Generative Adversarial Networks (GAN)\uc740 Ian Goodfellow\uc640 \uadf8\uc758 \ub3d9\ub8cc\ub4e4\uc774 2014\ub144\uc5d0 \uc18c\uac1c\ud55c \ud601\uc2e0\uc801\uc778 \ub525\ub7ec\ub2dd \uae30\ubc95\uc785\ub2c8\ub2e4. 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. \uc774 \ub450 \ub124\ud2b8\uc6cc\ud06c\ub294 \uc11c\ub85c \uacbd\uc7c1\ud558\uba70 \ud559\uc2b5\ud558\uc5ec \uace0\ud488\uc9c8\uc758 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub294 \ub370 \ubaa9\uc801\uc744 \ub450\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 GAN\uc758 \ub3d9\uc791 \uc6d0\ub9ac, \uad6c\uc131 \uc694\uc18c, \ud6c8\ub828 \uacfc\uc815 \ubc0f PyTorch\ub97c \uc0ac\uc6a9\ud55c \uad6c\ud604 \uc608\uc81c\ub97c \uc790\uc138\ud788 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
GAN\uc740 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd, \uc989 \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790 \uac04\uc758 \uacbd\uc7c1 \uad6c\uc870\ub85c \uc124\uc815\ub429\ub2c8\ub2e4. \uc774 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc774 \uc791\ub3d9\ud569\ub2c8\ub2e4:<\/p>\n
\uc774 \ub450 \ub124\ud2b8\uc6cc\ud06c\ub294 \ub3d9\uc2dc\uc5d0 \ud6c8\ub828\ub418\uba70, \uc0dd\uc131\uc790\ub294 \ud310\ubcc4\uc790\ub97c \uc18d\uc774\ub294 \uac00\uc9dc \ub370\uc774\ud130\ub97c \ub9cc\ub4e4\uae30 \uc704\ud574 \uac1c\uc120\ub418\uace0, \ud310\ubcc4\uc790\ub294 \uac00\uc9dc\uc640 \uc2e4\uc81c \ub370\uc774\ud130\ub97c \uc798 \uad6c\ubd84\ud558\uae30 \uc704\ud574 \uac1c\uc120\ub429\ub2c8\ub2e4.<\/p>\n
GAN\uc758 \ubaa9\ud45c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ube44\uc6a9 \ud568\uc218\ub97c \ucd5c\uc18c\ud654\ud558\ub294 \uac83\uc785\ub2c8\ub2e4:<\/p>\n
\nD\\*(x) = log(D(x)) + log(1 - D(G(z)))\n <\/code><\/pre>\n\uc5ec\uae30\uc11c,<\/p>\n
\nD(x):<\/code> \uc2e4\uc81c \ub370\uc774\ud130 x\uc5d0 \ub300\ud55c \ud310\ubcc4\uc790\uc758 \ucd9c\ub825\uc785\ub2c8\ub2e4. (1\uc5d0 \uac00\uae4c\uc6b0\uba74 \uc2e4\uc81c \ub370\uc774\ud130, 0\uc5d0 \uac00\uae4c\uc6b0\uba74 \uac00\uc9dc \ub370\uc774\ud130)<\/li>\nG(z):<\/code> \uc0dd\uc131\uc790\uac00 \ub79c\ub364 \ub178\uc774\uc988 z\ub97c \ud1b5\ud574 \uc0dd\uc131\ud55c \ub370\uc774\ud130\uc785\ub2c8\ub2e4.<\/li>\nD(G(z)):<\/code> \ud310\ubcc4\uc790\uac00 \uc0dd\uc131\ub41c \ub370\uc774\ud130\uc5d0 \ub300\ud574 \ubc18\ud658\ud55c \ud655\ub960\uc785\ub2c8\ub2e4.<\/li>\n<\/ul>\n\ubaa9\ud45c\ub294 \ud310\ubcc4\uc790\uac00 \uc2e4\uc81c \ub370\uc774\ud130\uc5d0 \ub300\ud574 1, \uc0dd\uc131\ub41c \ub370\uc774\ud130\uc5d0 \ub300\ud574 0\uc744 \ucd9c\ub825\ud558\ub3c4\ub85d \ud558\ub294 \uac83\uc785\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uc0dd\uc131\uc790\ub294 \uc810\uc810 \ub354 \uc9c4\uc9dc\uc640 \uc720\uc0ac\ud55c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
3. GAN\uc758 \uad6c\uc131 \uc694\uc18c<\/h2>\n3.1 \uc0dd\uc131\uc790(Generator)<\/h3>\n
\uc0dd\uc131\uc790\ub294 \uc77c\ubc18\uc801\uc73c\ub85c \uc644\uc804 \uc5f0\uacb0(fully connected) \ub808\uc774\uc5b4 \ub610\ub294 \ucee8\ubcfc\ub8e8\uc158 \ub808\uc774\uc5b4\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4. \uc785\ub825\uc73c\ub85c \ub79c\ub364 \ubca1\ud130 z\ub97c \ubc1b\uace0, \uc774\ub97c \ud1b5\ud574 \uc2e4\uc81c \ub370\uc774\ud130\uc640 \uc720\uc0ac\ud55c \uc815\ubcf4\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/p>\n
