\n\uc791\uc131\uc790: \ub2f9\uc2e0\uc758 \uc774\ub984<\/p>\n
\uc791\uc131\uc77c: 2023\ub144 10\uc6d4 1\uc77c<\/p>\n<\/header>\n\n1. GAN(Generative Adversarial Networks) \uc18c\uac1c<\/h2>\n
Generative Adversarial Networks(GAN)\uc740 Ian Goodfellow\uac00 2014\ub144\uc5d0 \uc81c\uc548\ud55c \uae30\uacc4 \ud559\uc2b5 \ubaa8\ub378\ub85c, \ub450 \uac1c\uc758 \uc2e0\uacbd\ub9dd \ubaa8\ub378\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4: \uc0dd\uc131\uae30(generator)\uc640 \ud310\ubcc4\uae30(discriminator). \uc0dd\uc131\uae30\ub294 \ud6c8\ub828 \ub370\uc774\ud130\ub97c \uae30\ubc18\uc73c\ub85c \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0, \ud310\ubcc4\uae30\ub294 \uc8fc\uc5b4\uc9c4 \ub370\uc774\ud130\uac00 \uc2e4\uc81c \ub370\uc774\ud130\uc778\uc9c0 \uc0dd\uc131\ub41c \ub370\uc774\ud130\uc778\uc9c0\ub97c \ud310\ubcc4\ud558\ub294 \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4. \uc774 \ub450 \uac00\uc9c0 \ub124\ud2b8\uc6cc\ud06c\ub294 \uc11c\ub85c \uacbd\uc7c1\uc744 \ud558\uba74\uc11c \ub3d9\uc2dc\uc5d0 \ud559\uc2b5\ud558\uac8c \ub429\ub2c8\ub2e4.<\/p>\n
GAN\uc758 \uae30\ubcf8 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \uc0dd\uc131\uae30: \ub79c\ub364\ud55c \ub178\uc774\uc988 \ubca1\ud130\ub97c \ubc1b\uc544\uc11c, \uc774\ub97c \uae30\ubc18\uc73c\ub85c \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\ub97c \uc0dd\uc131.<\/li>\n
- \ud310\ubcc4\uae30: \uc2e4\uc81c \ub370\uc774\ud130\uc640 \uc0dd\uc131\ub41c \ub370\uc774\ud130\ub97c \uc785\ub825\ubc1b\uc544, \uadf8\uac83\uc774 \uc9c4\uc9dc\uc778\uc9c0 \uac00\uc9dc\uc778\uc9c0 \uad6c\ubcc4.<\/li>\n<\/ul>\n
\uc774\ub7ec\ud55c \uacbd\uc7c1 \uad6c\uc870\ub294 \uc0dd\uc131\uae30\uac00 \uc810\uc810 \ub354 \uc2e4\uc81c \ub370\uc774\ud130\uc640 \uc720\uc0ac\ud55c \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\ub3c4\ub85d \uc720\ub3c4\ud558\uba70, \uacb0\uad6d \ub9e4\uc6b0 \ud604\uc2e4\uc801\uc778 \ub370\uc774\ud130 \uc0dd\uc131\uc774 \uac00\ub2a5\ud574\uc9d1\ub2c8\ub2e4.<\/p>\n<\/section>\n\n2. MuseGAN \uc18c\uac1c<\/h2>\n
MuseGAN\uc740 \uc74c\uc545 \uc0dd\uc131\uc5d0 \ud2b9\ud654\ub41c GAN\uc758 \ud55c \uc608\uc785\ub2c8\ub2e4. MuseGAN\uc740 \uc8fc\ub85c MIDI \ud30c\uc77c\uc744 \uae30\ubc18\uc73c\ub85c \ud55c \uc74c\uc545 \uc0dd\uc131 \ubaa8\ub378\ub85c, \ub2e4\uc591\ud55c \uc74c\uc545 \uc694\uc18c\ub4e4\uc744 \ud30c\uc545\ud558\uace0 \ud559\uc2b5\ud568\uc73c\ub85c\uc368 \uc0c8\ub85c\uc6b4 \uc74c\uc545\uc744 \ucc3d\uc791\ud560 \uc218 \uc788\ub3c4\ub85d \uc124\uacc4\ub418\uc5c8\uc2b5\ub2c8\ub2e4. MuseGAN\uc740 \ud2b9\ud788 \ub2e4\uc131(multi-track) \uc74c\uc545\uc744 \uc0dd\uc131\ud558\ub294 \ub370 \uac15\uc810\uc744 \ubcf4\uc774\uba70, \uc0dd\uc131\ub41c \uc74c\uc545\uc758 \uac01 \ud2b8\ub799\uc774 \uc11c\ub85c \uc870\ud654\ub86d\uac8c \uc5f0\uc8fc\ub428\uc744 \ubaa9\ud45c\ub85c \ud569\ub2c8\ub2e4.<\/p>\n
