\n Generative Adversarial Networks (GANs)\uc740 \uc0dd\uc131 \ubaa8\ub378\ub9c1 \ubd84\uc57c\uc5d0\uc11c \ud601\uc2e0\uc801\uc778 \ubc1c\uc804\uc744 \uc774\ub8ec \ub525\ub7ec\ub2dd \uc544\ud0a4\ud14d\ucc98\uc785\ub2c8\ub2e4. GAN\uc740 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd, \uc989 \uc0dd\uc131\uc790(Generator)\uc640 \ud310\ubcc4\uc790(Discriminator) \uac04\uc758 \uacbd\uc7c1\uc744 \ud1b5\ud574 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \ub370 \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \ubaa8\ub378\uc740 \uc8fc\uc5b4\uc9c4 \ub370\uc774\ud130 \ubd84\ud3ec\ub97c \ud559\uc2b5\ud558\uc5ec \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \uac83\uc774 \uac00\ub2a5\ud55c\ub370, \uc774\ub7ec\ud55c \ud2b9\uc131\uc744 \ud65c\uc6a9\ud558\uc5ec \ub2e4\uc591\ud55c \uc751\uc6a9 \ubd84\uc57c\uc5d0 \uc801\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \ubb38\uc11c\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58(pytorch) \ud504\ub808\uc784\uc6cc\ud06c\ub97c \uc0ac\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud558\uace0, \ub79c\ub364\ud55c \ub864\uc544\uc6c3 \ub370\uc774\ud130 \uc218\uc9d1(Rollout Data Collection)\uc774\ub77c\ub294 \uac1c\ub150\uc5d0 \ub300\ud574 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n GAN\uc740 Ian Goodfellow\uc5d0 \uc758\ud574 \uc81c\uc548\ub41c \ubaa8\ub378\uc774\uba70, \ub450 \uac1c\uc758 \ub124\ud2b8\uc6cc\ud06c\uac00 \uc11c\ub85c \uacbd\uc7c1\ud558\ub294 \uad6c\uc870\ub85c \ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4.\n <\/p>\n
\n \uc0dd\uc131\uc790\ub294 \ud310\ubcc4\uc790\ub97c \uc18d\uc774\uae30 \uc704\ud574 \uc810\uc810 \ub354 \uc0ac\uc2e4\uc801\uc778 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub824\uace0 \ud558\uba70, \ud310\ubcc4\uc790\ub294 \uc0dd\uc131\uc790\ub97c \ub354 \uc798 \uc2dd\ubcc4\ud558\uae30 \uc704\ud574 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774 \ub450 \ub124\ud2b8\uc6cc\ud06c\ub294 \uac01\uac01\uc758 \uc190\uc2e4 \ud568\uc218\ub97c \ucd5c\uc18c\ud654\ud558\uba70 \uacbd\uc7c1\ud569\ub2c8\ub2e4. \uacb0\uad6d, \uc0dd\uc131\uc790\ub294 \ud310\ubcc4\uc790\uac00 \uad6c\ubcc4\ud558\uc9c0 \ubabb\ud560 \uc815\ub3c4\uc758 \uc0ac\uc2e4\uc801\uc778 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uac8c \ub429\ub2c8\ub2e4.\n <\/p>\n
\n GAN\uc758 \uc190\uc2e4 \ud568\uc218\ub294 \uc544\ub798\uc640 \uac19\uc774 \uc815\uc758\ub429\ub2c8\ub2e4.\n <\/p>\n
\n LD<\/sub> = - Ex~pdata<\/sub><\/sub>[log(D(x))] - Ez~pz<\/sub><\/sub>[log(1 - D(G(z)))] \n LG<\/sub> = - Ez~pz<\/sub><\/sub>[log(D(G(z)))]\n <\/code><\/pre>\n\n \uc5ec\uae30\uc11c D\ub294 \ud310\ubcc4\uc790, G\ub294 \uc0dd\uc131\uc790\uc785\ub2c8\ub2e4. D\ub294 \uc9c4\uc9dc \ub370\uc774\ud130\uc640 \uac00\uc9dc \ub370\uc774\ud130\ub97c \uad6c\ubcc4\ud558\uace0, G\ub294 D\ub97c \uc18d\uc774\uae30 \uc704\ud574 \ud559\uc2b5\ud569\ub2c8\ub2e4.\n <\/p>\n
2. GAN \uad6c\ud604\ud558\uae30<\/h2>\n2.1 \ud658\uacbd \uc124\uc815<\/h3>\n
\n \ubcf8 \uc608\uc81c\uc5d0\uc11c\ub294 PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud569\ub2c8\ub2e4. \uba3c\uc800 PyTorch\uc640 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud569\ub2c8\ub2e4.\n <\/p>\n
