub85c PyTorch\uc640 NumPy, Matplotlib \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. \uc544\ub798 \uba85\ub839\uc5b4\ub85c \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \n pip install torch torchvision numpy matplotlib\n <\/code>\n <\/pre>\n3.2 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\n cGAN\uc744 \uad6c\ud604\ud558\uae30 \uc704\ud574 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. MNIST\ub294 0\ubd80\ud130 9\uae4c\uc9c0\uc758 \uc190\uae00\uc528 \uc22b\uc790 \uc774\ubbf8\uc9c0\ub85c \uad6c\uc131\ub41c \ub370\uc774\ud130\uc14b\uc785\ub2c8\ub2e4. PyTorch\uc758 torchvision\uc5d0\uc11c \ud574\ub2f9 \ub370\uc774\ud130\uc14b\uc744 \ubd88\ub7ec\uc62c \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \nimport torch\nfrom torchvision import datasets, transforms\n\n# \ub370\uc774\ud130\uc14b \ubd88\ub7ec\uc624\uae30\ntransform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])\ntrain_dataset = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=64, shuffle=True)\n <\/code>\n <\/pre>\n4. cGAN \uc544\ud0a4\ud14d\ucc98 \uad6c\ud604<\/h2>\n4.1 \uc0dd\uc131\uc790(Generator)<\/h3>\n
\n \uc0dd\uc131\uc790\ub294 \ub79c\ub364 \ub178\uc774\uc988\uc640 \uc870\uac74 \uc815\ubcf4\ub97c \uc785\ub825\uc73c\ub85c \ubc1b\uc544 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4. \uc0dd\uc131\uc790 \ubaa8\ub378\uc740 \uc8fc\ub85c \uc5ec\ub7ec \uac1c\uc758 \uc120\ud615 \uacc4\uce35\uacfc ReLU \ud65c\uc131\ud654 \ud568\uc218\ub97c \uc0ac\uc6a9\ud558\uc5ec \uad6c\ucd95\ub429\ub2c8\ub2e4.\n <\/p>\n
\n \nimport torch.nn as nn\n\nclass Generator(nn.Module):\n def __init__(self, z_dim, num_classes):\n super(Generator, self).__init__()\n self.label_embedding = nn.Embedding(num_classes, num_classes)\n self.model = nn.Sequential(\n nn.Linear(z_dim + num_classes, 128),\n nn.ReLU(),\n nn.Linear(128, 256),\n nn.ReLU(),\n nn.Linear(256, 512),\n nn.ReLU(),\n nn.Linear(512, 1 * 28 * 28),\n nn.Tanh()\n )\n\n def forward(self, noise, labels):\n label_input = self.label_embedding(labels)\n input = torch.cat((noise, label_input), dim=1)\n img = self.model(input)\n img = img.view(img.size(0), 1, 28, 28)\n return img\n <\/code>\n <\/pre>\n4.2 \ud310\ubcc4\uc790(Discriminator)<\/h3>\n
\n \ud310\ubcc4\uc790\ub294 \uc774\ubbf8\uc9c0\uc640 \uc870\uac74 \uc815\ubcf4\ub97c \ud568\uaed8 \ubc1b\uc544\ub4e4\uc5ec \uc9c4\uc9dc\uc640 \uac00\uc9dc\ub97c \ud310\ubcc4\ud558\ub294 \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4. \ubc14\ub2e5\uce35\uc5d0\uc11c \uc2dc\uc791\ud558\uc5ec \uc810\ucc28 \uae4a\uc5b4\uc9c0\ub294 \uad6c\uc870\ub85c \uc124\uacc4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \nclass Discriminator(nn.Module):\n def __init__(self, num_classes):\n super(Discriminator, self).__init__()\n self.label_embedding = nn.Embedding(num_classes, num_classes)\n self.model = nn.Sequential(\n nn.Linear(1 * 28 * 28 + num_classes, 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, labels):\n label_input = self.label_embedding(labels)\n img_flat = img.view(img.size(0), -1)\n input = torch.cat((img_flat, label_input), dim=1)\n validity = self.model(input)\n return validity\n <\/code>\n <\/pre>\n5. \uc190\uc2e4 \ud568\uc218 \ubc0f \ucd5c\uc801\ud654<\/h2>\n
\n cGAN\uc758 \uc190\uc2e4 \ud568\uc218\ub294 \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud569\ub2c8\ub2e4. \uc8fc\ub85c \uc774\uc9c4_cross-entropy \uc190\uc2e4 \ud568\uc218\ub97c \uc0ac\uc6a9\ud558\uba70, \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub294 \uc11c\ub85c \ubc18\ub300\ub418\ub294 \ubaa9\ud45c\ub97c \uac00\uc9d1\ub2c8\ub2e4.\n <\/p>\n
\n \nimport torch.optim as optim\n\ndef build_optimizers(generator, discriminator, lr=0.0002, beta1=0.5):\n g_optimizer = optim.Adam(generator.parameters(), lr=lr, betas=(beta1, 0.999))\n d_optimizer = optim.Adam(discriminator.parameters(), lr=lr, betas=(beta1, 0.999))\n return g_optimizer, d_optimizer\n <\/code>\n <\/pre>\n6. cGAN \ud6c8\ub828<\/h2>\n
