Generative Adversarial Networks (GANs)\uc740 \ucd5c\uadfc \uba87 \ub144 \uc0ac\uc774\uc5d0 \uc774\ubbf8\uc9c0\uc640 \ube44\ub514\uc624 \ub4f1 \ub2e4\uc591\ud55c \uc0dd\uc131 \uc791\uc5c5\uc5d0\uc11c \ud070 \uc8fc\ubaa9\uc744 \ubc1b\uc544\uc654\uc2b5\ub2c8\ub2e4. GAN\uc740 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd, \uc989 \uc0dd\uc131\uc790(generator)\uc640 \uc2dd\ubcc4\uc790(discriminator)\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc73c\uba70, \uc774 \ub450 \ubaa8\ub378\uc740 \uc11c\ub85c \uacbd\uc7c1\ud558\uba70 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 GAN\uc758 \uae30\ubcf8 \uac1c\ub150\uc744 \uc18c\uac1c\ud558\uace0, MuseGAN\uc774\ub77c\ub294 \ud2b9\uc815 GAN \uc544\ud0a4\ud14d\ucc98\ub97c \uc0b4\ud3b4\ubcf8 \ud6c4, \ud30c\uc774\ud1a0\uce58(PyTorch)\ub97c \ud65c\uc6a9\ud55c \uac04\ub2e8\ud55c \uc608\uc81c\ub97c \ud1b5\ud574 GAN\uc744 \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
GAN\uc740 Ian Goodfellow\uac00 2014\ub144\uc5d0 \uc81c\uc548\ud55c \uc54c\uace0\ub9ac\uc998\uc73c\ub85c, \uc8fc\ub85c \uc774\ubbf8\uc9c0 \uc0dd\uc131, \uc774\ubbf8\uc9c0 \ubcc0\ud658, \uc2a4\ud0c0\uc77c \uc804\uc774 \ub4f1 \ub2e4\uc591\ud55c \ubb38\uc81c\uac00 \uc788\uc2b5\ub2c8\ub2e4. GAN\uc758 \ud575\uc2ec\uc740 \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd\uc774 \uc11c\ub85c\ub97c “\uacf5\uaca9”\ud558\ub294 \uad6c\uc870\uc785\ub2c8\ub2e4.<\/p>\n
\uc774 \ub450 \ub124\ud2b8\uc6cc\ud06c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uc190\uc2e4 \ud568\uc218\ub97c \ud1b5\ud574 \ud559\uc2b5\ud569\ub2c8\ub2e4.<\/p>\n
\n\ub9cc\uc6d0\uc758\uc2dd\ubcc4\uc790\uc2dd\ubcc4\uc790)\uc5d0 \ub300\ud55c \uc190\uc2e4? = log(D(x)) + log(1 - D(G(z)))\n<\/pre>\n\uc5ec\uae30\uc11c
D(x)<\/code>\ub294 \uc2e4\uc81c \ub370\uc774\ud130 x\uc5d0 \ub300\ud55c \uc2dd\ubcc4\uc790\uc758 \ud655\ub960,G(z)<\/code>\ub294 \uc0dd\uc131\uc790\uc5d0 \uc758\ud574 \uc0dd\uc131\ub41c \ub370\uc774\ud130,D(G(z))<\/code>\ub294 \uc0dd\uc131\uc790 \ucd9c\ub825\uc5d0 \ub300\ud55c \uc2dd\ubcc4\uc790\uc758 \ud655\ub960\uc785\ub2c8\ub2e4.<\/p>\n2. MuseGAN \uc774\ud574\ud558\uae30<\/h2>\n
MuseGAN\uc740 \uc74c\uc545 \uc0dd\uc131 \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 GAN \uc544\ud0a4\ud14d\ucc98\ub97c \ud655\uc7a5\ud55c \uac83\uc785\ub2c8\ub2e4. MuseGAN\uc740 \uc74c\uc131\uacfc \uc545\uae30 \ud3ec\ud568\ud558\uc5ec \ub2e4\ucc44\ub85c\uc6b4 \uc74c\uc545 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 \ud2b9\ud788 MIDI \ud615\uc2dd\uc758 \uc74c\uc545 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \uac15\uc810\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
2.1 MuseGAN \uc544\ud0a4\ud14d\ucc98<\/h3>\n
