Generative Adversarial Networks (GANs)\ub294 2014\ub144 Ian Goodfellow\uc5d0 \uc758\ud574 \uc81c\uc548\ub41c \ubaa8\ub378\ub85c, \ub450 \uc2e0\uacbd\ub9dd \uac04\uc758 \uacbd\uc7c1\uc744 \ud1b5\ud574 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \ubaa8\ub378\uc785\ub2c8\ub2e4. GAN\uc740 \ud2b9\ud788 \uc774\ubbf8\uc9c0 \uc0dd\uc131, \uc2a4\ud0c0\uc77c \ubcc0\ud658, \ub370\uc774\ud130 \uc99d\uac15 \ub4f1\uc5d0 \ub110\ub9ac \uc0ac\uc6a9\ub418\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ubc88 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 GAN\uc758 \uae30\ubcf8 \uad6c\uc870, \ud30c\uc774\ud1a0\uce58\ub97c \uc774\uc6a9\ud55c \uad6c\ud604 \ubc29\ubc95, \uac15\ud654\ud559\uc2b5\uc758 \uae30\ubcf8 \uac1c\ub150 \ubc0f \uc5ec\ub7ec \uc751\uc6a9 \uc0ac\ub840\ub97c \uc18c\uac1c\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
GAN\uc740 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc73c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4: \uc0dd\uc131\uc790(Generator)\uc640 \ud310\ubcc4\uc790(Discriminator)\uc785\ub2c8\ub2e4. \uc0dd\uc131\uc790\ub294 \ubb34\uc791\uc704 \ub178\uc774\uc988\ub97c \uc785\ub825\uc73c\ub85c \ubc1b\uc544 \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \ud310\ubcc4\uc790\ub294 \uc785\ub825 \ub370\uc774\ud130\uac00 \uc9c4\uc9dc \ub370\uc774\ud130\uc778\uc9c0 \uc0dd\uc131\ub41c \ub370\uc774\ud130\uc778\uc9c0\ub97c \uad6c\ubcc4\ud569\ub2c8\ub2e4. \uc774 \ub450 \ub124\ud2b8\uc6cc\ud06c\ub294 \uc11c\ub85c \uacbd\uc7c1\ud558\uba74\uc11c \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/p>\n
\uc0dd\uc131\uc790\ub294 \ub178\uc774\uc988 \ubca1\ud130\ub97c \ubc1b\uc544 \uc9c4\uc9dc\ucc98\ub7fc \ubcf4\uc774\ub294 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4. \ubaa9\ud45c\ub294 \ud310\ubcc4\uc790\ub97c \uc18d\uc774\ub294 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
\ud310\ubcc4\uc790\ub294 \uc785\ub825 \ub370\uc774\ud130\uc758 \uc9c4\uc704 \uc5ec\ubd80\ub97c \ud310\ub2e8\ud569\ub2c8\ub2e4. \uc9c4\uc9dc \ub370\uc774\ud130\uc77c \uacbd\uc6b0 1, \uc0dd\uc131\ub41c \ub370\uc774\ud130\uc77c \uacbd\uc6b0 0\uc744 \ucd9c\ub825\ud569\ub2c8\ub2e4.<\/p>\n
GAN\uc758 \uc190\uc2e4 \ud568\uc218\ub294 \ub2e4\uc74c\uacfc \uac19\uc774 \uc124\uc815\ub429\ub2c8\ub2e4:<\/p>\n
min_G max_D V(D, G) = E[log(D(x))] + E[log(1 - D(G(z)))]<\/code><\/pre>\n\uc5ec\uae30\uc11c E<\/code>\ub294 \uae30\ub300\uac12\uc744 \ub098\ud0c0\ub0b4\uba70, x<\/code>\ub294 \uc9c4\uc9dc \ub370\uc774\ud130, G(z)<\/code>\ub294 \uc0dd\uc131\uc790\uac00 \uc0dd\uc131\ud55c \ub370\uc774\ud130\uc785\ub2c8\ub2e4. \uc0dd\uc131\uc790\ub294 \uc190\uc2e4\uc744 \ucd5c\uc18c\ud654\ud558\ub824\uace0 \ud558\uace0, \ud310\ubcc4\uc790\ub294 \uc190\uc2e4\uc744 \ucd5c\ub300\ud654\ud558\ub824 \ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n3. \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c GAN \uad6c\ud604<\/h2>\n
