\ub525\ub7ec\ub2dd\uc758 \ubc1c\uc804\uacfc \ub2e4\uc591\ud55c \uc751\uc6a9\uc73c\ub85c \uc778\ud574, \ub370\uc774\ud130\uc14b\uc774 \ucee4\uc9c0\uace0 \ubaa8\ub378\uc758 \ubcf5\uc7a1\uc131\uc774 \uc99d\uac00\ud568\uc5d0 \ub530\ub77c, \ub354 \ub9ce\uc740 \uacc4\uc0b0 \uc790\uc6d0\uc774 \ud544\uc694\ud558\uac8c \ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uae4a\uc740 \uc2e0\uacbd\ub9dd\uc744 \ud559\uc2b5\uc2dc\ud0a4\ub294 \ub370 \uc788\uc5b4, GPU\uc758 \ud65c\uc6a9\uc740 \ud544\uc218\uc801\uc785\ub2c8\ub2e4. \uc774 \uac15\uc88c\uc5d0\uc11c\ub294 PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec GPU\ub97c \ud65c\uc6a9\ud55c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \uc5b4\ub5bb\uac8c \ucd5c\uc801\ud654\ud560 \uc218 \uc788\ub294\uc9c0\uc5d0 \ub300\ud574 \ub2e4\ub8f0 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
GPU(\uadf8\ub798\ud53d \ucc98\ub9ac \uc7a5\uce58)\ub294 \ubcd1\ub82c \ucc98\ub9ac\uc5d0 \ucd5c\uc801\ud654\ub41c \ucef4\ud4e8\ud305 \uc720\ub2db\uc73c\ub85c, \ub9ce\uc740 \uc218\uc758 \uc5f0\uc0b0\uc744 \ub3d9\uc2dc\uc5d0 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 \ud2b9\ud788 \ub525\ub7ec\ub2dd\uacfc \uac19\uc740 \ub300\uaddc\ubaa8 \uacc4\uc0b0\uc5d0\uc11c \ud6a8\uacfc\uc801\uc785\ub2c8\ub2e4. CPU(\uc911\uc559 \ucc98\ub9ac \uc7a5\uce58)\uc5d0 \ube44\ud574 GPU\ub294 \uc218\ucc9c \uac1c\uc758 \ucf54\uc5b4\ub97c \uac16\ucd94\uace0 \uc788\uc5b4, \ub300\ub7c9\uc758 \ud589\ub82c \uc5f0\uc0b0\uc744 \uc2e0\uc18d\ud558\uac8c \ucc98\ub9ac\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
PyTorch\ub294 GPU\uc5d0\uc11c\uc758 \uc5f0\uc0b0\uc744 \uc9c0\uc6d0\ud558\ub294 \ub6f0\uc5b4\ub09c \ud504\ub808\uc784\uc6cc\ud06c\uc785\ub2c8\ub2e4. GPU\ub97c \uc0ac\uc6a9\ud558\uae30 \uc704\ud574\uc11c\ub294 \uba3c\uc800 PyTorch\uac00 GPU\ub97c \uc9c0\uc6d0\ud558\ub294 \ubc84\uc804\uc774 \uc124\uce58\ub418\uc5b4 \uc788\uc5b4\uc57c \ud558\uba70, CUDA\uac00 \uc124\uce58\ub41c NVIDIA GPU\uac00 \ud544\uc694\ud569\ub2c8\ub2e4.<\/p>\n
PyTorch\ub97c \uc124\uce58\ud558\uae30 \uc704\ud574\uc11c\ub294 \uc544\ub798\uc758 \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \uc124\uce58 \uc2dc, CUDA \ubc84\uc804\uc744 \uc120\ud0dd\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
pip install torch torchvision torchaudio --extra-index-url https:\/\/download.pytorch.org\/whl\/cu113<\/code><\/pre>\n3. \ubaa8\ub378\uacfc \ub370\uc774\ud130\uc758 GPU \uc774\ub3d9<\/h2>\n
PyTorch\uc5d0\uc11c\ub294 `.to()` \uba54\uc18c\ub4dc\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud150\uc11c\uc640 \ubaa8\ub378\uc744 GPU\ub85c \uc774\ub3d9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\uc758 \uc608\uc81c\ub97c \ud1b5\ud574 \uc774 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc608\uc81c \ucf54\ub4dc: \ud150\uc11c\uc640 \ubaa8\ub378\uc744 GPU\ub85c \uc774\ub3d9<\/h3>\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\n\n# GPU \uc0ac\uc6a9 \uac00\ub2a5 \uc5ec\ubd80 \ud655\uc778\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n\n# \uac04\ub2e8\ud55c \uc2e0\uacbd\ub9dd \ubaa8\ub378 \uc815\uc758\nclass SimpleNN(nn.Module):\n def __init__(self):\n super(SimpleNN, self).