\ub525\ub7ec\ub2dd\uc740 \uae30\uacc4 \ud559\uc2b5\uc758 \ud55c \ubd84\uc57c\ub85c, \uc2e0\uacbd\ub9dd\uc744 \ud65c\uc6a9\ud558\uc5ec \ub370\uc774\ud130\ub85c\ubd80\ud130 \ud328\ud134\uc744 \ud559\uc2b5\ud558\ub294 \uae30\ubc95\uc785\ub2c8\ub2e4. \uc624\ub298 \uc774 \uae00\uc5d0\uc11c\ub294 \ubcc0\ud615 \uc624\ud1a0\uc778\ucf54\ub354(Variational Autoencoder, VAE)\uc5d0 \ub300\ud574 \uae4a\uc774 \uc788\uac8c \ub2e4\ub8e8\uc5b4 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc624\ud1a0\uc778\ucf54\ub354\ub294 \ube44\uc9c0\ub3c4 \ud559\uc2b5 \ubc29\ubc95\uc73c\ub85c \uc77c\ubc18\uc801\uc73c\ub85c \uc785\ub825 \ub370\uc774\ud130\ub97c \uc555\ucd95\ud55c \ud6c4 \ubcf5\uc6d0\ud558\ub294 \uacfc\uc815\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc624\ud1a0\uc778\ucf54\ub354\ub294 \uc778\ucf54\ub354\uc640 \ub514\ucf54\ub354 \ub450 \ubd80\ubd84\uc73c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4.<\/p>\n
\n \uc624\ud1a0\uc778\ucf54\ub354\uc758 \ud6c8\ub828 \uacfc\uc815\uc740 \uc785\ub825 \ub370\uc774\ud130\uc640 \ucd9c\ub825 \ub370\uc774\ud130\uc758 \ucc28\uc774\ub97c \uc904\uc774\ub294 \ubc29\ud5a5\uc73c\ub85c \uc9c4\ud589\ub429\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \uc190\uc2e4 \ud568\uc218(loss function)\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc2e4\uc81c \ucd9c\ub825\uacfc \uc608\uce21 \ucd9c\ub825 \uac04\uc758 \ucc28\uc774\ub97c \uce21\uc815\ud569\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc73c\ub85c Mean Squared Error (MSE) \uc190\uc2e4 \ud568\uc218\uac00 \ub9ce\uc774 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.\n <\/p>\n
\ubcc0\ud615 \uc624\ud1a0\uc778\ucf54\ub354\ub294 \uae30\uc874\uc758 \uc624\ud1a0\uc778\ucf54\ub354\ub97c \ud655\uc7a5\ud55c \ubaa8\ub378\ub85c, \uc785\ub825 \ub370\uc774\ud130\uc758 \ud655\ub960 \ubd84\ud3ec\ub97c \ucd94\uc815\ud569\ub2c8\ub2e4. VAE\ub294 \uc0dd\uc131 \ubaa8\ub378\ub85c\uc11c, \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\ub294 \ub2a5\ub825\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
VAE\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ub450 \uac00\uc9c0 \uc8fc\uc694 \uc694\uc18c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4:<\/p>\n
\n VAE\uc758 \uc190\uc2e4 \ud568\uc218\ub294 \ub450 \ubd80\ubd84\uc73c\ub85c \ub098\ub20c \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\nL = E[log p(x|z)] - D_{KL}(q(z|x) || p(z))\n <\/pre>\n\uc5ec\uae30\uc11c:<\/p>\n
\uc774\uc81c \ubcc0\ud615 \uc624\ud1a0\uc778\ucf54\ub354\uc758 \uae30\ubcf8 \uad6c\uc131 \uc694\uc18c\uc640 \uc190\uc2e4 \ud568\uc218\ub97c \uc774\ud574\ud588\uc73c\ubbc0\ub85c, \ud30c\uc774\ud1a0\uce58\ub85c VAE\ub97c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\npip install torch torchvision matplotlib\n <\/pre>\n3.2 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc190\uae00\uc528 \uc22b\uc790\ub97c \uc778\uc2dd\ud558\ub294 VAE\ub97c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. MNIST\ub294 28×28 \ud53d\uc140\uc758 \ud751\ubc31 \uc774\ubbf8\uc9c0\ub85c \uad6c\uc131\ub41c \ub370\uc774\ud130\uc14b\uc785\ub2c8\ub2e4.<\/p>\n