\uc774\uc81c GAN\uc744 \ud30c\uc774\ud1a0\uce58\ub85c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \ub370\uc774\ud130\uc14b\uc73c\ub85c\ub294 MNIST \uc190\uae00\uc528 \uc22b\uc790 \ub370\uc774\ud130\ub97c \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
3.1 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\nimport torch\nimport torchvision\nfrom torchvision import datasets, transforms\n\n# \ub370\uc774\ud130 \ubcc0\ud658 \ubc0f \ub2e4\uc6b4\ub85c\ub4dc\ntransform = transforms.Compose([\n transforms.ToTensor(),\n transforms.Normalize((0.5,), (0.5,))\n])\n\n# MNIST \ub370\uc774\ud130\uc14b\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)<\/code><\/pre>\n3.2 \uc0dd\uc131\uc790 (Generator) \ubaa8\ub378 \uc815\uc758<\/h3>\nimport torch.nn as nn\n\nclass Generator(nn.Module):\n def __init__(self):\n super(Generator, self).__init__()\n self.layer1 = nn.Sequential(\n nn.Linear(100, 256),\n nn.ReLU(True)\n )\n self.layer2 = nn.Sequential(\n nn.Linear(256, 512),\n nn.ReLU(True)\n )\n self.layer3 = nn.Sequential(\n nn.Linear(512, 1024),\n nn.ReLU(True)\n )\n self.layer4 = nn.Sequential(\n nn.Linear(1024, 28*28),\n nn.Tanh() # \ud53d\uc140 \uac12\uc740 -1\uacfc 1 \uc0ac\uc774\n )\n \n def forward(self, z):\n out = self.layer1(z)\n out = self.layer2(out)\n out = self.layer3(out)\n out = self.layer4(out)\n return out.view(-1, 1, 28, 28) # \uc774\ubbf8\uc9c0 \ud615\ud0dc\ub85c \ubcc0\ud658<\/code><\/pre>\n3.3 \ud310\ubcc4\uc790 (Discriminator) \ubaa8\ub378 \uc815\uc758<\/h3>\nclass Discriminator(nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.layer1 = nn.Sequential(\n nn.Linear(28*28, 1024),\n nn.LeakyReLU(0.2, inplace=True)\n )\n self.layer2 = nn.Sequential(\n nn.Linear(1024, 512),\n nn.LeakyReLU(0.2, inplace=True)\n )\n self.layer3 = nn.Sequential(\n nn.Linear(512, 256),\n nn.LeakyReLU(0.2, inplace=True)\n )\n self.layer4 = nn.Sequential(\n nn.Linear(256, 1),\n nn.Sigmoid() # \ucd9c\ub825\uac12\uc744 0\uacfc 1 \uc0ac\uc774\ub85c\n )\n \n def forward(self, x):\n out = self.layer1(x.view(-1, 28*28)) # \ud3c9\ud0c4\ud654\n out = self.layer2(out)\n out = self.layer3(out)\n out = self.layer4(out)\n return out<\/code><\/pre>\n3.4 \ubaa8\ub378 \ud6c8\ub828<\/h3>\nimport torch.optim as optim\n\n# \ubaa8\ub378 \ucd08\uae30\ud654\ngenerator = Generator()\ndiscriminator = Discriminator()\n\n# \uc190\uc2e4 \ud568\uc218 \ubc0f \ucd5c\uc801\ud654\uae30 \uc124\uc815\ncriterion = nn.BCELoss() # Binary Cross Entropy Loss\noptimizer_g = optim.Adam(generator.parameters(), lr=0.0002)\noptimizer_d = optim.Adam(discriminator.parameters(), lr=0.0002)\n\n# \ud6c8\ub828\nnum_epochs = 200\nfor epoch in range(num_epochs):\n for i, (images, _) in enumerate(train_loader):\n # \uc9c4\uc9dc \ub370\uc774\ud130 \ub808\uc774\ube14\n real_labels = torch.ones(images.size(0), 1)\n fake_labels = torch.zeros(images.size(0), 1)\n\n # \ud310\ubcc4\uc790 \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 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 d_loss_fake.backward()\n \n optimizer_d.step()\n \n # \uc0dd\uc131\uc790 \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 if (epoch+1) % 10 == 0:\n print(f'Epoch [{epoch+1}\/{num_epochs}], d_loss: {d_loss_real.item() + d_loss_fake.item():.4f}, g_loss: {g_loss.item():.4f}')<\/code><\/pre>\n3.5 \uacb0\uacfc \uc2dc\uac01\ud654<\/h3>\nimport matplotlib.pyplot as plt\n\n# \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0 \uc2dc\uac01\ud654 \ud568\uc218\ndef plot_generated_images(generator, n=10):\n z = torch.randn(n, 100)\n with torch.no_grad():\n generated_images = generator(z).cpu()\n generated_images = generated_images.view(-1, 28, 28)\n \n plt.figure(figsize=(10, 1))\n for i in range(n):\n plt.subplot(1, n, i+1)\n plt.imshow(generated_images[i], cmap='gray')\n plt.axis('off')\n plt.show()\n\n# \uc774\ubbf8\uc9c0 \uc0dd\uc131\nplot_generated_images(generator)<\/code><\/pre>\n4. \uac15\ud654\ud559\uc2b5\uc758 \uae30\ubcf8 \uac1c\ub150<\/h2>\n
\uac15\ud654\ud559\uc2b5(Reinforcement Learning, RL)\uc740 \uc5d0\uc774\uc804\ud2b8\uac00 \ud658\uacbd\uacfc \uc0c1\ud638\uc791\uc6a9\ud558\uba70 \ucd5c\uc801\uc758 \ud589\ub3d9\uc744 \ud559\uc2b5\ud558\ub294 \uae30\uacc4 \ud559\uc2b5\uc758 \ud55c \ubd84\uc57c\uc785\ub2c8\ub2e4. \uc5d0\uc774\uc804\ud2b8\ub294 \uc0c1\ud0dc\ub97c \uad00\ucc30\ud558\uace0, \ud589\ub3d9\uc744 \uc120\ud0dd\ud558\uace0, \ubcf4\uc0c1\uc744 \ubc1b\uc73c\uba70, \uc774\ub97c \ud1b5\ud574 \ucd5c\uc801\uc758 \uc815\ucc45\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/p>\n
4.1 \uac15\ud654\ud559\uc2b5\uc758 \uad6c\uc131 \uc694\uc18c<\/h3>\n\n- \uc0c1\ud0dc (State):<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \ud604\uc7ac\uc758 \ud658\uacbd\uc744 \ub098\ud0c0\ub0b4\ub294 \uc815\ubcf4\uc785\ub2c8\ub2e4.<\/li>\n
- \ud589\ub3d9 (Action):<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \ud604\uc7ac \uc0c1\ud0dc\uc5d0\uc11c \uc218\ud589\ud560 \uc218 \uc788\ub294 \uc791\uc5c5\uc785\ub2c8\ub2e4.<\/li>\n
- \ubcf4\uc0c1 (Reward):<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \ud589\ub3d9\uc744 \uc218\ud589\ud55c \ud6c4\uc5d0 \ud658\uacbd\uc73c\ub85c\ubd80\ud130 \ubc1b\ub294 \ud53c\ub4dc\ubc31\uc785\ub2c8\ub2e4.<\/li>\n
- \uc815\ucc45 (Policy):<\/strong> \uc5d0\uc774\uc804\ud2b8\uac00 \uac01 \uc0c1\ud0dc\uc5d0\uc11c \ucde8\ud560 \ud589\ub3d9\uc758 \ud655\ub960 \ubd84\ud3ec\ub97c \ub098\ud0c0\ub0c5\ub2c8\ub2e4.<\/li>\n<\/ul>\n
4.2 \uac15\ud654\ud559\uc2b5 \uc54c\uace0\ub9ac\uc998<\/h3>\n\n- Q-Learning:<\/strong> \uac00\uce58 \uae30\ubc18 \ubc29\ubc95\uc73c\ub85c, Q \uac12\uc744 \ud559\uc2b5\ud558\uc5ec \ucd5c\uc801\uc758 \uc815\ucc45\uc744 \uc720\ub3c4\ud569\ub2c8\ub2e4.<\/li>\n
- \uc815\ucc45 \uacbd\uc0ac \ubc29\ubc95 (Policy Gradient):<\/strong> \uc815\ucc45\uc744 \uc9c1\uc811 \ud559\uc2b5\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/li>\n
