\ub525\ub7ec\ub2dd\uc740 \ucd5c\uadfc \uba87 \ub144\uac04 \ub370\uc774\ud130 \uacfc\ud559\uc758 \uc5ec\ub7ec \ubd84\uc57c\uc5d0\uc11c \uc5c4\uccad\ub09c \uc778\uae30\ub97c \ub04c\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc591\ud55c \ubd84\uc57c\uc758 \ubb38\uc81c\ub97c \ud574\uacb0\ud558\ub294 \ub370 \uc788\uc5b4 \ub9e4\uc6b0 \uc720\uc6a9\ud55c \ub3c4\uad6c\ub85c \uc790\ub9ac\uc7a1\uac8c \ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc774 \uac15\uc88c\uc5d0\uc11c\ub294 \uc720\uba85\ud55c \ub525\ub7ec\ub2dd \uc544\ud0a4\ud14d\ucc98 \uc911 \ud558\ub098\uc778 LeNet-5\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
LeNet-5\ub294 Yann LeCun\uc744 \ud3ec\ud568\ud55c \uc5f0\uad6c\uc790\ub4e4\uc774 1998\ub144\uc5d0 \uac1c\ubc1c\ud55c \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd(CNN) \uc544\ud0a4\ud14d\ucc98\uc785\ub2c8\ub2e4. \uc774\ubbf8\uc9c0\ub97c \uc778\uc2dd\ud558\ub294 \ub370 \uc720\uc6a9\ud55c \ubaa8\ub378\ub85c, \uc8fc\ub85c \uc190\uae00\uc528 \uc22b\uc790 \uc778\uc2dd\uc5d0 \uc0ac\uc6a9\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 CNN\uc758 \uae30\ubcf8 \uad6c\uc870\ub97c \ub530\ub974\uba70, \uc5ec\ub7ec \uacc4\uce35\uc73c\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. LeNet-5\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uacc4\uce35\uc73c\ub85c \uc774\ub8e8\uc5b4\uc838 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
LeNet-5\ub294 CNN\uc758 \uae30\ubcf8 \uc544\ud0a4\ud14d\ucc98 \uc911 \ud558\ub098\ub85c, \ub9ce\uc740 \uae4a\uc740 \ub124\ud2b8\uc6cc\ud06c\uc758 \uae30\ubc18\uc774 \ub418\uc5c8\uc2b5\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740 \uc774\ubbf8\uc9c0 \uc778\uc2dd \ubd84\uc57c\uc5d0\uc11c \ub9ce\uc740 \ud601\uc2e0\uc744 \uac00\uc838\uc654\uace0, \ud604\uc7ac\uc5d0\ub3c4 \ub2e4\uc591\ud55c \ubcc0\ud615 \ubaa8\ub378\uc774 \uc874\uc7ac\ud569\ub2c8\ub2e4. LeNet-5\uc758 \ub2e8\uc21c\uc131\uacfc \ud6a8\uc728\uc131 \ub355\ubd84\uc5d0, \ub9ce\uc740 \ub370\uc774\ud130\uc14b\uc5d0\uc11c \uc88b\uc740 \uc131\ub2a5\uc744 \ubc1c\ud718\ud569\ub2c8\ub2e4.<\/p>\n
\uc774\uc81c PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec LeNet-5\ub97c \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. PyTorch\ub294 \uc0ac\uc6a9\uc790 \uce5c\ud654\uc801\uc778 \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c\ub85c, \ub2e4\uc591\ud55c \uc5f0\uad6c\uc640 \uc0b0\uc5c5\uc5d0\uc11c \ub110\ub9ac \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \ub610\ud55c, PyTorch\ub294 \ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504\ub97c \uc0ac\uc6a9\ud558\ub294 \uc7a5\uc810\uc774 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uba3c\uc800, \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud558\uace0 \ud658\uacbd\uc744 \uc124\uc815\ud574\uc57c \ud569\ub2c8\ub2e4. \ub2e4\uc74c \ucf54\ub4dc\ub97c \uc0ac\uc6a9\ud558\uc5ec PyTorch\uc640 torchvision\uc744 \uc124\uce58\ud558\uc138\uc694:<\/p>\n
pip install torch torchvision<\/code><\/pre>\nLeNet-5 \ubaa8\ub378 \uad6c\ud604<\/h3>\n
\uc774\uc81c LeNet-5\uc758 \uad6c\uc870\ub97c \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4:<\/p>\n
import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass LeNet5(nn.Module):\n def __init__(self):\n super(LeNet5, self).__init__()\n self.conv1 = nn.Conv2d(1, 6, kernel_size=5)\n self.conv2 = nn.Conv2d(6, 16, kernel_size=5)\n self.fc1 = nn.Linear(16 * 5 * 5, 120)\n self.fc2 = nn.Linear(120, 84)\n self.fc3 = nn.Linear(84, 10)\n\n def forward(self, x):\n x = F.relu(self.conv1(x))\n x = F.avg_pool2d(x, kernel_size=2, stride=2)\n x = F.relu(self.conv2(x))\n x = F.avg_pool2d(x, kernel_size=2, stride=2)\n x = x.view(-1, 16 * 5 * 5)\n x = F.relu(self.fc1(x))\n x = F.relu(self.fc2(x))\n x = self.fc3(x)\n return x\n<\/code><\/pre>\n\ubaa8\ub378 \ud6c8\ub828 \uc704\ud55c \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