MuseGAN\uc758 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n- \ub178\uc774\uc988 \uc785\ub825: \ub79c\ub364\ud55c \ub178\uc774\uc988 \ubca1\ud130.<\/li>\n
- \ud2b8\ub799 \uc0dd\uc131\uae30: \uc5ec\ub7ec \ud2b8\ub799(\uc608: \ub4dc\ub7fc, \ubca0\uc774\uc2a4, \uba5c\ub85c\ub514)\uc744 \uc0dd\uc131.<\/li>\n
- \uc0c1\ud669(Context) \ud2b9\uc131: \ud2b8\ub799 \uac04\uc758 \uc0c1\uad00\uad00\uacc4\ub97c \ud559\uc2b5\ud558\uc5ec \uc790\uc5f0\uc2a4\ub7ec\uc6b4 \uc74c\uc545\uc744 \uc0dd\uc131.<\/li>\n<\/ul>\n
\uc774\ub7ec\ud55c \uc694\uc18c\ub4e4\uc740 MuseGAN\uc774 \ud50c\ub808\uc774\uc5b4 \ub610\ub294 \uc791\uace1\uac00\uc640 \uac19\uc740 \uc5ed\ud560\uc744 \ud558\uba74\uc11c\ub3c4, \uc778\uac04\uc774 \ub290\ub07c\ub294 \uac10\uc815\uacfc \uc74c\uc545\uc801 \ub17c\ub9ac\ub97c \ud559\uc2b5\ud560 \uc218 \uc788\ub3c4\ub85d \ub3d5\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n\n3. \ud30c\uc774\ud1a0\uce58(PyTorch)\ub85c MuseGAN \uad6c\ud604\ud558\uae30<\/h2>\n
\uc774\uc81c MuseGAN\uc744 \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud558\uc5ec \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. MuseGAN\uc744 \uad6c\ud604\ud558\uae30 \uc704\ud574\uc11c\ub294 \uae30\ubcf8\uc801\uc73c\ub85c \ub450 \uac1c\uc758 \ub124\ud2b8\uc6cc\ud06c(\uc0dd\uc131\uae30\uc640 \ud310\ubcc4\uae30)\uac00 \ud544\uc694\ud569\ub2c8\ub2e4.<\/p>\n
\uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uace0 \uac00\uc838\uc640\uc57c \ud569\ub2c8\ub2e4:<\/p>\n
!pip install torch torchvision<\/code><\/pre>\n\uc774\uc81c \uc0dd\uc131\uae30\uc640 \ud310\ubcc4\uae30\ub97c \uc704\ud55c \uae30\ubcf8\uc801\uc778 \ud074\ub798\uc2a4 \uad6c\uc870\ub97c \uc124\uc815\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4:<\/p>\n
import torch\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, 128),\n nn.ReLU(),\n nn.Linear(128, 256),\n nn.ReLU(),\n nn.Linear(256, 512),\n nn.ReLU(),\n nn.Linear(512, 88), # MIDI \uc74c\uc5ed\uc5d0 \ub9de\ub294 \ucd9c\ub825 \ud06c\uae30\n nn.Tanh() # \uc74c\uc758 \ubc94\uc704\ub97c -1\uc5d0\uc11c 1\ub85c \uc870\uc815\n )\n\n def forward(self, z):\n return self.model(z)\n\nclass Discriminator(nn.Module):\n def __init__(self):\n super(Discriminator, self).__init__()\n self.model = nn.Sequential(\n nn.Linear(88, 512),\n nn.ReLU(),\n nn.Linear(512, 256),\n nn.ReLU(),\n nn.Linear(256, 128),\n nn.ReLU(),\n nn.Linear(128, 1),\n nn.Sigmoid() # \ucd9c\ub825\uac12\uc744 0\uacfc 1 \uc0ac\uc774\ub85c \uc81c\ud55c\n )\n\n def forward(self, x):\n return self.model(x)\n <\/code><\/pre>\n\uc704\uc758 \ucf54\ub4dc\ub294 \uae30\ubcf8\uc801\uc778 \uc0dd\uc131\uae30\uc640 \ud310\ubcc4\uae30 \uad6c\uc870\ub97c \uc815\uc758\ud569\ub2c8\ub2e4. \uc0dd\uc131\uae30\ub294 \ub79c\ub364\ud55c \ub178\uc774\uc988\ub97c \uc785\ub825 \ubc1b\uc544 MIDI \ud615\uc2dd\uc758 \ub370\uc774\ud130\ub97c \ucd9c\ub825\ud558\uba70, \ud310\ubcc4\uae30\ub294 \uc774\ub7ec\ud55c \ub370\uc774\ud130\ub97c \ubc1b\uc544 \uc9c4\uc9dc \ub370\uc774\ud130\uc778\uc9c0 \uac00\uc9dc \ub370\uc774\ud130\uc778\uc9c0\ub97c \ud310\ub2e8\ud569\ub2c8\ub2e4.<\/p>\n