MuseGAN\uc740 \uc77c\ubc18 GAN\uc758 \uad6c\uc870\ub97c \uae30\ubc18\uc73c\ub85c \ud558\uba74\uc11c\ub3c4 \ub2e4\uc74c\uacfc \uac19\uc740 \uc694\uc18c\ub97c \ud3ec\ud568\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\n
- \uc8fc\uc694 \uc0dd\uc131\uae30\n
- \ub2e4\ub2e8\uacc4 \uac10\uc9c0\uae30<\/strong>: \uc0dd\uc131\ub41c \uc74c\uc545\uc758 \uc5ec\ub7ec \uce21\uba74\uc744 \ud3c9\uac00\ud558\uae30 \uc704\ud55c \uc5ec\ub7ec \ub124\ud2b8\uc6cc\ud06c\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/li>\n
<\/strong><\/li>\n<\/ul>\n
2.2 MuseGAN\uc758 \ub370\uc774\ud130\uc14b<\/h3>\n
MuseGAN\uc744 \ud559\uc2b5\uc2dc\ud0a4\uae30 \uc704\ud574\uc11c\ub294 MIDI \ud615\uc2dd\uc758 \ub370\uc774\ud130\uc14b\uc774 \ud544\uc694\ud569\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc73c\ub85c Lakh MIDI Dataset\uacfc \uac19\uc740 \ub370\uc774\ud130\uc14b\uc774 \ud65c\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n
3. PyTorch\ub85c GAN \uad6c\ud604\ud558\uae30<\/h2>\n
\uc774\uc81c GAN\uc758 \uae30\ubcf8 \uac1c\ub150\uc744 \uc774\ud574\ud588\uc73c\ub2c8, PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c GAN\uc744 \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
3.1 \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58<\/h3>\n
\uc6b0\uc120 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud569\ub2c8\ub2e4. PyTorch\uc640 \uad00\ub828\ub41c \ubaa8\ub4c8\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc5b4\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
pip install torch torchvision matplotlib<\/code><\/pre>\n3.2 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\uc5ec\uae30\uc11c\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud558\uac8c \uad6c\ud604\ud558\uaca0\uc2b5\ub2c8\ub2e4. MNIST\ub294 \uc190\uc73c\ub85c \uc4f4 \uc22b\uc790 \uc774\ubbf8\uc9c0 \ub370\uc774\ud130\uc14b\uc785\ub2c8\ub2e4.<\/p>\n
\nimport torch\nfrom torchvision import datasets, transforms\n\n# MNIST \ub370\uc774\ud130\uc14b \ub85c\ub4dc\ntransform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])\ndataset = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ndataloader = torch.utils.data.DataLoader(dataset, batch_size=64, shuffle=True)\n<\/code><\/pre>\n3.3 \uc0dd\uc131\uc790 \ubc0f \uc2dd\ubcc4\uc790 \ubaa8\ub378 \uc815\uc758<\/h3>\n
\ub2e4\uc74c\uc73c\ub85c \uc0dd\uc131\uc790\uc640 \uc2dd\ubcc4\uc790 \ubaa8\ub378\uc744 \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\n