3.2 \ud310\ubcc4\uc790(Discriminator)<\/h3>\n
\ud310\ubcc4\uc790\ub294 \uc785\ub825 \ub370\uc774\ud130(\uc2e4\uc81c \ub610\ub294 \uc0dd\uc131\ub41c)\ub97c \ubc1b\uc544\uc11c \uc774\ub97c \uc9c4\uc9dc\uc778\uc9c0 \uac00\uc9dc\uc778\uc9c0 \ud310\ub2e8\ud569\ub2c8\ub2e4. \uc774 \ub610\ud55c \uc644\uc804 \uc5f0\uacb0 \ub610\ub294 \ucee8\ubcfc\ub8e8\uc158 \ub124\ud2b8\uc6cc\ud06c\ub85c \uc124\uacc4\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
4. GAN\uc758 \ud6c8\ub828 \uacfc\uc815<\/h2>\n
GAN\uc758 \ud6c8\ub828\uc740 \ub2e4\uc74c \ub2e8\uacc4\ub85c \uc774\ub8e8\uc5b4\uc9d1\ub2c8\ub2e4:<\/p>\n
\n- \uc2e4\uc81c \ub370\uc774\ud130\ub97c \uc120\ud0dd\ud558\uace0, \ub79c\ub364 \ub178\uc774\uc988 \ubca1\ud130 z\ub97c \uc0d8\ud50c\ub9c1\ud569\ub2c8\ub2e4.<\/li>\n
- \uc0dd\uc131\uc790\uac00 \ub178\uc774\uc988 z\ub97c \uc785\ub825\ubc1b\uc544 \uac00\uc9dc \ub370\uc774\ud130\ub97c \ub9cc\ub4ed\ub2c8\ub2e4.<\/li>\n
- \ud310\ubcc4\uc790\ub294 \uc2e4\uc81c \ub370\uc774\ud130\uc640 \uc0dd\uc131\uc790\uac00 \ub9cc\ub4e0 \ub370\uc774\ud130\ub97c \ud3c9\uac00\ud569\ub2c8\ub2e4.<\/li>\n
- \ud310\ubcc4\uc790\uc758 \uc190\uc2e4\uc744 \uacc4\uc0b0\ud558\uace0, \uc5ed\uc804\ud30c\ud558\uc5ec \ud310\ubcc4\uc790\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/li>\n
- \uc0dd\uc131\uc790\uc758 \uc190\uc2e4\uc744 \uacc4\uc0b0\ud558\uace0, \uc5ed\uc804\ud30c\ud558\uc5ec \uc0dd\uc131\uc790\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/li>\n<\/ol>\n
\uc774 \uacfc\uc815\uc744 \ubc18\ubcf5\ud558\uba70 \ub450 \ub124\ud2b8\uc6cc\ud06c \ubaa8\ub450 \uac1c\uc120\ud569\ub2c8\ub2e4.<\/p>\n
5. GAN\uc758 PyTorch \uad6c\ud604 \uc608\uc81c<\/h2>\n
\ub2e4\uc74c\uc740 PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec GAN\uc744 \uad6c\ud604\ud558\ub294 \uac04\ub2e8\ud55c \uc608\uc81c\uc785\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc22b\uc790 \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud558\ub294 \ubaa8\ub378\uc744 \ub9cc\ub4e4\uc5b4 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
5.1 \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58 \ubc0f \ub370\uc774\ud130\uc14b \ub85c\ub529<\/h3>\n\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torchvision import datasets, transforms\nfrom torch.utils.data import DataLoader\n <\/code><\/pre>\n\uba3c\uc800, \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc784\ud3ec\ud2b8\ud558\uace0 MNIST \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud569\ub2c8\ub2e4.<\/p>\n
\n# MNIST \ub370\uc774\ud130\uc14b \ub2e4\uc6b4\ub85c\ub4dc \ubc0f \ub85c\ub4dc\ntransform = transforms.Compose([\n transforms.Resize(28),\n transforms.ToTensor(),\n])\n\nmnist = datasets.MNIST(root='.\/data', train=True, transform=transform, download=True)\ndataloader = DataLoader(mnist, batch_size=64, shuffle=True)\n <\/code><\/pre>\n5.2 \uc0dd\uc131\uc790(Generator) \ubaa8\ub378 \uc815\uc758<\/h3>\n