__init__()\n self.fc1 = nn.Linear(10, 50)\n self.fc2 = nn.Linear(50, 1)\n\n def forward(self, x):\n x = torch.relu(self.fc1(x))\n x = self.fc2(x)\n return x\n\n# \ubaa8\ub378 \uc778\uc2a4\ud134\uc2a4\ud654 \ubc0f GPU\ub85c \uc774\ub3d9\nmodel = SimpleNN().to(device)\n\n# \ub370\uc774\ud130 \ud150\uc11c \uc815\uc758 \ubc0f GPU\ub85c \uc774\ub3d9\ndata = torch.randn(64, 10).to(device)\noutput = model(data)\nprint(output.shape) # (64, 1)\n<\/code><\/pre>\n4. \uc131\ub2a5 \ucd5c\uc801\ud654 \uae30\uc220<\/h2>\n
GPU\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \ud65c\uc6a9\ud558\uae30 \uc704\ud574 \uba87 \uac00\uc9c0 \uc131\ub2a5 \ucd5c\uc801\ud654 \uae30\uc220\uc744 \uace0\ub824\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
4.1 \ubc30\uce58 \ucc98\ub9ac<\/h3>\n
\uc77c\ubc18\uc801\uc73c\ub85c \ub354 \ud070 \ubc30\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec GPU\ub97c \ucd5c\ub300\ud55c \ud65c\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud558\uc9c0\ub9cc \ubc30\uce58 \ud06c\uae30\ub97c \ub108\ubb34 \ud06c\uac8c \uc124\uc815\ud558\uba74 GPU \uba54\ubaa8\ub9ac\uac00 \ubd80\uc871\ud560 \uc218 \uc788\uc73c\ubbc0\ub85c \uc801\uc808\ud55c \ud06c\uae30\ub97c \ud655\uc778\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
4.2 Mixed Precision Training<\/h3>\n
Mixed Precision Training\uc740 \uc5f0\uc0b0\uc744 16\ube44\ud2b8\uc640 32\ube44\ud2b8 \ud63c\ud569\uc73c\ub85c \ucc98\ub9ac\ud558\ub294 \ubc29\ubc95\uc73c\ub85c, \uba54\ubaa8\ub9ac \uc0ac\uc6a9\ub7c9\uc744 \uc904\uc774\uace0 \uc131\ub2a5\uc744 \uac1c\uc120\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 NVIDIA\uc758 Apex \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
!pip install nvidia-apex<\/code><\/pre>\n4.3 Gradient Accumulation<\/h3>\n
\uba54\ubaa8\ub9ac \uc81c\ud55c\uc73c\ub85c \uc778\ud574 \ubc30\uce58 \ud06c\uae30\ub97c \ub298\ub9b4 \uc218 \uc5c6\ub294 \uacbd\uc6b0, \uc5ec\ub7ec \uc2a4\ud15d\uc758 gradients\ub97c \ub204\uc801\ud558\uc5ec \ucd5c\uc885 \uc5c5\ub370\uc774\ud2b8\ub97c \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub807\uac8c \ud558\uba74 \ub098\uc911\uc5d0 \ub354 \ud070 \ubc30\uce58 \ud06c\uae30\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
4.4 \ub370\uc774\ud130 \ub85c\ub529 \ucd5c\uc801\ud654<\/h3>\n
DataLoader\uc758 num_workers<\/code> \uc18d\uc131\uc744 \ud65c\uc6a9\ud558\uc5ec \ub370\uc774\ud130\ub97c \ubcd1\ub82c\ub85c \ub85c\ub4dc\ud568\uc73c\ub85c\uc368 \ub370\uc774\ud130 \uc900\ube44 \uc2dc\uac04\uc744 \ub2e8\ucd95\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\nfrom torch.utils.data import DataLoader, TensorDataset\n\ndataset = TensorDataset(data, target)\ndataloader = DataLoader(dataset, batch_size=64, num_workers=4)\n<\/code><\/pre>\n5. \uc608\uc81c \ucf54\ub4dc<\/h2>\n
