\uc778\uacf5\uc9c0\ub2a5(AI) \ud6c8\ub828\uc744 \uc704\ud574 \uac15\ub825\ud55c \uc11c\ubc84 \ud658\uacbd\uc744 \uad6c\ucd95\ud574\uc57c \ud569\ub2c8\ub2e4. \ud2b9\ud788 \ub300\uaddc\ubaa8 \ub370\uc774\ud130\uc14b\uc774\ub098 \ubcf5\uc7a1\ud55c \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud560 \ub54c\ub294 \ud558\ub098\uc758 GPU\ub098 \ub178\ub4dc\ub9cc\uc73c\ub85c\ub294 \ucda9\ubd84\ud558\uc9c0 \uc54a\uc744 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 \uc6b0\ubd84\ud22c \uc11c\ubc84\uc5d0\uc11c \ub2e4\uc911 GPU<\/strong> \ubc0f \ub2e4\uc911 \ub178\ub4dc<\/strong>\ub97c \ud65c\uc6a9\ud55c \ubd84\uc0b0 \ud6c8\ub828\uc744 \uc124\uc815\ud558\ub294 \ubc29\ubc95\uc744 \uc790\uc138\ud788 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n \ubd84\uc0b0 \ud6c8\ub828\uc740 \uc5ec\ub7ec \uac1c\uc758 \uacc4\uc0b0 \uc790\uc6d0\uc744 \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378 \ud6c8\ub828\uc744 \ubcd1\ub82c\ub85c \uc218\ud589\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \ud6c8\ub828 \uc2dc\uac04\uc744 \ub2e8\ucd95\ud558\uace0, \ub354 \ub9ce\uc740 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubd84\uc0b0 \ud6c8\ub828\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \uacbd\uc6b0 \uc720\uc6a9\ud569\ub2c8\ub2e4:<\/p>\n \uc6b0\ubd84\ud22c \uc11c\ubc84\uc5d0\uc11c \ubd84\uc0b0 \ud6c8\ub828\uc744 \uc124\uc815\ud558\uae30 \uc704\ud574\uc11c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uc900\ube44\uac00 \ud544\uc694\ud569\ub2c8\ub2e4:<\/p>\n \uc6b0\ubd84\ud22c \uc11c\ubc84\ub97c \uc124\uce58\ud558\uace0, SSH\ub97c \ud1b5\ud574 \uc811\uc18d\ud569\ub2c8\ub2e4. \uae30\ubcf8\uc801\uc778 \uc2dc\uc2a4\ud15c \uc5c5\ub370\uc774\ud2b8\ub97c \uc9c4\ud589\ud569\ub2c8\ub2e4:<\/p>\n GPU \uac00\uc18d\uc744 \uc704\ud574 NVIDIA\uc758 CUDA \ubc0f cuDNN\uc744 \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. NVIDIA\uc758 \uacf5\uc2dd \uc6f9\uc0ac\uc774\ud2b8\uc5d0\uc11c \uc124\uce58 \ubc29\ubc95\uc744 \ud655\uc778\ud558\uace0, \ud544\uc694\ud55c \ud328\ud0a4\uc9c0\ub97c \uc124\uce58\ud569\ub2c8\ub2e4. CUDA \uc124\uce58 \uc608\uc2dc\ub294 \uc544\ub798\uc640 \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n CUDA\uac00 \uc124\uce58\ub418\uba74, cuDNN\ub3c4 \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. CUDA Toolkit\uc5d0 \ub9de\ub294 \ubc84\uc804\uc758 cuDNN\uc744 NVIDIA\uc5d0\uc11c \ub2e4\uc6b4\ub85c\ub4dc\ud558\uc5ec \uc124\uce58\ud569\ub2c8\ub2e4.<\/p>\n Pytorch\uc640 TensorFlow \ub450 \uac00\uc9c0\uc758 \ub300\ud45c\uc801\uc778 \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\ub97c \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 Pytorch \uc608\uc81c\ub97c \uc18c\uac1c\ud558\uaca0\uc2b5\ub2c8\ub2e4:<\/p>\n \ub2e4\uc911 \ub178\ub4dc\ub97c \uc0ac\uc6a9\ud560 \uacbd\uc6b0, \uc11c\ub85c\uc758 \ub178\ub4dc\uc5d0 SSH\ub85c \uc811\uc18d\ud560 \uc218 \uc788\uc5b4\uc57c \ud569\ub2c8\ub2e4. SSH \ud0a4\ub97c \uc0dd\uc131\ud558\uace0, \ud074\ub77c\uc774\uc5b8\ud2b8 \ub178\ub4dc\uc758 \ud37c\ube14\ub9ad \ud0a4\ub97c \uac01 \uc11c\ubc84 \ub178\ub4dc\uc5d0 \ucd94\uac00\ud569\ub2c8\ub2e4:<\/p>\n \uae30\ubcf8\uc801\uc778 \ud328\ud0a4\uc9c0\ub97c \uc124\uce58\ud55c \ud6c4, \uba40\ud2f0 GPU\ub97c \ud65c\uc6a9\ud558\uae30 \uc704\ud55c \ud328\ud0a4\uc9c0\ub97c \ucd94\uac00\ub85c \uc124\uce58\ud569\ub2c8\ub2e4. \uac00\uc0c1\ud658\uacbd\uc744 \uc0ac\uc6a9\ud558\ub294 \uac83\uc774 \uc88b\uc2b5\ub2c8\ub2e4:<\/p>\n \uc774\uc81c \uc2e4\uc81c\ub85c \ubd84\uc0b0 \ud6c8\ub828\uc744 \uc124\uc815\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. Pytorch\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub2e4\uc911 GPU\uc5d0\uc11c \ud6c8\ub828\ud558\ub294 \ubc29\ubc95\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4.