- Actor-Critic:<\/strong> \uac00\uce58 \ud568\uc218\uc640 \uc815\ucc45\uc744 \ub3d9\uc2dc\uc5d0 \ud559\uc2b5\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/li>\n<\/ul>\n
4.3 \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c \uac15\ud654\ud559\uc2b5 \uad6c\ud604<\/h3>\n
\uac04\ub2e8\ud55c \uac15\ud654\ud559\uc2b5 \uad6c\ud604\uc744 \uc704\ud574 OpenAI\uc758 Gym \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 CartPole \ud658\uacbd\uc744 \ub2e4\ub8e8\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
4.3.1 Gym \ud658\uacbd \uc124\uc815<\/h4>\nimport gym\n\n# Gym \ud658\uacbd \uc0dd\uc131\nenv = gym.make('CartPole-v1') # CartPole \ud658\uacbd<\/code><\/pre>\n4.3.2 DQN \ubaa8\ub378 \uc815\uc758<\/h4>\nclass DQN(nn.Module):\n def __init__(self, input_size, num_actions):\n super(DQN, self).__init__()\n self.fc1 = nn.Linear(input_size, 24)\n self.fc2 = nn.Linear(24, 24)\n self.fc3 = nn.Linear(24, num_actions)\n\n def forward(self, x):\n x = F.relu(self.fc1(x))\n x = F.relu(self.fc2(x))\n return self.fc3(x)<\/code><\/pre>\n4.3.3 \ubaa8\ub378 \ud6c8\ub828<\/h4>\ndef train_dqn(env, num_episodes):\n model = DQN(input_size=env.observation_space.shape[0], num_actions=env.action_space.n)\n optimizer = optim.Adam(model.parameters())\n criterion = nn.MSELoss()\n\n for episode in range(num_episodes):\n state = env.reset()\n state = torch.FloatTensor(state)\n done = False\n total_reward = 0\n\n while not done:\n q_values = model(state)\n action = torch.argmax(q_values).item() # \ub610\ub294 epsilon-greedy \uc815\ucc45 \uc0ac\uc6a9\n\n next_state, reward, done, _ = env.step(action)\n next_state = torch.FloatTensor(next_state)\n\n total_reward += reward\n\n # DQN \uc5c5\ub370\uc774\ud2b8 \ub85c\uc9c1 \ucd94\uac00 \ud544\uc694\n\n state = next_state\n\n print(f'Episode {episode+1}, Total Reward: {total_reward}') \n\n return model\n\n# DQN \ud6c8\ub828 \uc2dc\uc791\ntrain_dqn(env, num_episodes=1000)<\/code><\/pre>\n5. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \ud3ec\uc2a4\ud305\uc5d0\uc11c\ub294 GAN\uacfc \uac15\ud654\ud559\uc2b5\uc758 \uae30\ubcf8 \uac1c\ub150 \ubc0f \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud55c \uad6c\ud604 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. GAN\uc740 \ub370\uc774\ud130 \uc0dd\uc131\uc5d0 \ub9e4\uc6b0 \uc720\uc6a9\ud55c \ubaa8\ub378\uc774\uace0, \uac15\ud654\ud559\uc2b5\uc740 \uc5d0\uc774\uc804\ud2b8\uac00 \ucd5c\uc801\uc758 \uc815\ucc45\uc744 \ud559\uc2b5\ud558\ub3c4\ub85d \ub3d5\ub294 \uae30\ubc95\uc785\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uae30\uc220\ub4e4\uc740 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \uc751\uc6a9\ub420 \uc218 \uc788\uc73c\uba70, \uc55e\uc73c\ub85c\uc758 \uc5f0\uad6c\uc640 \ubc1c\uc804\uc774 \uae30\ub300\ub429\ub2c8\ub2e4.<\/p>\n
6. \ucc38\uace0 \uc790\ub8cc<\/h2>\n