\nimport torch.nn as nn\n\nclass 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(),\n nn.Linear(256, 512),\n nn.ReLU(),\n nn.Linear(512, 1024),\n nn.ReLU(),\n nn.Linear(1024, 28 * 28),\n nn.Tanh()\n )\n\n def forward(self, z):\n return self.model(z).view(-1, 1, 28, 28)\n\nclass Discriminator(nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.model = nn.Sequential(\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()\n )\n\n def forward(self, img):\n return self.model(img.view(-1, 28 * 28))\n<\/code><\/pre>\n3.4 \uc190\uc2e4 \ud568\uc218 \ubc0f \uc635\ud2f0\ub9c8\uc774\uc800 \uc124\uc815<\/h3>\n
GAN\uc758 \ud559\uc2b5\uc5d0 \uc0ac\uc6a9\ub420 \uc190\uc2e4 \ud568\uc218\ub294 Binary Cross Entropy\uc774\uba70, \uc635\ud2f0\ub9c8\uc774\uc800\ub85c\ub294 Adam\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n
\nimport torch.optim as optim\n\n# \ubaa8\ub378 \ucd08\uae30\ud654\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>\n3.5 GAN \ud559\uc2b5<\/h3>\n
\uc774\uc81c GAN\uc744 \ud559\uc2b5\ud560 \uc900\ube44\uac00 \ub418\uc5c8\uc2b5\ub2c8\ub2e4. \ud559\uc2b5 \uacfc\uc815\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\nimport numpy as np\nimport matplotlib.pyplot as plt\n\nnum_epochs = 50\nsample_interval = 1000\nz_dim = 100\nbatch_size = 64\n\nfor epoch in range(num_epochs):\n for i, (imgs, _) in enumerate(dataloader):\n # \uc9c4\uc9dc \uc774\ubbf8\uc9c0\uc640 \uac00\uc9dc \uc774\ubbf8\uc9c0\uc758 \ub77c\ubca8 \uc900\ube44\n real_labels = torch.ones(imgs.size(0), 1)\n fake_labels = torch.zeros(imgs.size(0), 1)\n \n # Discriminator \ud559\uc2b5\n optimizer_d.zero_grad()\n \n outputs = discriminator(imgs)\n d_loss_real = criterion(outputs, real_labels)\n d_loss_real.backward()\n \n z = torch.randn(imgs.size(0), z_dim)\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 optimizer_d.step()\n \n # Generator \ud559\uc2b5\n optimizer_g.zero_grad()\n \n outputs = discriminator(fake_images)\n g_loss = criterion(outputs, real_labels)\n g_loss.backward()\n optimizer_g.step()\n \n if i % sample_interval == 0:\n print(f'Epoch [{epoch}\/{num_epochs}] Batch [{i}\/{len(dataloader)}] \\\n Loss D: {d_loss_real.item() + d_loss_fake.item()}, Loss G: {g_loss.item()}')\n<\/code><\/pre>\n3.6 \uacb0\uacfc \uc2dc\uac01\ud654<\/h3>\n
\ud559\uc2b5\uc774 \ub05d\ub09c \ud6c4\uc5d0\ub294 \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0\ub97c \uc2dc\uac01\ud654\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\n# \uc0dd\uc131\ub41c \uc774\ubbf8\uc9c0 \uc2dc\uac01\ud654\nz = torch.randn(100, z_dim)\ngenerated_images = generator(z)\n\n# \uc774\ubbf8\uc9c0 \ucd9c\ub825\ngrid_img = make_grid(generated_images, nrow=10, normalize=True)\nplt.imshow(grid_img.permute(1, 2, 0).detach().numpy())\nplt.axis('off')\nplt.show()\n<\/code><\/pre>\n4. MuseGAN \uad6c\ud604<\/h2>\n