\n !pip install torch torchvision matplotlib\n <\/code><\/pre>\n2.2 \ub370\uc774\ud130 \uc900\ube44<\/h3>\n
\n MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec GAN\uc744 \ud559\uc2b5\ud558\uaca0\uc2b5\ub2c8\ub2e4. PyTorch\uc758 torchvision \ud328\ud0a4\uc9c0\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub370\uc774\ud130\ub97c \uc27d\uac8c \ub2e4\uc6b4\ub85c\ub4dc\ud558\uace0 \ubd88\ub7ec\uc62c \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n import torch\n import torchvision.transforms as transforms\n from torchvision import datasets\n\n # \ub370\uc774\ud130\uc14b \ub2e4\uc6b4\ub85c\ub4dc\n transform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])\n train_dataset = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\n\n train_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=64, shuffle=True)\n <\/code><\/pre>\n2.3 GAN \ubaa8\ub378 \uc815\uc758<\/h3>\n
\n \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790 \ub124\ud2b8\uc6cc\ud06c\ub97c \uc815\uc758\ud569\ub2c8\ub2e4. \uac01\uac01\uc758 \ub124\ud2b8\uc6cc\ud06c\ub294 PyTorch\uc758 nn.Module\uc744 \uc0c1\uc18d\ubc1b\uc544 \uad6c\ud604\ud569\ub2c8\ub2e4.\n <\/p>\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(True),\n nn.Linear(256, 512),\n nn.ReLU(True),\n nn.Linear(512, 1024),\n nn.ReLU(True),\n nn.Linear(1024, 784),\n nn.Tanh() # MNIST\ub294 -1\uc5d0\uc11c 1 \uc0ac\uc774\uc758 \uac12\uc744 \uac00\uc9d0\n )\n \n def forward(self, z):\n return self.model(z)\n\n class Discriminator(nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.model = nn.Sequential(\n nn.Linear(784, 1024),\n nn.LeakyReLU(0.2, inplace=True),\n nn.Linear(1024, 512),\n nn.LeakyReLU(0.2, inplace=True),\n nn.Linear(512, 256),\n nn.LeakyReLU(0.2, inplace=True),\n nn.Linear(256, 1),\n nn.Sigmoid() # \ucd5c\uc885 \ucd9c\ub825\uc740 0\uacfc 1 \uc0ac\uc774\n )\n\n def forward(self, x):\n return self.model(x)\n <\/code><\/pre>\n2.4 GAN \ud6c8\ub828 \ub8e8\ud504 \uc124\uc815<\/h3>\n
\n \ud6c8\ub828 \ub8e8\ud504\ub97c \uc124\uc815\ud558\uc5ec GAN\uc774 \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790\uac00 \uc11c\ub85c \uacbd\uc7c1\ud558\ub3c4\ub85d \ud569\ub2c8\ub2e4. GAN \ud6c8\ub828\uc740 \ubc18\ubcf5\uc801\uc774\uba70, \ub9e4 \ubc18\ubcf5\ub9c8\ub2e4 \ud310\ubcc4\uc790\ub294 \uc2e4\uc81c \ub370\uc774\ud130\uc640 \uc0dd\uc131\ub41c \ub370\uc774\ud130\ub97c \uad6c\ubcc4\ud558\ub294 \ubc95\uc744 \ud559\uc2b5\ud558\uace0, \uc0dd\uc131\uc790\ub294 \ud310\ubcc4\uc790\ub97c \uc18d\uc774\uae30 \uc704\ud574 \ub178\ub825\ud569\ub2c8\ub2e4.\n <\/p>\n
\n generator = Generator()\n discriminator = Discriminator()\n\n criterion = nn.BCELoss() # Binary Cross Entropy Loss\n lr = 0.0002\n num_epochs = 200\n g_optimizer = torch.optim.Adam(generator.parameters(), lr=lr)\n d_optimizer = torch.optim.Adam(discriminator.parameters(), lr=lr)\n\n for epoch in range(num_epochs):\n for i, (images, _) in enumerate(train_loader):\n # \uc2e4\uc81c \ub370\uc774\ud130 \ub77c\ubca8\n real_labels = torch.ones(images.size(0), 1)\n # \uac00\uc9dc \ub370\uc774\ud130 \ub77c\ubca8\n fake_labels = torch.zeros(images.size(0), 1)\n\n # \ud310\ubcc4\uc790 \ud559\uc2b5\n outputs = discriminator(images.view(-1, 