\uc0dd\uc131\uc790 \ubaa8\ub378\uc740 \ub79c\ub364 \ub178\uc774\uc988\ub97c \uc785\ub825\uc73c\ub85c \ubc1b\uc544 \uc2e4\uc81c\uc640 \uc720\uc0ac\ud55c \uc774\ubbf8\uc9c0\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/p>\n
\nclass Generator(nn.Module):\n def __init__(self):\n super(Generator, self).__init__()\n self.model = nn.Sequential(\n nn.Linear(100, 128),\n nn.ReLU(),\n nn.Linear(128, 256),\n nn.ReLU(),\n nn.Linear(256, 512),\n nn.ReLU(),\n nn.Linear(512, 28*28), # MNIST \uc774\ubbf8\uc9c0 \ud06c\uae30\n nn.Tanh() # \ud53d\uc140 \uac12 \ubc94\uc704\ub97c [-1, 1]\ub85c \uc870\uc815\n )\n\n def forward(self, z):\n return self.model(z).view(-1, 1, 28, 28)\n <\/code><\/pre>\n5.3 \ud310\ubcc4\uc790(Discriminator) \ubaa8\ub378 \uc815\uc758<\/h3>\n
\ud310\ubcc4\uc790 \ubaa8\ub378\uc740 \uc785\ub825 \uc774\ubbf8\uc9c0\ub97c \ubc1b\uc544 \uadf8\uac83\uc774 \uc2e4\uc81c\uc778\uc9c0 \uc0dd\uc131\ub41c \uac83\uc778\uc9c0 \ud310\ubcc4\ud569\ub2c8\ub2e4.<\/p>\n
\nclass Discriminator(nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.model = nn.Sequential(\n nn.Flatten(), # \uc774\ubbf8\uc9c0 \ud615\ud0dc\ub97c \uc77c\ucc28\uc6d0\uc73c\ub85c \ubcc0\ud658\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() # \ucd9c\ub825 \ud655\ub960\n )\n\n def forward(self, x):\n return self.model(x)\n <\/code><\/pre>\n5.4 \uc190\uc2e4 \ud568\uc218\uc640 \uc635\ud2f0\ub9c8\uc774\uc800 \uc815\uc758<\/h3>\n\n# \uc0dd\uc131\uc790\uc640 \ud310\ubcc4\uc790 \uc0dd\uc131\ngenerator = Generator()\ndiscriminator = Discriminator()\n\n# \uc190\uc2e4 \ud568\uc218\ncriterion = nn.BCELoss()\n\n# \uc635\ud2f0\ub9c8\uc774\uc800\noptimizer_g = optim.Adam(generator.parameters(), lr=0.0002, betas=(0.5, 0.999))\noptimizer_d = optim.Adam(discriminator.parameters(), lr=0.0002, betas=(0.5, 0.999))\n <\/code><\/pre>\n5.5 GAN \ud6c8\ub828 \ub2e8\uacc4<\/h3>\n
\uc774\uc81c GAN\uc744 \ud6c8\ub828\ud558\ub294 \ub8e8\ud504\ub97c \uc815\uc758\ud569\ub2c8\ub2e4. \uac01 \uc5d0\ud3ec\ud06c\ub9c8\ub2e4 \ud310\ubcc4\uc790\uc640 \uc0dd\uc131\uc790\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/p>\n
\nnum_epochs = 50\n\nfor epoch in range(num_epochs):\n for real_images, _ in dataloader:\n batch_size = real_images.size(0)\n\n # \uc9c4\uc9dc \uc774\ubbf8\uc9c0\uc5d0 \ub300\ud55c \ub808\uc774\ube14\n real_labels = torch.ones(batch_size, 1)\n # \uac00\uc9dc \uc774\ubbf8\uc9c0\uc5d0 \ub300\ud55c \ub808\uc774\ube14\n fake_labels = torch.zeros(batch_size, 1)\n\n # \ud310\ubcc4\uc790 \ud6c8\ub828\n discriminator.zero_grad()\n outputs = discriminator(real_images)\n d_loss_real = criterion(outputs, real_labels)\n d_loss_real.backward()\n\n # \uac00\uc9dc \ub370\uc774\ud130 \uc0dd\uc131\n noise = torch.randn(batch_size, 100)\n fake_images = generator(noise)\n\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 generator.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+1}\/{num_epochs}], d_loss: {d_loss_real.item() + d_loss_fake.item()}, g_loss: {g_loss.item()}')\n <\/code><\/pre>\n6. GAN\uc758 \ud65c\uc6a9 \ubd84\uc57c<\/h2>\n
GAN\uc740 \uc5ec\ub7ec \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc77c\ubd80 \uc608\uc2dc\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n