{"id":30152,"date":"2024-10-28T03:19:50","date_gmt":"2024-10-28T03:19:50","guid":{"rendered":"http:\/\/atmokpo.com\/w\/?p=30152"},"modified":"2024-11-26T06:49:41","modified_gmt":"2024-11-26T06:49:41","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%95%a9%ec%84%b1%ea%b3%b1%ec%b8%b5%ec%9d%98-%ed%95%84%ec%9a%94%ec%84%b1","status":"publish","type":"post","link":"https:\/\/atmokpo.com\/w\/30152\/","title":{"rendered":"\ub525\ub7ec\ub2dd \ud30c\uc774\ud1a0\uce58 \uac15\uc88c, \ud569\uc131\uacf1\uce35\uc758 \ud544\uc694\uc131"},"content":{"rendered":"

<\/p>\n

\ub525\ub7ec\ub2dd\uc740 \uba38\uc2e0\ub7ec\ub2dd\uc758 \ud55c \ubd84\uc57c\ub85c, \uc5ec\ub7ec \uce35\uc758 \ub274\ub7f0\uc744 \ud1b5\ud574 \ub370\uc774\ud130\uc5d0\uc11c \ud328\ud134\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4. \ub2e4\uc591\ud55c \ub525\ub7ec\ub2dd \ubaa8\ub378\uc774 \uc874\uc7ac\ud558\uc9c0\ub9cc, \uadf8 \uc911\uc5d0\uc11c\ub3c4 \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd(Convolutional Neural Networks, CNN)<\/strong>\uc740 \uc8fc\ub85c \uc774\ubbf8\uc9c0 \ucc98\ub9ac\uc5d0 \uc801\ud569\ud55c \uc544\ud0a4\ud14d\ucc98\uc785\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 \ud569\uc131\uacf1\uce35\uc758 \ud544\uc694\uc131\uacfc \uc791\ub3d9 \uc6d0\ub9ac\ub97c \uc774\ud574\ud558\uace0, \uc774\ub97c \ud30c\uc774\ud1a0\uce58(PyTorch)\ub97c \uc0ac\uc6a9\ud558\uc5ec \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n

1. \ud569\uc131\uacf1\uce35\uc758 \uac1c\ub150<\/h2>\n

\ud569\uc131\uacf1\uce35\uc740 \uc774\ubbf8\uc9c0\uc758 \ud2b9\uc9d5\uc744 \ucd94\ucd9c\ud558\uae30 \uc704\ud574 \uc124\uacc4\ub41c \uce35\uc73c\ub85c, \uc77c\ubc18\uc801\uc778 \uc804\uacb0\ud569\uce35(Fully Connected Layer)\uacfc\ub294 \ub2e4\ub978 \ubc29\uc2dd\uc73c\ub85c \uc791\ub3d9\ud569\ub2c8\ub2e4. \ud569\uc131\uacf1\uce35\uc740 \uc785\ub825 \uc774\ubbf8\uc9c0\uc5d0 \ub300\ud574 \ucee4\ub110(kernel) \ub610\ub294 \ud544\ud130(filter)\ub77c \ubd88\ub9ac\ub294 \ub9e4\uac1c\ubcc0\uc218\ub97c \uc0ac\uc6a9\ud558\uc5ec, \uc774\ubbf8\uc9c0\uc758 \uc9c0\uc5ed\uc801\uc778 \ud2b9\uc9d5\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc5d0\uc11c\ub294 \uc9c0\uc5ed\uc801\uc778 \uc601\uc5ed\uc744 \ubd84\uc11d\ud558\uc5ec \ud544\ud130\uc640\uc758 \ud569\uc131\uacf1(convolution)\uc744 \uc218\ud589\ud569\ub2c8\ub2e4.<\/p>\n

1.1. \ud569\uc131\uacf1 \uc5f0\uc0b0<\/h3>\n

\ud569\uc131\uacf1 \uc5f0\uc0b0\uc740 \uc785\ub825 \uc774\ubbf8\uc9c0\uc5d0 \ub300\ud574 \ucee4\ub110\uc744 \uc2ac\ub77c\uc774\ub529\ud558\uc5ec \uc9c0\uc5ed\uc801\uc778 \ud53c\uccd0\ub97c \ucd94\ucd9c\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4. \uad6c\uccb4\uc801\uc73c\ub85c, \ucee4\ub110\uc774 \uc774\ubbf8\uc9c0\uc758 \ud2b9\uc815 \uc601\uc5ed\uc5d0 \uc704\uce58\ud560 \ub54c, \ud574\ub2f9 \uc601\uc5ed\uc758 \ud53d\uc140 \uac12\uacfc \ucee4\ub110\uc758 \uac12\ub4e4\uc744 \uacf1\ud558\uace0, \uadf8 \uacb0\uacfc\ub97c \ubaa8\ub450 \ub354\ud558\uc5ec \uc0c8\ub85c\uc6b4 Pixel Value\ub97c \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/p>\n

