\ucd5c\uadfc \uba87 \ub144 \uac04 \uc778\uacf5\uc9c0\ub2a5 \ubd84\uc57c\uc5d0\uc11c \uc0dd\uc131\uc801 \uc801\ub300 \uc2e0\uacbd\ub9dd(Generative Adversarial Networks, GAN)\uacfc \ubcc0\ubd84 \uc624\ud1a0\uc778\ucf54\ub354(Variational Autoencoder, VAE)\ub294 \ub370\uc774\ud130 \uc0dd\uc131 \ubc0f \ubcc0\ud615\uc5d0\uc11c \ud601\uc2e0\uc801\uc778 \uae30\uc220\ub85c \uc790\ub9ac \uc7a1\uc558\uc2b5\ub2c8\ub2e4. \uc774\ub4e4 \ubaa8\ub378\uc740 \uc11c\ub85c \ub2e4\ub978 \ubc29\uc2dd\uc73c\ub85c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294\ub370, GAN\uc740 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc774 \uacbd\uc7c1\ud558\ub294 \uad6c\uc870\ub85c \uc774\ub8e8\uc5b4\uc838 \uc788\uace0, VAE\ub294 \ud655\ub960\uc801 \ubaa8\ub378\ub85c \ub370\uc774\ud130\ub97c \uc555\ucd95\ud558\uace0 \uc0dd\uc131\ud558\ub294 \ubc29\uc2dd\uc73c\ub85c \uc791\ub3d9\ud569\ub2c8\ub2e4.<\/p>\n
GAN\uc740 Ian Goodfellow\uac00 2014\ub144\uc5d0 \uc81c\uc548\ud55c \ubaa8\ub378\ub85c, \uc0dd\uc131\uae30(Generator)\uc640 \ud310\ubcc4\uae30(Discriminator)\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4. \uc0dd\uc131\uae30\ub294 \ubb34\uc791\uc704 \ub178\uc774\uc988\ub97c \uc785\ub825\ubc1b\uc544 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \ud310\ubcc4\uae30\ub294 \uc785\ub825\ubc1b\uc740 \ub370\uc774\ud130\uac00 \uc9c4\uc9dc\uc778\uc9c0 \uac00\uc9dc\uc778\uc9c0 \ud310\ub2e8\ud569\ub2c8\ub2e4. \uc774 \ub450 \ub124\ud2b8\uc6cc\ud06c\ub294 \uc11c\ub85c \uacbd\uc7c1\ud558\uba70 \ud559\uc2b5\ud558\uac8c \ub418\uba70, \uc774 \uacfc\uc815\uc5d0\uc11c \uc0dd\uc131\uae30\ub294 \uc810\uc810 \ub354 \uc0ac\uc2e4\uc801\uc778 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
GAN\uc758 \ud6c8\ub828 \uacfc\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\ubcc0\ubd84 \uc624\ud1a0\uc778\ucf54\ub354(VAE)\ub294 \uc624\ud1a0\uc778\ucf54\ub354\uc758 \ubcc0\ud615\uc73c\ub85c, \ub370\uc774\ud130\uc758 \ubd84\ud3ec\ub97c \ubaa8\ub378\ub9c1\ud558\uc5ec \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \ub2a5\ub825\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. VAE\ub294 \uc778\ucf54\ub354(Encoder)\uc640 \ub514\ucf54\ub354(Decoder)\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc73c\uba70, \ub370\uc774\ud130\uc758 \uc7a0\uc7ac \uacf5\uac04(latent space)\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/p>\n
VAE\uc758 \ud6c8\ub828 \uacfc\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
GAN\uacfc VAE\ub294 \ubaa8\ub450 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \ubaa8\ub378\uc774\uc9c0\ub9cc, \uadf8 \uc811\uadfc \ubc29\uc2dd\uc5d0 \uc788\uc5b4 \uba87 \uac00\uc9c0 \uc911\uc694\ud55c \ucc28\uc774\uc810\uc774 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\uc774\uc81c \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. MNIST \ub370\uc774\ud130\uc14b\uc744 \ud1b5\ud574 \uc190\uae00\uc528 \uc22b\uc790 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \uc608\uc81c\ub97c \uc0b4\ud3b4\ubd05\ub2c8\ub2e4.<\/p>\n
