{"id":30138,"date":"2024-10-28T03:19:46","date_gmt":"2024-10-28T03:19:46","guid":{"rendered":"http:\/\/atmokpo.com\/w\/?p=30138"},"modified":"2024-11-26T06:49:46","modified_gmt":"2024-11-26T06:49:46","slug":"%eb%94%a5%eb%9f%ac%eb%8b%9d-%ed%8c%8c%ec%9d%b4%ed%86%a0%ec%b9%98-%ea%b0%95%ec%a2%8c-%ed%8c%8c%ec%9d%b4%ed%86%a0%ec%b9%98-%ed%8a%b9%ec%a7%95-%eb%b0%8f-%ec%9e%a5%ec%a0%90","status":"publish","type":"post","link":"https:\/\/atmokpo.com\/w\/30138\/","title":{"rendered":"\ub525\ub7ec\ub2dd \ud30c\uc774\ud1a0\uce58 \uac15\uc88c, \ud30c\uc774\ud1a0\uce58 \ud2b9\uc9d5 \ubc0f \uc7a5\uc810"},"content":{"rendered":"

<\/p>\n

\ub525\ub7ec\ub2dd\uc740 \ud604\ub300 \uae30\uacc4\ud559\uc2b5\uc758 \ud575\uc2ec \uae30\uc220\ub85c \uc790\ub9ac\uc7a1\uc558\uc2b5\ub2c8\ub2e4. \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud65c\uc6a9\ub418\uba70, \uc774\ubbf8 \uc6b0\ub9ac\uc758 \uc77c\uc0c1\uc0dd\ud65c\uc5d0\ub3c4 \uae4a\uc219\uc774 \uce68\ud22c\ud574 \uc788\uc2b5\ub2c8\ub2e4. \ud2b9\ud788, \ud30c\uc774\ud1a0\uce58(PyTorch)\ub294 \uc774\ub7ec\ud55c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \uad6c\ucd95\ud558\ub294 \ub370 \uc788\uc5b4 \ub9e4\uc6b0 \uc778\uae30 \uc788\ub294 \ud504\ub808\uc784\uc6cc\ud06c \uc911 \ud558\ub098\uc785\ub2c8\ub2e4. \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58\uc758 \ud2b9\uc9d5\uacfc \uc7a5\uc810\uc5d0 \ub300\ud574 \uc0c1\uc138\ud788 \uc54c\uc544\ubcf4\ub3c4\ub85d \ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n

1. \ud30c\uc774\ud1a0\uce58\ub780?<\/h2>\n

\ud30c\uc774\ud1a0\uce58\ub294 Facebook AI Research Lab(FAIR)\uc5d0\uc11c \uac1c\ubc1c\ud55c \uc624\ud508\uc18c\uc2a4 \uba38\uc2e0\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\uc785\ub2c8\ub2e4. \uc774 \ud504\ub808\uc784\uc6cc\ud06c\ub294 \ub525\ub7ec\ub2dd \uc5f0\uad6c\uc790\uc640 \uac1c\ubc1c\uc790\ub4e4\uc5d0\uac8c \uac15\ub825\ud558\uace0 \uc720\uc5f0\ud55c \ub3c4\uad6c\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4. \ud30c\uc774\ud1a0\uce58\ub294 Python \uae30\ubc18\uc73c\ub85c \ub418\uc5b4 \uc788\uc73c\uba70, \ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504(dynamic computation graph)\ub77c\ub294 \ub3c5\ud2b9\ud55c \ud2b9\uc9d5\uc744 \uac16\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n