784))\n d_loss_real = criterion(outputs, real_labels)\n real_score = outputs\n\n z = torch.randn(images.size(0), 100)\n fake_images = generator(z)\n outputs = discriminator(fake_images.detach())\n d_loss_fake = criterion(outputs, fake_labels)\n fake_score = outputs\n\n d_loss = d_loss_real + d_loss_fake\n d_optimizer.zero_grad()\n d_loss.backward()\n d_optimizer.step()\n\n # \uc0dd\uc131\uc790 \ud559\uc2b5\n outputs = discriminator(fake_images)\n g_loss = criterion(outputs, real_labels)\n\n g_optimizer.zero_grad()\n g_loss.backward()\n g_optimizer.step()\n \n if (epoch+1) % 10 == 0:\n print(f'Epoch [{epoch+1}\/{num_epochs}], d_loss: {d_loss.item():.4f}, g_loss: {g_loss.item():.4f}, '\n f'D(x): {real_score.mean().item():.2f}, D(G(z)): {fake_score.mean().item():.2f}')\n <\/code><\/pre>\n2.5 \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0 \uc2dc\uac01\ud654<\/h3>\n
\n GAN \ud6c8\ub828\uc774 \uc644\ub8cc\ub41c \ud6c4 \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \uc2dc\uac01\ud654\ud558\uc5ec \uc131\ub2a5\uc744 \ud3c9\uac00\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n import matplotlib.pyplot as plt\n\n z = torch.randn(64, 100)\n generated_images = generator(z).view(-1, 1, 28, 28).detach().numpy()\n\n plt.figure(figsize=(8, 8))\n for i in range(64):\n plt.subplot(8, 8, i+1)\n plt.imshow(generated_images[i][0], cmap='gray')\n plt.axis('off')\n plt.show()\n <\/code><\/pre>\n3. \ub79c\ub364\ud55c \ub864\uc544\uc6c3 \ub370\uc774\ud130 \uc218\uc9d1<\/h2>\n
\n GAN\uc744 \ud65c\uc6a9\ud558\uc5ec \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub294 \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \ub370 \uc720\uc6a9\ud569\ub2c8\ub2e4. \uadf8\ub7ec\ub098 \uc774\ub7ec\ud55c \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud558\ub294 \uac83\uc740 \ud2b9\uc815 \ud658\uacbd\uc774\ub098 \uc815\ucc45\uc5d0 \uc758\uc874\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.
\n \ub864\uc544\uc6c3 \ub370\uc774\ud130 \uc218\uc9d1(rollout data collection)<\/strong>\uc740 \uc8fc\uc5b4\uc9c4 \uc815\ucc45 \ub610\ub294 \ud658\uacbd\uc5d0\uc11c \uc0dd\uc131\ub41c \ub370\uc774\ud130\ub97c \uc218\uc9d1\ud558\ub294 \uacfc\uc815\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4. \uba38\uc2e0\ub7ec\ub2dd \ubc0f \uac15\ud654\ud559\uc2b5\uc5d0\uc11c \uc911\uc694\ud55c \uac1c\ub150\uc73c\ub85c, \ub370\uc774\ud130\uc758 \ub2e4\uc591\uc131\uc744 \ud655\ubcf4\ud558\uace0 \ud559\uc2b5 \uc131\ub2a5\uc744 \ub192\uc774\uae30 \uc704\ud574 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.\n <\/p>\n\n \uc608\ub97c \ub4e4\uc5b4, \uac15\ud654\ud559\uc2b5 \uc5d0\uc774\uc804\ud2b8\ub97c \ud6c8\ub828\ud560 \ub54c, \ub864\uc544\uc6c3 \ub370\uc774\ud130 \uc218\uc9d1\uc744 \ud1b5\ud574 \uc5d0\uc774\uc804\ud2b8\uac00 \ub2e4\uc591\ud55c \uc0c1\ud669\uc744 \uacbd\ud5d8\ud558\ub3c4\ub85d \ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4. \uc774\ub294 \uc5d0\uc774\uc804\ud2b8\uac00 \ub2e4\uc591\ud55c \uc0c1\ud0dc\uc640 \ud589\ub3d9 \uc30d\uc744 \ud559\uc2b5\ud558\uac8c \ud558\uc5ec \ub354 \uc77c\ubc18\ud654\ub41c \uc815\ucc45\uc744 \uc0dd\uc131\ud558\ub294 \ub370 \ub3c4\uc6c0\uc744 \uc90d\ub2c8\ub2e4.\n <\/p>\n