1.2. \ud480\ub9c1 \ub808\uc774\uc5b4(Pooling Layer)<\/h3>\n

\ud569\uc131\uacf1 \uc5f0\uc0b0 \ud6c4, \ud480\ub9c1 \ub808\uc774\uc5b4\ub97c \uc0ac\uc6a9\ud558\uc5ec \ucc28\uc6d0\uc744 \ucd95\uc18c\ud558\uace0 \uacc4\uc0b0\ub7c9\uc744 \uc904\uc774\uba70, \ub178\uc774\uc988\uc5d0 \uac15\ud55c \ud2b9\uc9d5\uc744 \uc720\uc9c0\ud560 \uc218 \uc788\uac8c \ub429\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc73c\ub85c \ucd5c\ub300 \ud480\ub9c1(Max Pooling)\uc774\ub098 \ud3c9\uade0 \ud480\ub9c1(Average Pooling)\uc744 \uc0ac\uc6a9\ud569\ub2c8\ub2e4. \ud480\ub9c1\uc740 \uc774\ubbf8\uc9c0\uc758 \ud2b9\uc815 \ud2b9\uc9d5\uc744 \uac15\uc870\ud558\uace0, \uc704\uce58 \ubd88\ubcc0\uc131\uc744 \ub354\uc6b1 \uac15\ud654\ud569\ub2c8\ub2e4.<\/p>\n

2. \ud569\uc131\uacf1\uce35\uc758 \ud544\uc694\uc131<\/h2>\n

2.1. \ud30c\ub77c\ubbf8\ud130 \uc218 \uac10\uc18c<\/h3>\n

\uc804\uacb0\ud569\uce35\uc740 \ubaa8\ub4e0 \uc785\ub825 \ub178\ub4dc\uac00 \ubaa8\ub4e0 \ucd9c\ub825 \ub178\ub4dc\uc640 \uc5f0\uacb0\ub418\ubbc0\ub85c, \ub9e4\uac1c\ubcc0\uc218\uc758 \uc218\uac00 \uae09\uaca9\ud788 \uc99d\uac00\ud569\ub2c8\ub2e4. \ubc18\uba74, \ud569\uc131\uacf1\uce35\uc740 \ucee4\ub110\uc758 \ud06c\uae30(\uc608: 3×3)\ub9cc\ud07c\uc758 \ub9e4\uac1c\ubcc0\uc218\ub9cc \ud544\uc694\ud558\ubbc0\ub85c, \uc804\uacb0\ud569\uce35\uc5d0 \ube44\ud574 \ud6e8\uc52c \uc801\uc740 \uc218\uc758 \ud30c\ub77c\ubbf8\ud130\ub85c\ub3c4 \ud6a8\uacfc\uc801\uc73c\ub85c \ud2b9\uc9d5\uc744 \ucd94\ucd9c\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n

2.2. \uc9c0\uc5ed\uc801\uc778 \ud2b9\uc9d5 \ucd94\ucd9c<\/h3>\n

\uc774\ubbf8\uc9c0\ub294 \uc8fc\ub85c \uc9c0\uc5ed\uc801\uc778 \uc815\ubcf4\ub97c \uc9c0\ub2c8\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uc774\ubbf8\uc9c0\uc758 \uc5b4\ub5a4 \uad6d\uc18c \uc601\uc5ed\uc774 \ud2b9\uc9d5 \uac1d\uccb4\ub97c \ud3ec\ud568\ud558\uace0 \uc788\uc744 \uacbd\uc6b0, \ud574\ub2f9 \uc601\uc5ed\uc758 \ud2b9\uc9d5\uc744 \uc798 \ucd94\ucd9c\ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4. \ud569\uc131\uacf1\uce35\uc740 \uc774\ub7ec\ud55c \uc9c0\uc5ed\uc801\uc778 \ud328\ud134\uc744 \uc798 \ud559\uc2b5\ud558\uc5ec \uc815\ud655\ud55c \uc608\uce21\uc744 \uac00\ub2a5\ud558\uac8c \ud569\ub2c8\ub2e4.<\/p>\n

2.3. \uc704\uce58 \ubd88\ubcc0\uc131<\/h3>\n

\ud569\uc131\uacf1\uce35\uacfc \ud480\ub9c1\uce35\uc744 \ud1b5\ud574 \ud559\uc2b5\ub41c \ud2b9\uc9d5\ub4e4\uc740 \uc774\ubbf8\uc9c0 \ub0b4\uc5d0\uc11c\uc758 \uc704\uce58\uc5d0 \ub3c5\ub9bd\uc801\uc785\ub2c8\ub2e4. \uc989, \uc774\ubbf8\uc9c0\uc758 \uac1d\uccb4\uac00 \uc5b4\ub514\uc5d0 \uc704\uce58\ud558\ub354\ub77c\ub3c4 \ud574\ub2f9 \ud2b9\uc9d5\uc744 \uc798 \uc778\uc2dd\ud560 \uc218 \uc788\uac8c \ub429\ub2c8\ub2e4. \uc774\ub294 \uc774\ubbf8\uc9c0 \ubd84\ub958\uc640 \uac19\uc740 \uc791\uc5c5\uc5d0\uc11c \ud070 \uc7a5\uc810\uc774 \ub429\ub2c8\ub2e4.<\/p>\n