\nimport torch\nfrom torchvision import datasets, transforms\n\ntransform = transforms.Compose([\n transforms.ToTensor(),\n transforms.Lambda(lambda x: x.view(-1))\n])\n\nmnist_train = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = torch.utils.data.DataLoader(mnist_train, batch_size=128, shuffle=True)\n <\/pre>\n3.3 \ubaa8\ub378 \uc815\uc758<\/h3>\n
\ubcc0\ud615 \uc624\ud1a0\uc778\ucf54\ub354 \ubaa8\ub378\uc744 \uad6c\uc131\ud558\uae30 \uc704\ud574, \uc778\ucf54\ub354\uc640 \ub514\ucf54\ub354 \ud074\ub798\uc2a4 \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\n
\nimport torch.nn as nn\n\nclass Encoder(nn.Module):\n def __init__(self, input_dim, latent_dim):\n super(Encoder, self).__init__()\n self.fc1 = nn.Linear(input_dim, 400)\n self.fc21 = nn.Linear(400, latent_dim) # \ud3c9\uade0\n self.fc22 = nn.Linear(400, latent_dim) # \ub85c\uadf8 \ubd84\uc0b0\n \n def forward(self, x):\n h1 = torch.relu(self.fc1(x))\n mu = self.fc21(h1)\n logvar = self.fc22(h1)\n return mu, logvar\n\n\nclass Decoder(nn.Module):\n def __init__(self, latent_dim, output_dim):\n super(Decoder, self).__init__()\n self.fc1 = nn.Linear(latent_dim, 400)\n self.fc2 = nn.Linear(400, output_dim)\n \n def forward(self, z):\n h2 = torch.relu(self.fc1(z))\n return torch.sigmoid(self.fc2(h2))\n \nclass VAE(nn.Module):\n def __init__(self, input_dim, latent_dim):\n super(VAE, self).__init__()\n self.encoder = Encoder(input_dim, latent_dim)\n self.decoder = Decoder(latent_dim, input_dim)\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 forward(self, x):\n mu, logvar = self.encoder(x)\n z = self.reparameterize(mu, logvar)\n return self.decoder(z), mu, logvar\n <\/pre>\n3.4 \uc190\uc2e4 \ud568\uc218 \uc815\uc758<\/h3>\n
VAE\uc5d0\uc11c \uc190\uc2e4 \ud568\uc218\ub97c \uc815\uc758\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 pytorch\uc758 \uae30\ub2a5\uc744 \uc774\uc6a9\ud574 \uad6c\ud604\ud569\ub2c8\ub2e4.<\/p>\n
\ndef vae_loss(recon_x, x, mu, logvar):\n BCE = nn.functional.binary_cross_entropy(recon_x, x, reduction='sum')\n KLD = -0.5 * torch.sum(1 + logvar - mu.pow(2) - logvar.exp())\n return BCE + KLD\n <\/pre>\n3.5 \ubaa8\ub378 \ud6c8\ub828<\/h3>\n
\ud6c8\ub828 \ub8e8\ud504\ub97c \uc0ac\uc6a9\ud574 \ubaa8\ub378\uc744 \ud6c8\ub828\ud569\ub2c8\ub2e4. \uc190\uc2e4 \ud568\uc218\ub97c \ub2e4\uc2dc \uacc4\uc0b0\ud558\uace0, \uc5ed\uc804\ud30c\ub97c \uc218\ud589\ud558\uc5ec \uac00\uc911\uce58\ub97c \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/p>\n