2. \ud30c\uc774\ud1a0\uce58\uc758 \uc8fc\uc694 \ud2b9\uc9d5<\/h2>\n

2.1 \ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504<\/h3>\n

\ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504\ub294 \ubaa8\ub378\uc758 \uad6c\uc870\uac00 \uc2e4\ud589 \uc911\uc5d0 \ubcc0\uacbd\ub420 \uc218 \uc788\uc74c\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4. \uc774\ub294 \uac1c\ubc1c\uc790\uac00 \uc870\uac74\ubd80 \ub85c\uc9c1\uc774\ub098 \ubc18\ubcf5 \uad6c\uc870\ub97c \uad6c\ud604\ud560 \ub54c \ub9e4\uc6b0 \uc720\uc6a9\ud569\ub2c8\ub2e4. \uc989, \uac01 iteration\ub9c8\ub2e4 \uacc4\uc0b0 \uadf8\ub798\ud504\ub97c \ubcc0\uacbd\ud560 \uc218 \uc788\uc73c\ubbc0\ub85c, \ucf54\ub4dc\uc758 \uc720\uc5f0\uc131\uc774 \ub9e4\uc6b0 \ub192\uc544\uc9d1\ub2c8\ub2e4.<\/p>\n

import torch\n\n# \uc608\uc2dc: \ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504\nx = torch.tensor(1.0, requires_grad=True)\ny = x ** 2\ny.backward()\nprint(x.grad)  # \ucd9c\ub825: tensor(2.0)\n    <\/code><\/pre>\n
2.2 \uc9c1\uad00\uc801\uc778 API<\/h3>\n
\ud30c\uc774\ud1a0\uce58\ub294 \uc0ac\uc6a9\ud558\uae30 \uc26c\uc6b4 API\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4. Tensor \uc5f0\uc0b0, \uc790\ub3d9 \ubbf8\ubd84, \ubaa8\ub378 \uc0dd\uc131\uc744 \uc9c1\uad00\uc801\uc774\uace0 \uac04\ub2e8\ud558\uac8c \ud560 \uc218 \uc788\ub3c4\ub85d \ud574\uc90d\ub2c8\ub2e4. \uc774\ub294 \ucd08\ubcf4\uc790\uac00 \uacf5\ubd80\ud558\uae30\uc5d0\ub3c4 \uc88b\uc740 \ud658\uacbd\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n
import torch.nn as nn\n\n# \uac04\ub2e8\ud55c \uc2e0\uacbd\ub9dd \uc815\uc758\nclass SimpleNN(nn.Module):\n    def __init__(self):\n        super(SimpleNN, self).__init__()\n        self.fc = nn.Linear(10, 1)\n\n    def forward(self, x):\n        return self.fc(x)\n\nmodel = SimpleNN()\nprint(model)\n    <\/code><\/pre>\n
2.3 GPU \uac00\uc18d<\/h3>\n
\ud30c\uc774\ud1a0\uce58\ub294 GPU\ub97c \ud1b5\ud55c \uc5f0\uc0b0\uc744 \uc27d\uac8c \uad6c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. Tensor\ub97c GPU\ub85c \uc62e\uae30\ub294 \uac83\uc740 \ub2e8\uc21c\ud55c \ucf54\ub4dc \ubcc0\uacbd\uc73c\ub85c \uac00\ub2a5\ud558\uba70, \uc774\ub294 \ub300\uaddc\ubaa8 \ub370\uc774\ud130\uc14b\uc774\ub098 \ubcf5\uc7a1\ud55c \ubaa8\ub378\uc744 \ub2e4\ub8f0 \ub54c \uc131\ub2a5\uc744 \uadf9\ub300\ud654\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
device = torch.device(\"cuda\" if torch.cuda.is_available() else \"cpu\")\nx = torch.tensor([1.0, 2.0]).to(device)  # GPU\ub85c \ud150\uc11c\ub97c \uc774\ub3d9\nprint(x)\n    <\/code><\/pre>\n
2.4 \ud48d\ubd80\ud55c \ucee4\ubba4\ub2c8\ud2f0\uc640 \ubb38\uc11c<\/h3>\n