\n \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub294 \uc11c\ub85c \uacbd\uc7c1\ud558\uba70 \ud6c8\ub828\ub429\ub2c8\ub2e4. \ub9e4 \ubc18\ubcf5\ub9c8\ub2e4 \ud310\ubcc4\uc790\ub294 \uc2e4\uc81c \uc774\ubbf8\uc9c0\uc5d0 \ub300\ud574 \ub192\uc740 \uc2e0\ub8b0\ub3c4\ub97c \ubcf4\uc774\ub294 \ub3d9\uc2dc\uc5d0 \uc0dd\uc131\uc790\uac00 \ub9cc\ub4e0 \uc774\ubbf8\uc9c0\uc5d0 \ub300\ud574\uc11c\ub294 \ub0ae\uc740 \uc2e0\ub8b0\ub3c4\ub97c \uac00\uc9c0\ub3c4\ub85d \uc218\uc815\ub429\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \ud6c8\ub828 \ub8e8\ud504 \uc608\uc81c\uc785\ub2c8\ub2e4.\n <\/p>\n
\n \nnum_classes = 10\nz_dim = 100\n\ngenerator = Generator(z_dim, num_classes)\ndiscriminator = Discriminator(num_classes)\n\ng_optimizer, d_optimizer = build_optimizers(generator, discriminator)\n\ncriterion = nn.BCELoss()\n\n# \ud6c8\ub828 \ub8e8\ud504\nnum_epochs = 200\nfor epoch in range(num_epochs):\n for imgs, labels in train_loader:\n batch_size = imgs.size(0)\n\n # \uc9c4\uc9dc \uc774\ubbf8\uc9c0\uc640 \uac00\uc9dc \uc774\ubbf8\uc9c0 \ub808\uc774\ube14 \uc900\ube44\n real_labels = torch.ones(batch_size, 1)\n fake_labels = torch.zeros(batch_size, 1)\n\n # \ud310\ubcc4\uc790 \ud6c8\ub828\n discriminator.zero_grad()\n outputs = discriminator(imgs, labels)\n d_loss_real = criterion(outputs, real_labels)\n d_loss_real.backward()\n\n noise = torch.randn(batch_size, z_dim)\n random_labels = torch.randint(0, num_classes, (batch_size,))\n generated_imgs = generator(noise, random_labels)\n\n outputs = discriminator(generated_imgs, random_labels)\n d_loss_fake = criterion(outputs, fake_labels)\n d_loss_fake.backward()\n\n d_optimizer.step()\n d_loss = d_loss_real + d_loss_fake\n \n # \uc0dd\uc131\uc790 \ud6c8\ub828\n generator.zero_grad()\n noise = torch.randn(batch_size, z_dim)\n generated_imgs = generator(noise, random_labels)\n outputs = discriminator(generated_imgs, random_labels)\n g_loss = criterion(outputs, real_labels)\n g_loss.backward()\n g_optimizer.step()\n\n print(f'Epoch [{epoch}\/{num_epochs}], d_loss: {d_loss.item()}, g_loss: {g_loss.item()}')\n <\/code>\n <\/pre>\n7. \uacb0\uacfc \uc2dc\uac01\ud654<\/h2>\n
\n \ud6c8\ub828\uc774 \uc644\ub8cc\ub41c \ud6c4, \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \uc2dc\uac01\ud654\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. Matplotlib\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud2b9\uc815 \ud074\ub798\uc2a4\uc758 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\uace0 \ubcf4\uc5ec\uc904 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \nimport matplotlib.pyplot as plt\n\ndef generate_and_show_images(generator, num_images=10):\n noise = torch.randn(num_images, z_dim)\n labels = torch.randint(0, num_classes, (num_images,))\n generated_images = generator(noise, labels)\n\n for i in range(num_images):\n img = generated_images[i].detach().numpy().reshape(28, 28)\n plt.subplot(2, 5, i + 1)\n plt.imshow(img, cmap='gray')\n plt.axis('off')\n plt.show()\n\ngenerate_and_show_images(generator)\n <\/code>\n <\/pre>\n8. \uacb0\ub860<\/h2>\n
\n \ubcf8 \uae00\uc5d0\uc11c\ub294 \uc870\uac74\ubd80 \uc0dd\uc131\uc801 \uc801\ub300 \uc2e0\uacbd\ub9dd(cGAN)\uc758 \uac1c\ub150 \ubc0f \uad6c\ud604 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. cGAN\uc740 \ud2b9\uc815 \uc870\uac74\uc5d0 \ub530\ub77c \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\ub294 \uac15\ub825\ud55c \ubc29\ubc95\uc73c\ub85c, \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \uc751\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \uac83\ubfd0\ub9cc \uc544\ub2c8\ub77c \uc774\ubbf8\uc9c0 \ubcc0\ud658, \uc2a4\ud0c0\uc77c \uc804\uc774 \ub4f1\uc758 \uc791\uc5c5\uc5d0\uc11c\ub3c4 \ud65c\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec cGAN\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc54c\uc544\ubcf4\uc558\uc73c\ubbc0\ub85c, \uc55e\uc73c\ub85c \ub354 \ubc1c\uc804\ub41c \ubaa8\ub378\uc774\ub098 \ub2e4\uc591\ud55c \uc751\uc6a9\uc774 \uc774\ub8e8\uc5b4\uc9c8 \uc218 \uc788\uae30\ub97c \uae30\ub300\ud569\ub2c8\ub2e4.\n <\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"