\nimport torch.optim as optim\n\ndevice = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nmodel = VAE(784, 20).to(device)\noptimizer = optim.Adam(model.parameters(), lr=1e-3)\n\nmodel.train()\nfor epoch in range(10):\n train_loss = 0\n for batch_idx, (data, _) in enumerate(train_loader):\n data = data.to(device)\n optimizer.zero_grad()\n recon_batch, mu, logvar = model(data)\n loss = vae_loss(recon_batch, data, mu, logvar)\n loss.backward()\n train_loss += loss.item()\n optimizer.step()\n \n print(f'Epoch {epoch+1}, Loss: {train_loss \/ len(train_loader.dataset)}')\n <\/pre>\n3.6 \uacb0\uacfc \ud655\uc778<\/h3>\n
\ud6c8\ub828\uc774 \uc644\ub8cc\ub418\uba74, \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \ud6c8\ub828 \ub370\uc774\ud130\uc640 \uc5bc\ub9c8\ub098 \uc720\uc0ac\ud55c\uc9c0 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\nimport matplotlib.pyplot as plt\n\ndef visualize_results(model, num_images=10):\n with torch.no_grad():\n z = torch.randn(num_images, 20).to(device)\n sample = model.decoder(z).cpu()\n sample = sample.view(num_images, 1, 28, 28)\n \n plt.figure(figsize=(10, 1))\n for i in range(num_images):\n plt.subplot(1, num_images, i + 1)\n plt.imshow(sample[i].squeeze(), cmap='gray')\n plt.axis('off')\n plt.show()\n\nvisualize_results(model)\n <\/pre>\n4. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ubcc0\ud615 \uc624\ud1a0\uc778\ucf54\ub354\uc758 \uac1c\ub150\uacfc \ud30c\uc774\ud1a0\uce58\ub85c \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. VAE\ub294 \ub370\uc774\ud130\uc758 \uc7a0\uc7ac\uc801\uc778 \ubd84\ud3ec\ub97c \ud559\uc2b5\ud558\uace0 \uc0c8\ub85c\uc6b4 \uc0d8\ud50c\uc744 \uc0dd\uc131\ud560 \uc218 \uc788\ub294 \uae30\ub2a5\uc774 \uc788\uc5b4 \ub2e4\uc591\ud55c \uc0dd\uc131 \ubaa8\ub378\ub9c1\uc5d0 \ud65c\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uae30\uc220\uc744 \ud65c\uc6a9\ud558\uc5ec \uc774\ubbf8\uc9c0, \ud14d\uc2a4\ud2b8 \ubc0f \uc624\ub514\uc624 \ub370\uc774\ud130 \uc0dd\uc131\uacfc \uac19\uc740 \uc5ec\ub7ec \uac00\uc9c0 \ud765\ubbf8\ub85c\uc6b4 \uc791\uc5c5\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\ub354 \ub098\uc544\uac00, VAE\ub294 GAN\uacfc \uac19\uc740 \ub2e4\ub978 \uc0dd\uc131 \ubaa8\ub378\uacfc \uacb0\ud569\ud558\uc5ec \ub354\uc6b1 \uac15\ub825\ud558\uace0 \ub2e4\uc591\ud55c \uc0dd\uc131 \ubaa8\ub378\uc744 \uad6c\ud604\ud558\ub294 \ub370 \uae30\uc5ec\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, VAE\ub294 \uace0\ucc28\uc6d0 \ub370\uc774\ud130\uc758 \uc7a0\uc7ac \uacf5\uac04\uc744 \ud0d0\uc0c9\ud558\uace0 \uc0d8\ud50c\ub9c1\ud560 \uc218 \uc788\ub3c4\ub85d \ub3c4\uc640\uc90d\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0 \ubb38\ud5cc<\/h2>\n
\n
- Kingma, D. P., & Welling, M. (2014). Auto-Encoding Variational Bayes. arXiv:1312.6114<\/i><\/li>\n
- PyTorch Documentation: https:\/\/pytorch.org\/docs\/stable\/index.html<\/a><\/li>\n<\/ul>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"