MuseGAN\uc758 \uc804\uccb4\uc801\uc778 \uad6c\uc870\uc640 \ub370\uc774\ud130 \ucc98\ub9ac \ud6c4, \uc2e4\uc81c\ub85c MuseGAN\uc744 \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \ud2b9\uc815\ud55c \uad6c\ud604 \uc138\ubd80 \uc0ac\ud56d\uc740 \ub2e4\ub97c \uc218 \uc788\uc9c0\ub9cc, MuseGAN\uc758 \uc8fc\uc694 \uad6c\uc131\uc694\uc18c\ub4e4\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
4.1 MuseGAN \uc544\ud0a4\ud14d\ucc98 \uc124\uacc4<\/h3>\n
MuseGAN\uc758 \ub370\uc774\ud130\ub294 MIDI \ud30c\uc77c \ud615\uc2dd\uc774\uba70, \uc774\ub97c \ucc98\ub9ac\ud558\uae30 \uc704\ud574 MIDI \ub370\uc774\ud130 \ub85c\ub354\uc640 \ub2e4\uc591\ud55c \ub808\uc774\uc5b4 \uad6c\uc870\ub97c \uc124\uacc4\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
4.2 MIDI \ub370\uc774\ud130 \ub85c\ub529<\/h3>\n
\nimport pretty_midi\n\ndef load_midi(file_path):\n midi_data = pretty_midi.PrettyMIDI(file_path)\n # MIDI \ub370\uc774\ud130 \ucc98\ub9ac \ub85c\uc9c1 \uad6c\ud604\n return midi_data\n<\/code><\/pre>\n4.3 MuseGAN\uc758 \ud6c8\ub828 \ub8e8\ud504<\/h3>\n
\ubba4\uc9c0\uceec \uc0dd\uc131\uae30\uc758 \ud6c8\ub828\uc740 GAN\uc758 \uc6d0\ub9ac\uc640 \uc720\uc0ac\ud558\uba70, \uc798 \uc815\uc758\ub41c \uc190\uc2e4 \ud568\uc218\uc640 \ucd5c\uc801\ud654 \uacfc\uc815\uc774 \ud544\uc694\ud569\ub2c8\ub2e4.<\/p>\n
\n# MuseGAN \ud6c8\ub828 \ub8e8\ud504 \uc608\uc2dc\nfor epoch in range(num_epochs):\n for midi_input in midi_dataset:\n # \ubaa8\ub378 \ud6c8\ub828 \ub85c\uc9c1 \uad6c\ud604\n pass\n<\/code><\/pre>\n4.4 \uacb0\uacfc \uc0dd\uc131 \ubc0f \ud3c9\uac00<\/h3>\n
\ud6c8\ub828\uc774 \ub05d\ub09c \ud6c4, MuseGAN\uc73c\ub85c \uc0dd\uc131\ub41c MIDI \ud30c\uc77c\uc744 \ud655\uc778\ud558\uace0 \ud3c9\uac00\ud574 \ubd05\ub2c8\ub2e4. \ud3c9\uac00\ub97c \ud1b5\ud574 \ubaa8\ub378\uc744 \uac1c\uc120\ud560 \uc218 \uc788\ub294 \ud53c\ub4dc\ubc31\uc744 \ubc1b\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
5. \uacb0\ub860<\/h2>\n
\ubcf8 \uae00\uc5d0\uc11c\ub294 GAN\uc758 \uae30\ucd08\ubd80\ud130 \uc2dc\uc791\ud558\uc5ec MuseGAN\uc758 \uad6c\uc870\uc640 \uc791\ub3d9 \uc6d0\ub9ac\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \ub610\ud55c PyTorch\ub97c \ud1b5\ud574 \uac04\ub2e8\ud55c GAN \uad6c\ud604\uc744 \uc2dc\ub3c4\ud558\uace0, \uc2e4\uc9c8\uc801\uc73c\ub85c \uc74c\uc545 \uc0dd\uc131 \ubb38\uc81c\uc5d0 \uc811\uadfc\ud558\ub294 \ubc29\ubc95\uc744 \uc18c\uac1c\ud588\uc2b5\ub2c8\ub2e4. GAN\uc758 \ubc1c\uc804\uacfc \uadf8 \uc751\uc6a9 \ubd84\uc57c\ub294 \uc55e\uc73c\ub85c\ub3c4 \uacc4\uc18d\ud574\uc11c \ud655\uc7a5\ub420 \uac83\uc73c\ub85c \uae30\ub300\ub429\ub2c8\ub2e4.<\/p>\n
\ud53c\ub4dc\ubc31\uc774\ub098 \uc9c8\ubb38\uc774 \uc788\ub2e4\uba74 \uc5b8\uc81c\ub4e0\uc9c0 \ub313\uae00\ub85c \ub0a8\uaca8\uc8fc\uc138\uc694!<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"