\uc774\uc81c GAN\uc744 \ud559\uc2b5\ud558\ub294 \uacfc\uc815\uc744 \uc815\uc758\ud574\uc57c \ud569\ub2c8\ub2e4. \ud559\uc2b5\uc5d0\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ub2e8\uacc4\uac00 \ud544\uc694\ud569\ub2c8\ub2e4:<\/p>\n
\n- \uba3c\uc800, \uc2e4\uc81c \ub370\uc774\ud130\uc640 \uac00\uc9dc \ub370\uc774\ud130\ub97c \uc0dd\uc131\ud558\uace0 \ud310\ubcc4\uae30\ub85c \uc785\ub825\ud569\ub2c8\ub2e4.<\/li>\n
- \ud310\ubcc4\uae30\uc758 \uc190\uc2e4(loss)\uc744 \uacc4\uc0b0\ud558\uace0 \uc5ed\uc804\ud30c(backpropagation)\ub97c \ud1b5\ud574 \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/li>\n
- \uc0dd\uc131\uae30\uc758 \uc190\uc2e4\uc744 \uacc4\uc0b0\ud558\uace0 \ub610\ub2e4\uc2dc \uc5ed\uc804\ud30c\ub97c \ud1b5\ud574 \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/li>\n<\/ul>\n
\ub2e4\uc74c\uc740 GAN\uc758 \ud559\uc2b5 \ub8e8\ud504\ub97c \uad6c\ud604\ud55c \ucf54\ub4dc\uc785\ub2c8\ub2e4:<\/p>\n
def train_gan(generator, discriminator, data_loader, num_epochs=100, lr=0.0002):\n criterion = nn.BCELoss() # Binary Cross Entropy Loss\n optimizer_G = torch.optim.Adam(generator.parameters(), lr=lr)\n optimizer_D = torch.optim.Adam(discriminator.parameters(), lr=lr)\n\n for epoch in range(num_epochs):\n for real_data in data_loader:\n batch_size = real_data.size(0)\n\n # \uc9c4\uc9dc \ub370\uc774\ud130\uc640 \uac00\uc9dc \ub370\uc774\ud130\uc758 \ub808\uc774\ube14 \uc0dd\uc131\n real_labels = torch.ones(batch_size, 1)\n fake_labels = torch.zeros(batch_size, 1)\n\n # \ud310\ubcc4\uae30 \ud559\uc2b5\n optimizer_D.zero_grad()\n outputs = discriminator(real_data)\n d_loss_real = criterion(outputs, real_labels)\n d_loss_real.backward()\n\n z = torch.randn(batch_size, 100) # \ub79c\ub364 \ub178\uc774\uc988 \uc0dd\uc131\n fake_data = generator(z)\n outputs = discriminator(fake_data.detach())\n d_loss_fake = criterion(outputs, fake_labels)\n d_loss_fake.backward()\n\n optimizer_D.step()\n\n # \uc0dd\uc131\uae30 \ud559\uc2b5\n optimizer_G.zero_grad()\n outputs = discriminator(fake_data)\n g_loss = criterion(outputs, real_labels)\n g_loss.backward()\n optimizer_G.step()\n \n if epoch % 10 == 0:\n print(f'Epoch [{epoch}\/{num_epochs}], d_loss: {d_loss_real.item() + d_loss_fake.item()}, g_loss: {g_loss.item()}')\n <\/code><\/pre>\n\uc5ec\uae30\uc11c, train_gan<\/code> \ud568\uc218\ub294 \uc0dd\uc131\uae30\uc640 \ud310\ubcc4\uae30\ub97c \ud559\uc2b5\uc2dc\ud0a4\ub294 \ub8e8\ud504\ub97c \uad6c\ud604\ud569\ub2c8\ub2e4. \uc774 \ub8e8\ud504\ub294 data_loader<\/code> \ub97c \ud1b5\ud574 \uc2e4\uc81c \ub370\uc774\ud130\ub97c \ubc1b\uc544\uc624\uace0, \uac01 \ub124\ud2b8\uc6cc\ud06c\uc758 \uc190\uc2e4\uc744 \uacc4\uc0b0\ud558\uc5ec \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/p>\n\uc774\uc81c MuseGAN\uc744 \uc644\uc804\ud788 \uad6c\ud604\ud558\uace0 \ub098\uba74, \ub2e4\uc591\ud55c MIDI \ud30c\uc77c\uc744 \uc0dd\uc131\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \uc0dd\uc131\ub41c \ub370\uc774\ud130\ub97c MIDI \ud615\uc2dd\uc73c\ub85c \ubcc0\ud658\ud558\uc5ec \ucd9c\ub825\ud574\uc57c \ud569\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \uac04\ub2e8\ud55c MIDI \ud30c\uc77c\uc744 \uc0dd\uc131\ud558\ub294 \ucf54\ub4dc\uc785\ub2c8\ub2e4:<\/p>\n
from mido import Message, MidiFile\n\ndef save_to_midi(generated_data, filename='output.mid'):\n mid = MidiFile()\n track = mid.add_track('Generated Music')\n\n for note in generated_data:\n track.append(Message('note_on', note=int(note), velocity=64, time=0))\n track.append(Message('note_off', note=int(note), velocity=64, time=32))\n\n mid.save(filename)\n\n# GAN\uc744 \ud559\uc2b5\ud55c \ud6c4 \uc0dd\uc131\ub41c \ub370\uc774\ud130\ub97c MIDI \ud30c\uc77c\ub85c \uc800\uc7a5\ngenerated_data = generator(torch.randn(16, 100)).detach().numpy()\nsave_to_midi(generated_data[0]) # \uccab \ubc88\uc9f8 \uc0dd\uc131\ub41c \uc74c\uc545\uc744 \uc800\uc7a5\n <\/code><\/pre>\nMuseGAN\uc744 \ud1b5\ud574 \uc0dd\uc131\ub41c \uc74c\uc545\uc744 \uc2e4\uc81c\ub85c \ub4e4\uc5b4\ubcf4\uba74 \ud765\ubbf8\ub85c\uc6b4 \uacb0\uacfc\ub97c \uc5bb\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\uc81c \uc5ec\ub7ec\ubd84\ub3c4 GAN\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc74c\uc545 \uc0dd\uc131\uc774\ub77c\ub294 \ucc3d\uc758\uc801\uc778 \uc791\uc5c5\uc5d0 \ub3c4\uc804\ud574\ubcf4\uc138\uc694!<\/p>\n<\/section>\n\n4. \uacb0\ub860<\/h2>\n
MuseGAN\uacfc \uac19\uc740 GAN \uae30\ubc18 \ubaa8\ub378\uc740 \uc74c\uc545 \uc0dd\uc131 \ubfd0\ub9cc \uc544\ub2c8\ub77c \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. GAN\uc758 \uc6d0\ub9ac\uc640 MuseGAN\uc758 \uad6c\uc870\ub97c \uc774\ud574\ud568\uc73c\ub85c\uc368 \uc6b0\ub9ac\ub294 \ub525\ub7ec\ub2dd\uc758 \uae30\ucd08\ub97c \ub2e4\uc9c0\uace0, \ucc3d\uc758\uc801\uc778 \ud504\ub85c\uc81d\ud2b8\ub97c \ub9cc\ub4e4 \uc218 \uc788\ub294 \uae30\ucd08\ub97c \ub9c8\ub828\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub354 \ub9ce\uc740 \uc5f0\uad6c\uc640 \uac1c\ubc1c\uc774 \uc774\ub8e8\uc5b4\uc9c8 \uac83\uc774\uba70, \ub525\ub7ec\ub2dd\uacfc GAN\uc758 \ubbf8\ub798\ub294 \ub354\uc6b1 \ubc1d\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n