3.1 \ub864\uc544\uc6c3 \ub370\uc774\ud130 \uc218\uc9d1\uc744 \uc704\ud55c \ud658\uacbd \uad6c\ud604<\/h3>\n
\n OpenAI\uc758 Gym\uacfc \uac19\uc740 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \ud1b5\ud574 \uac15\ud654\ud559\uc2b5 \ud658\uacbd\uc744 \uc27d\uac8c \uad6c\ucd95\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\ub294 \uac04\ub2e8\ud55c \ub864\uc544\uc6c3 \uc218\uc9d1\uc758 \uc608\uc785\ub2c8\ub2e4.\n <\/p>\n
\n import gym\n\n env = gym.make('CartPole-v1')\n\n def collect_rollouts(env, num_rollouts=5):\n rollouts = []\n for _ in range(num_rollouts):\n state = env.reset()\n done = False\n rollout = []\n while not done:\n action = env.action_space.sample() # \ub79c\ub364\ud55c \ud589\ub3d9 \uc120\ud0dd\n next_state, reward, done, _ = env.step(action)\n rollout.append((state, action, reward, next_state))\n state = next_state\n rollouts.append(rollout)\n \n return rollouts\n\n rollouts = collect_rollouts(env, num_rollouts=10)\n print(rollouts)\n <\/code><\/pre>\n3.2 \uc218\uc9d1\ub41c \ub370\uc774\ud130 \ud65c\uc6a9<\/h3>\n
\n \uc218\uc9d1\ub41c \ub864\uc544\uc6c3 \ub370\uc774\ud130\ub294 GAN\uc758 \ud559\uc2b5\uc5d0 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc218\uc9d1\ub41c \ub370\uc774\ud130\ub97c \ud1b5\ud574 \ubaa8\ub378\uc740 \ub354\uc6b1 \ub2e4\uc591\ud55c \uc0c1\ud669\uc5d0\uc11c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.
\n GAN \ubaa8\ub378\uc758 \uc785\ub825\uc73c\ub85c \uc0ac\uc6a9\ud558\uc5ec \ub2e4\uc591\ud55c \uc0c1\ud0dc\ub97c \ub79c\ub364\uc73c\ub85c \uc0dd\uc131\ud558\uac70\ub098, \ud2b9\uc815 \uc0c1\ud0dc\uc5d0 \ub300\ud55c \uc801\uc808\ud55c \ud589\ub3d9\uc744 \ud559\uc2b5\ud558\ub294 \ub370 \uc0ac\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\uacb0\ub860<\/h2>\n
\n \ubcf8 \uae00\uc5d0\uc11c\ub294 GAN\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \uad6c\ud604 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc124\uba85\ud558\uc600\uc73c\uba70, \uac15\ud654\ud559\uc2b5\uc5d0\uc11c\uc758 \ub79c\ub364\ud55c \ub864\uc544\uc6c3 \ub370\uc774\ud130 \uc218\uc9d1\uc758 \uc911\uc694\uc131\uc5d0 \ub300\ud574 \ub2e4\ub8e8\uc5c8\uc2b5\ub2c8\ub2e4. \ub610\ud55c, PyTorch\ub97c \ud65c\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud558\ub294 \uc608\uc81c\ub97c \ud1b5\ud574 \uc2e4\uc81c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4.
\n \uc774\ub7ec\ud55c \uae30\ubc95\ub4e4\uc740 \ub2e4\uc591\ud55c \uba38\uc2e0\ub7ec\ub2dd \ubc0f \ub525\ub7ec\ub2dd \uc751\uc6a9 \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub420 \uc218 \uc788\uc73c\uba70, \ud544\uc694\ud55c \uc0c1\ud669\uc5d0 \ub9de\ub294 \ub370\uc774\ud130\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \uc0dd\uc131\ud558\uace0 \ud65c\uc6a9\ud558\ub294 \ub370 \uae30\uc5ec\ud560 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n
\ucc38\uace0\ubb38\ud5cc<\/h2>\n\n- Goodfellow, Ian, et al. “Generative Adversarial Nets.” Advances in Neural Information Processing Systems, 2014.<\/li>\n
- Pytorch Documentation: https:\/\/pytorch.org\/docs\/stable\/index.html<\/a><\/li>\n
- OpenAI Gym Documentation: https:\/\/gym.openai.com\/<\/a><\/li>\n<\/ul>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"