<\/p>\n Pytorch\uc758 \ub2e4\uc911 \ub178\ub4dc\uc5d0\uc11c\uc758 \ud6c8\ub828\uc740 \ud6c8\ub828 \uacfc\uc815\uc5d0\uc11c \uc2a4\ud06c\ub9bd\ud2b8\ub97c \uc2e4\ud589\ud558\uae30 \uc704\ud574\uc11c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud569\ub2c8\ub2e4:<\/p>\n \uc5ec\uae30\uc11c G<\/strong>\ub294 \uac01 \ub178\ub4dc\ubcc4 GPU \uc218, N<\/strong>\uc740 \ucd1d \ub178\ub4dc \uc218, R<\/strong>\uc740 \ud604\uc7ac \ub178\ub4dc\uc758 \uc21c\ubc88\uc785\ub2c8\ub2e4.<\/p>\n \ubd84\uc0b0 \ud6c8\ub828\uc744 \uc131\uacf5\uc801\uc73c\ub85c \uad6c\ud604\ud55c \ud6c4, \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\uae30 \uc704\ud55c \uba87 \uac00\uc9c0 \ubc29\ubc95\uc774 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n\ubd84\uc0b0 \ud6c8\ub828\uc774\ub780?<\/h2>\n
\n
\uc0ac\uc804 \uc900\ube44 \uc0ac\ud56d<\/h2>\n
\n
1. \uc6b0\ubd84\ud22c \uc11c\ubc84 \uc124\uce58<\/h2>\n
sudo apt update && sudo apt upgrade -y<\/code><\/pre>\n2. CUDA \ubc0f cuDNN \uc124\uce58<\/h2>\n
sudo apt install nvidia-cuda-toolkit<\/code><\/pre>\n3. \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \uc124\uce58<\/h2>\n
pip install torch torchvision torchaudio --extra-index-url https:\/\/download.pytorch.org\/whl\/cu113<\/code><\/pre>\n4. SSH \uc124\uc815<\/h2>\n
ssh-keygen -t rsa<\/code><\/pre>\nssh-copy-id user@remote-server<\/code><\/pre>\n5. \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac \ubc0f \ud328\ud0a4\uc9c0 \uc124\uce58<\/h2>\n
python -m venv venv<\/code><\/pre>\nsource venv\/bin\/activate<\/code><\/pre>\npip install -U pip setuptools<\/code><\/pre>\n\ub2e4\uc911 GPU \ubc0f \ub2e4\uc911 \ub178\ub4dc\uc5d0\uc11c \ubd84\uc0b0 \ud6c8\ub828 \uc124\uc815\ud558\uae30<\/h2>\n
\ub2e4\uc911 GPU \ud6c8\ub828 \uc124\uc815 \uc608\uc81c<\/h3>\n
torch.nn.DataParallel<\/code> API\ub97c \uc0ac\uc6a9\ud558\uc5ec \ub2e4\uc911 GPU\uc5d0\uc11c \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\ub294 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/p>\n\nimport torch\nimport torch.nn as nn\nimport torchvision.models as models\n\n# \ubaa8\ub378 \uc815\uc758\nmodel = models.resnet50(pretrained=True)\nmodel = nn.DataParallel(model) # \ub370\uc774\ud130 \ubcd1\ub82c \ucc98\ub9ac\nmodel.to('cuda') # CUDA \ub514\ubc14\uc774\uc2a4\ub85c \uc774\ub3d9\n\n# \ud6c8\ub828 \ub8e8\ud504 (\uc608\uc2dc)\ndef train(model, dataloader, optimizer):\n model.train()\n for inputs, labels in dataloader:\n inputs, labels = inputs.to('cuda'), labels.to('cuda')\n optimizer.zero_grad()\n outputs = model(inputs)\n loss = criterion(outputs, labels)\n loss.backward()\n optimizer.step()\n<\/code><\/pre>\n\ub2e4\uc911 \ub178\ub4dc \ud6c8\ub828 \uc124\uc815 \uc608\uc81c<\/h3>\n
torch.distributed<\/code> API\ub97c \uc0ac\uc6a9\ud558\uc5ec \uad6c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \uba3c\uc800 \ud544\uc694\ud55c \ud658\uacbd \ubcc0\uc218\ub97c \uc124\uc815\ud569\ub2c8\ub2e4:<\/p>\n\nimport os\nimport torch\nimport torch.distributed as dist\n\n# hyperparameter\nworld_size = 2 # \ucd1d \ub178\ub4dc \uc218\nprocess_rank = int(os.environ['RANK']) # \ud604\uc7ac \ub178\ub4dc \ubc88\ud638\ndist.init_process_group(backend='nccl', world_size=world_size, rank=process_rank)\n\n# \ubaa8\ub378 \uc6d0\uaca9 \ucd08\uae30\ud654\nmodel = models.resnet50(pretrained=True).to(process_rank)\nmodel = nn.parallel.DistributedDataParallel(model, device_ids=[process_rank])\n\n# \ub370\uc774\ud130\ub85c\ub354 \ucd08\uae30\ud654\ntrain_sampler = torch.utils.data.distributed.DistributedSampler(train_dataset)\ntrain_loader = torch.utils.data.DataLoader(train_dataset, sampler=train_sampler)\n<\/code><\/pre>\n\ud6c8\ub828 \uc2a4\ud06c\ub9bd\ud2b8 \uc2e4\ud589\ud558\uae30<\/h2>\n
python -m torch.distributed.launch --nproc_per_node=G --nnodes=N --node_rank=R train.py<\/code><\/pre>\n\uc131\ub2a5 \ucd5c\uc801\ud654<\/h2>\n