pip install torch torchvision matplotlib<\/code><\/pre>\n5.2 \ub370\uc774\ud130\uc14b \ubd88\ub7ec\uc624\uae30<\/h3>\nimport torch\nfrom torchvision import datasets, transforms\n\ntransform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])\ntrain_data = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = torch.utils.data.DataLoader(train_data, batch_size=64, shuffle=True)<\/code><\/pre>\n5.3 GAN \ubaa8\ub378 \uc815\uc758<\/h3>\nclass Generator(torch.nn.Module):\n def __init__(self):\n super(Generator, self).__init__()\n self.model = torch.nn.Sequential(\n torch.nn.Linear(100, 256),\n torch.nn.ReLU(),\n torch.nn.Linear(256, 512),\n torch.nn.ReLU(),\n torch.nn.Linear(512, 1024),\n torch.nn.ReLU(),\n torch.nn.Linear(1024, 784),\n torch.nn.Tanh()\n )\n\n def forward(self, x):\n return self.model(x).view(-1, 1, 28, 28)\n\n class Discriminator(torch.nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.model = torch.nn.Sequential(\n torch.nn.Flatten(),\n torch.nn.Linear(784, 512),\n torch.nn.LeakyReLU(0.2),\n torch.nn.Linear(512, 256),\n torch.nn.LeakyReLU(0.2),\n torch.nn.Linear(256, 1),\n torch.nn.Sigmoid()\n )\n\n def forward(self, x):\n return self.model(x)<\/code><\/pre>\n5.4 \ubaa8\ub378 \ud6c8\ub828<\/h3>\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\ngenerator = Generator().to(device)\ndiscriminator = Discriminator().to(device)\n\ncriterion = torch.nn.BCELoss()\noptimizer_G = torch.optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))\noptimizer_D = torch.optim.Adam(discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999))\n\nnum_epochs = 200\nfor epoch in range(num_epochs):\n for i, (images, _) in enumerate(train_loader):\n images = images.to(device)\n batch_size = images.size(0)\n\n # \uc9c4\uc9dc \ubc0f \uac00\uc9dc \ub808\uc774\ube14 \uc0dd\uc131\n real_labels = torch.ones(batch_size, 1).to(device)\n fake_labels = torch.zeros(batch_size, 1).to(device)\n\n # \ud310\ubcc4\uae30 \ud6c8\ub828\n optimizer_D.zero_grad()\n outputs = discriminator(images)\n d_loss_real = criterion(outputs, real_labels)\n d_loss_real.backward()\n\n noise = torch.randn(batch_size, 100).to(device)\n fake_images = generator(noise)\n outputs = discriminator(fake_images.detach())\n d_loss_fake = criterion(outputs, fake_labels)\n d_loss_fake.backward()\n optimizer_D.step()\n\n # \uc0dd\uc131\uae30 \ud6c8\ub828\n optimizer_G.zero_grad()\n outputs = discriminator(fake_images)\n g_loss = criterion(outputs, real_labels)\n g_loss.backward()\n optimizer_G.step()\n\n print(f'Epoch [{epoch}\/{num_epochs}], d_loss: {d_loss_real.item() + d_loss_fake.item():.4f}, g_loss: {g_loss.item():.4f}')<\/code><\/pre>\n6. \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c VAE \uad6c\ud604<\/h2>\n