LeNet-5\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud6c8\ub828\ud560 \uac83\uc785\ub2c8\ub2e4. torchvision\uc744 \uc0ac\uc6a9\ud558\uc5ec \ub370\uc774\ud130\ub97c \uc27d\uac8c \ub2e4\uc6b4\ub85c\ub4dc\ud558\uace0 \ub85c\ub4dc\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c \ucf54\ub4dc\ub97c \uc0ac\uc6a9\ud558\uc5ec MNIST \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud558\uc138\uc694:<\/p>\n
from torchvision import datasets, transforms\n\ntransform = transforms.Compose([\n transforms.Grayscale(num_output_channels=1),\n transforms.Resize((32, 32)),\n transforms.ToTensor(),\n transforms.Normalize((0.5,), (0.5,))\n])\n\ntrain_dataset = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntest_dataset = datasets.MNIST(root='.\/data', train=False, download=True, transform=transform)\n\ntrain_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=64, shuffle=True)\ntest_loader = torch.utils.data.DataLoader(dataset=test_dataset, batch_size=64, shuffle=False)\n<\/code><\/pre>\n\ubaa8\ub378 \ud6c8\ub828<\/h3>\n
\ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\uae30 \uc704\ud574\uc11c\ub294 \uc190\uc2e4 \ud568\uc218\uc640 \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998\uc744 \uc124\uc815\ud574\uc57c \ud569\ub2c8\ub2e4. \uc5ec\uae30\uc11c\ub294 Cross Entropy Loss\uc640 Adam \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \uc0ac\uc6a9\ud558\uaca0\uc2b5\ub2c8\ub2e4:<\/p>\n
device = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\nmodel = LeNet5().to(device)\ncriterion = nn.CrossEntropyLoss()\noptimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n\nnum_epochs = 5\nfor epoch in range(num_epochs):\n for i, (images, labels) in enumerate(train_loader):\n images = images.to(device)\n labels = labels.to(device)\n\n # Forward pass\n outputs = model(images)\n loss = criterion(outputs, labels)\n\n # Backward and optimize\n optimizer.zero_grad()\n loss.backward()\n optimizer.step()\n\n if (i + 1) % 100 == 0:\n print(f'Epoch [{epoch + 1}\/{num_epochs}], Step [{i + 1}\/{len(train_loader)}], Loss: {loss.item():.4f}')\n<\/code><\/pre>\n\ubaa8\ub378 \ud3c9\uac00<\/h3>\n
\ud6c8\ub828\uc774 \uc644\ub8cc\ub41c \ud6c4, \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc815\ud655\ub3c4\ub97c \ud655\uc778\ud558\uaca0\uc2b5\ub2c8\ub2e4:<\/p>\n
model.eval()\nwith torch.no_grad():\n correct = 0\n total = 0\n for images, labels in test_loader:\n images = images.to(device)\n labels = labels.to(device)\n outputs = model(images)\n _, predicted = torch.max(outputs.data, 1)\n total += labels.size(0)\n correct += (predicted == labels).sum().item()\n\n print(f'Accuracy of the model on the test images: {100 * correct \/ total:.2f}%')\n<\/code><\/pre>\n\uacb0\ub860<\/h3>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 LeNet-5 \uc544\ud0a4\ud14d\ucc98\ub97c PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec \uad6c\ud604\ud558\uace0 \ud6c8\ub828\ud558\ub294 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. LeNet-5\ub294 CNN\uc758 \uae30\ucd08\ub97c \uc774\ud574\ud558\uace0 \uc2e4\uc2b5\ud560 \uc218 \uc788\ub294 \uc88b\uc740 \uc608\uc81c\uc785\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc744 \uae30\ubc18\uc73c\ub85c \ub354 \ubcf5\uc7a1\ud55c \ub124\ud2b8\uc6cc\ud06c \uc544\ud0a4\ud14d\ucc98\ub098 \ub2e4\uc591\ud55c \uc751\uc6a9\uc73c\ub85c \ubc1c\uc804\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c \ub2e8\uacc4\ub85c\ub294 \ub354 \uae4a\uc740 \ub124\ud2b8\uc6cc\ud06c \uad6c\uc870\ub098 \ub370\uc774\ud130\uc14b\uc744 \ud65c\uc6a9\ud574\ubcf4\ub294 \uac83\uc744 \ucd94\ucc9c\ub4dc\ub9bd\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0 \uc790\ub8cc<\/h3>\n