\uc544\ub798\uc758 \ucf54\ub4dc\ub294 \uc804\uccb4\uc801\uc778 \ud504\ub85c\uc138\uc2a4\ub97c \ubcf4\uc5ec\uc8fc\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c \ubaa8\ub378\uc744 \uc815\uc758\ud558\uace0, \ub370\uc774\ud130\ub97c \ub85c\ub4dc\ud558\uba70, GPU\uc5d0\uc11c \ud559\uc2b5\uc744 \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc744 \uc124\uba85\ud569\ub2c8\ub2e4.<\/p>\n
import torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torch.utils.data import DataLoader, TensorDataset\n\n# GPU \uc0ac\uc6a9 \uc5ec\ubd80 \ud655\uc778\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\n\n# \ub370\uc774\ud130\uc14b \uc0dd\uc131\nX = torch.randn(1000, 10).to(device)\ny = torch.randn(1000, 1).to(device)\n\n# TensorDataset \ubc0f DataLoader\ndataset = TensorDataset(X, y)\ndataloader = DataLoader(dataset, batch_size=64, num_workers=4)\n\n# \uc2e0\uacbd\ub9dd \ubaa8\ub378\nclass SimpleNN(nn.Module):\n def __init__(self):\n super(SimpleNN, self).__init__()\n self.fc1 = nn.Linear(10, 50)\n self.fc2 = nn.Linear(50, 1)\n\n def forward(self, x):\n x = torch.relu(self.fc1(x))\n return self.fc2(x)\n\n# \ubaa8\ub378 \uc778\uc2a4\ud134\uc2a4\ud654 \ubc0f optimizer \uc124\uc815\nmodel = SimpleNN().to(device)\noptimizer = optim.Adam(model.parameters(), lr=0.001)\ncriterion = nn.MSELoss()\n\n# \ud559\uc2b5 \ub8e8\ud504\nnum_epochs = 10\nfor epoch in range(num_epochs):\n for data, target in dataloader:\n # \ub370\uc774\ud130\uc640 \ud0c0\uac9f\uc744 GPU\ub85c \uc774\ub3d9\n data, target = data.to(device), target.to(device)\n\n optimizer.zero_grad() # gradient \ucd08\uae30\ud654\n output = model(data) # \ubaa8\ub378 forward \uc804\ud30c\n loss = criterion(output, target) # \uc190\uc2e4 \uacc4\uc0b0\n loss.backward() # backward \uc804\ud30c\n optimizer.step() # optimizer \uc5c5\ub370\uc774\ud2b8\n\n print(f'Epoch [{epoch+1}\/{num_epochs}], Loss: {loss.item():.4f}')\n<\/code><\/pre>\n6. \uacb0\ub860<\/h2>\n
\ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\ub294 \ub370 GPU\ub97c \uc0ac\uc6a9\ud558\ub294 \uac83\uc740 \ud544\uc218\uc801\uc774\uba70, PyTorch\ub294 \uc774\ub97c \uc704\ud55c \uac15\ub825\ud55c \ub3c4\uad6c\uc785\ub2c8\ub2e4. \ubaa8\ub378\uacfc \ub370\uc774\ud130\ub97c GPU\ub85c \uc774\ub3d9\uc2dc\ud0a4\uace0, \ubc30\uce58 \ucc98\ub9ac\ub97c \ud1b5\ud574 \uc131\ub2a5\uc744 \ucd5c\uc801\ud654\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \ucd94\uac00\ub85c Mixed Precision Training\uacfc Gradient Accumulation \uac19\uc740 \uae30\uc220\ub4e4\uc744 \ud65c\uc6a9\ud558\uc5ec \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \ub04c\uc5b4\ub0bc \uc218\ub3c4 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc774 \uac15\uc88c\ub97c \ud1b5\ud574 PyTorch\uc640 GPU\ub97c \ud65c\uc6a9\ud558\uc5ec \ub525\ub7ec\ub2dd \uc131\ub2a5\uc744 \ucd5c\uc801\ud654\ud558\ub294 \ubc29\ubc95\uc744 \uc798 \uc774\ud574\ud558\uc168\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4. \uc774\uc81c \uc5ec\ub7ec\ubd84\uc740 \ub354 \ubcf5\uc7a1\ud55c \ubaa8\ub378\uacfc \ub300\ub7c9\uc758 \ub370\uc774\ud130\ub85c \uc791\uc5c5\ud560 \uc900\ube44\uac00 \ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4!<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"