\uc774\uc81c VAE\ub97c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774\ubc88\uc5d0\ub3c4 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc190\uae00\uc528 \uc22b\uc790 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \uc608\uc81c\ub97c \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
6.1 VAE \ubaa8\ub378 \uc815\uc758<\/h3>\nclass VAE(torch.nn.Module):\n def __init__(self):\n super(VAE, self).__init__()\n self.encoder = torch.nn.Sequential(\n torch.nn.Flatten(),\n torch.nn.Linear(784, 400),\n torch.nn.ReLU()\n )\n\n self.fc_mu = torch.nn.Linear(400, 20)\n self.fc_var = torch.nn.Linear(400, 20)\n\n self.decoder = torch.nn.Sequential(\n torch.nn.Linear(20, 400),\n torch.nn.ReLU(),\n torch.nn.Linear(400, 784),\n torch.nn.Sigmoid()\n )\n\n def encode(self, x):\n h = self.encoder(x)\n return self.fc_mu(h), self.fc_var(h)\n\n def reparameterize(self, mu, logvar):\n std = torch.exp(0.5 * logvar)\n eps = torch.randn_like(std)\n return mu + eps * std\n\n def decode(self, z):\n return self.decoder(z)\n\n def forward(self, x):\n mu, logvar = self.encode(x)\n z = self.reparameterize(mu, logvar)\n recon_x = self.decode(z)\n return recon_x, mu, logvar<\/code><\/pre>\n6.2 VAE \uc190\uc2e4 \ud568\uc218<\/h3>\ndef vae_loss(recon_x, x, mu, logvar):\n BCE = torch.nn.functional.binary_cross_entropy(recon_x, x, reduction='sum')\n return BCE + 0.5 * torch.sum(torch.exp(logvar) + mu.pow(2) - 1 - logvar)<\/code><\/pre>\n6.3 VAE \ubaa8\ub378 \ud6c8\ub828<\/h3>\nvae = VAE().to(device)\noptimizer_VAE = torch.optim.Adam(vae.parameters(), lr=1e-3)\n\nnum_epochs = 100\nfor epoch in range(num_epochs):\n for i, (images, _) in enumerate(train_loader):\n images = images.to(device)\n\n optimizer_VAE.zero_grad()\n recon_images, mu, logvar = vae(images)\n loss = vae_loss(recon_images, images, mu, logvar)\n loss.backward()\n optimizer_VAE.step()\n\n print(f'Epoch [{epoch}\/{num_epochs}], Loss: {loss.item():.4f}')<\/code><\/pre>\n7. \uacb0\ub860<\/h2>\n
\ubcf8 \uae00\uc5d0\uc11c\ub294 GAN\uacfc VAE\uc758 \uac1c\ub150\uacfc \uad6c\uc870\uc5d0 \ub300\ud574 \uc0b4\ud3b4\ubcf4\uace0, \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud558\uc5ec \uc774\ub4e4 \ubaa8\ub378\uc744 \uad6c\ud604\ud574 \ubcf4\uc558\uc2b5\ub2c8\ub2e4. GAN\uc740 \uc0dd\uc131\uae30\uc640 \ud310\ubcc4\uae30\uac00 \uacbd\uc7c1\ud558\ub294 \uad6c\uc870\ub97c \ud1b5\ud574 \uc0ac\uc2e4\uc801\uc778 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \ub370 \uac15\ub825\ud55c \ubc18\uba74, VAE\ub294 \uc7a0\uc7ac \uacf5\uac04\uc744 \ud1b5\ud574 \ub370\uc774\ud130\ub97c \ubaa8\ub378\ub9c1\ud558\uace0 \uc0dd\uc131\ud558\ub294 \ub370 \uc6b0\uc218\ud55c \uc131\ub2a5\uc744 \ubcf4\uc785\ub2c8\ub2e4. \uc774 \ub450 \ubaa8\ub378\uc758 \ud2b9\uc9d5\uc744 \uc774\ud574\ud558\uace0 \ud65c\uc6a9\ud558\uba74 \ub2e4\uc591\ud55c \ub370\uc774\ud130 \uc0dd\uc131 \ubb38\uc81c\ub97c \ud574\uacb0\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
8. \ucc38\uace0 \ubb38\ud5cc<\/h2>\n