\ud30c\uc774\ud1a0\uce58\ub294 \ud65c\ubc1c\ud55c \uc0ac\uc6a9\uc790 \ucee4\ubba4\ub2c8\ud2f0\uc640 \uc798 \uc815\ub9ac\ub41c \ubb38\uc11c\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4. \uc774\ub294 \ubb38\uc81c \ud574\uacb0 \ubc0f \uc9c0\uc2dd \uacf5\uc720\uc5d0 \ud070 \ub3c4\uc6c0\uc774 \ub429\ub2c8\ub2e4. \uc0ac\uc6a9\uc790\ub4e4\uc740 \ub2e4\uc591\ud55c \uc608\uc81c\uc640 \ud29c\ud1a0\ub9ac\uc5bc\uc744 \ud1b5\ud574 \ud559\uc2b5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
3. \ud30c\uc774\ud1a0\uce58\uc758 \uc7a5\uc810<\/h2>\n
3.1 \uc720\uc5f0\uc131<\/h3>\n
\ud30c\uc774\ud1a0\uce58\ub294 \uc0ac\uc6a9\uc790\uc5d0\uac8c \ub9ce\uc740 \uc720\uc5f0\uc131\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. \ubaa8\ub378\uc744 \uc124\uacc4\ud560 \ub54c \ubcf5\uc7a1\ud55c \uad6c\uc870\ub97c \uc190\uc27d\uac8c \uad6c\ud604\ud560 \uc218 \uc788\uc5b4, \uc5f0\uad6c \ubc0f \uc2e4\ud5d8\uc5d0 \uc720\ub9ac\ud569\ub2c8\ub2e4.<\/p>\n
3.2 \uc0dd\uc0b0\uc131<\/h3>\n
\ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504\uc640 \uc9c1\uad00\uc801\uc778 API \ub355\ubd84\uc5d0 \uc9e7\uc740 \uc2dc\uac04 \uc548\uc5d0 \uc2e4\ud5d8\uc744 \uc218\ud589\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 \uc5f0\uad6c\uc790 \ubc0f \ub370\uc774\ud130 \uacfc\ud559\uc790\ub4e4\uc5d0\uac8c \ub9e4\uc6b0 \uc911\uc694\ud55c \uc694\uc18c\uc785\ub2c8\ub2e4.<\/p>\n
3.3 \uc131\ub2a5<\/h3>\n
\ud30c\uc774\ud1a0\uce58\ub294 GPU \uac00\uc18d\uc744 \uc9c0\uc6d0\ud558\uace0 \ucd5c\uc801\ud654\ub41c C++ \ubc31\uc5d4\ub4dc\ub85c \uc778\ud574 \ub192\uc740 \uc131\ub2a5\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. \ub300\uaddc\ubaa8 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \ucd5c\uc801\ud654\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
3.4 \ub2e4\uc591\ud55c \uc751\uc6a9 \ubd84\uc57c<\/h3>\n
\ud30c\uc774\ud1a0\uce58\ub294 \uc774\ubbf8\uc9c0 \uc778\uc2dd, \uc790\uc5f0\uc5b4 \ucc98\ub9ac, \ucd94\ucc9c \uc2dc\uc2a4\ud15c \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \uc5f0\uad6c \ub17c\ubb38\uc5d0\uc11c\ub3c4 \uad11\ubc94\uc704\ud558\uac8c \ucc44\ud0dd\ub418\uace0 \uc788\ub294 \ud504\ub808\uc784\uc6cc\ud06c\uc785\ub2c8\ub2e4.<\/p>\n
4. \ud30c\uc774\ud1a0\uce58 \uc124\uce58 \ubc29\ubc95<\/h2>\n
\ud30c\uc774\ud1a0\uce58\ub294 pip \ub610\ub294 conda\ub97c \ud1b5\ud574 \uc27d\uac8c \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc544\ub798\ub294 pip\ub97c \uc0ac\uc6a9\ud55c \uc124\uce58 \ubc29\ubc95\uc785\ub2c8\ub2e4.<\/p>\n
pip install torch torchvision torchaudio\n    <\/code><\/pre>\n
5. \ud30c\uc774\ud1a0\uce58 \uc2e4\uc2b5 \uc608\uc81c: \uac04\ub2e8\ud55c MNIST \ud074\ub798\uc2dc\ud53c\uc5b4 \ub9cc\ub4e4\uae30<\/h2>\n
\uc774\uc81c \uac04\ub2e8\ud55c MNIST \uc190\uae00\uc528 \ub370\uc774\ud130\uc14b\uc744 \ubd84\ub958\ud558\ub294 \ubaa8\ub378\uc744 \ub9cc\ub4e4\uc5b4 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
5.1 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
import torch\nimport torchvision.transforms as transforms\nfrom torchvision import datasets\nfrom torch.utils.data import DataLoader\n\n# \ub370\uc774\ud130 \ubcc0\ud658 \ubc0f \ub85c\ub4dc\ntransform = transforms.Compose([transforms.ToTensor()])\ntrain_dataset = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = DataLoader(dataset=train_dataset, batch_size=64, shuffle=True)\n    <\/code><\/pre>\n
5.2 \ubaa8\ub378 \uc815\uc758<\/h3>\n
class MNISTClassifier(nn.Module):\n    def __init__(self):\n        super(MNISTClassifier, self).__init__()\n        self.flatten = nn.Flatten()\n        self.fc1 = nn.Linear(28 * 28, 128)\n        self.fc2 = nn.Linear(128, 10)\n\n    def forward(self, x):\n        x = self.flatten(x)\n        x = torch.relu(self.fc1(x))\n        x = self.fc2(x)\n        return x\n\nmodel = MNISTClassifier()\n    <\/code><\/pre>\n
5.3 \uc190\uc2e4 \ud568\uc218 \ubc0f \uc635\ud2f0\ub9c8\uc774\uc800 \uc815\uc758<\/h3>\n
criterion = nn.CrossEntropyLoss()\noptimizer = torch.optim.SGD(model.parameters(), lr=0.01)\n    <\/code><\/pre>\n
5.4 \ubaa8\ub378 \ud559\uc2b5<\/h3>\n
num_epochs = 5\nfor epoch in range(num_epochs):\n    for images, labels in train_loader:\n        # Forward pass\n        outputs = model(images)\n        loss = criterion(outputs, labels)\n        \n        # Backward pass and optimization\n        optimizer.zero_grad()\n        loss.backward()\n        optimizer.step()\n    \n    print(f'Epoch [{epoch+1}\/{num_epochs}], Loss: {loss.item():.4f}')\n    <\/code><\/pre>\n
5.5 \ubaa8\ub378 \ud3c9\uac00<\/h3>\n
\ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\uae30 \uc704\ud574 \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
test_dataset = datasets.MNIST(root='.\/data', train=False, download=True, transform=transform)\ntest_loader = DataLoader(dataset=test_dataset, batch_size=64, shuffle=False)\n\ncorrect = 0\ntotal = 0\nwith torch.no_grad():\n    for images, labels in test_loader:\n        outputs = model(images)\n        _, predicted = torch.max(outputs.data, 1)\n        total += labels.size(0)\n        correct += (predicted == labels).sum().item()\n\nprint(f'Accuracy of the model on the test images: {100 * correct \/ total:.2f}%')\n    <\/code><\/pre>\n
6. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \ud3ec\uc2a4\ud305\uc744 \ud1b5\ud574 PyTorch\uc758 \uc8fc\uc694 \ud2b9\uc9d5\uacfc \uc7a5\uc810, \uac04\ub2e8\ud55c MNIST \ubaa8\ub378\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. PyTorch\uc758 \uc720\uc5f0\uc131\uacfc \uc9c1\uad00\uc801\uc778 API\ub294 \ub370\uc774\ud130 \uacfc\ud559\uc790\uc640 \uc5f0\uad6c\uc790\ub4e4\uc774 \uc2e4\ud5d8\ud558\uace0 \uc5f0\uad6c\ud558\ub294 \ub370 \ud070 \ub3c4\uc6c0\uc744 \uc90d\ub2c8\ub2e4. \uc55e\uc73c\ub85c\uc758 \ub525\ub7ec\ub2dd \uc5f0\uad6c\ub098 \ud504\ub85c\uc81d\ud2b8\uc5d0\uc11c PyTorch\ub97c \ud65c\uc6a9\ud574\ubcf4\uae30\ub97c \uad8c\uc7a5\ud569\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0 \uc790\ub8cc<\/h2>\n