2.4. \ub2e4\uc591\ud55c \uc751\uc6a9 \ubd84\uc57c<\/h3>\n

\ud569\uc131\uacf1\uce35\uc740 \uc774\ubbf8\uc9c0 \ubd84\ub958, \uac1d\uccb4 \ud0d0\uc9c0, \uc774\ubbf8\uc9c0 \uc0dd\uc131, \uc2ec\uc9c0\uc5b4 \uc790\uc5f0\uc5b4 \ucc98\ub9ac\uc640 \uac19\uc740 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0 \uac78\uccd0 \uc751\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc778\uacf5\uc9c0\ub2a5\uc758 \uae09\uc18d\ud55c \ubc1c\uc804\uc5d0\ub3c4 \ubd88\uad6c\ud558\uace0 CNN\uc758 \uae30\ubcf8 \uad6c\uc870\ub294 \uc5ec\uc804\ud788 \ub9ce\uc740 \ucd5c\uc2e0 \ubaa8\ub378\uc758 \ud575\uc2ec \uc694\uc18c\ub85c \ub0a8\uc544\uc788\uc2b5\ub2c8\ub2e4.<\/p>\n

3. \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c\uc758 \ud569\uc131\uacf1\uce35 \uad6c\ud604<\/h2>\n

\uc774\uc81c \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c CNN\uc744 \uad6c\ud604\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \uae30\ubcf8\uc801\uc778 \ud569\uc131\uacf1\uce35, \ud480\ub9c1\uce35, \uadf8\ub9ac\uace0 \uc804\uacb0\ud569\uce35\uc744 \ud3ec\ud568\ud55c CNN \ubaa8\ub378\uc758 \uc608\uc785\ub2c8\ub2e4.<\/p>\n

import torch\nimport torch.nn as nn\nimport torch.nn.functional as F\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets\nimport torch.optim as optim\n\n# CNN \ubaa8\ub378 \uc815\uc758\nclass SimpleCNN(nn.Module):\n    def __init__(self):\n        super(SimpleCNN, self).__init__()\n        self.conv1 = nn.Conv2d(in_channels=1, out_channels=32, kernel_size=3, stride=1, padding=1)  # \ud569\uc131\uacf1\uce35\n        self.pool = nn.MaxPool2d(kernel_size=2, stride=2, padding=0)  # \ud480\ub9c1\uce35\n        self.conv2 = nn.Conv2d(in_channels=32, out_channels=64, kernel_size=3, stride=1, padding=1)\n        self.fc1 = nn.Linear(64 * 7 * 7, 128)  # \uc804\uacb0\ud569\uce35\n        self.fc2 = nn.Linear(128, 10)  # \ucd9c\ub825\uce35\n\n    def forward(self, x):\n        x = self.pool(F.relu(self.conv1(x)))  # \uccab \ubc88\uc9f8 \ud569\uc131\uacf1 + \ud480\ub9c1\n        x = self.pool(F.relu(self.conv2(x)))  # \ub450 \ubc88\uc9f8 \ud569\uc131\uacf1 + \ud480\ub9c1\n        x = x.view(-1, 64 * 7 * 7)  # Flatten\n        x = F.relu(self.fc1(x))  # \uccab \ubc88\uc9f8 \uc804\uacb0\ud569\uce35\n        x = self.fc2(x)  # \ucd9c\ub825\uce35\n        return x\n\n# \ub370\uc774\ud130\uc14b \ubd88\ub7ec\uc624\uae30 \ubc0f \uc804\ucc98\ub9ac\ntransform = transforms.Compose([\n    transforms.ToTensor(),\n    transforms.Normalize((0.5,), (0.5,))  # \uc815\uaddc\ud654\n])\ntrainset = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=64, shuffle=True)\n\n# \ubaa8\ub378, \uc190\uc2e4 \ud568\uc218, \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998 \uc815\uc758\nmodel = SimpleCNN()\ncriterion = nn.CrossEntropyLoss()  # \uc190\uc2e4 \ud568\uc218\noptimizer = optim.Adam(model.parameters(), lr=0.001)  # \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998\n\n# \ud6c8\ub828 \ub8e8\ud504\nfor epoch in range(10):  # 10 epochs\n    for inputs, labels in trainloader:\n        optimizer.zero_grad()  # Gradient \ucd08\uae30\ud654\n        outputs = model(inputs)  # Forward pass\n        loss = criterion(outputs, labels)  # \uc190\uc2e4 \uacc4\uc0b0\n        loss.backward()  # Backward pass\n        optimizer.step()  # Weight \uc5c5\ub370\uc774\ud2b8\n\n    print(f'Epoch [{epoch+1}\/10], Loss: {loss.item():.4f}')\n\nprint('\ud6c8\ub828 \uc644\ub8cc!')<\/code><\/pre>\n
3.1. \ucf54\ub4dc \uc124\uba85<\/h3>\n
\uc704 \ucf54\ub4dc\uc5d0\uc11c\ub294 SimpleCNN<\/code> \ud074\ub798\uc2a4\ub97c \uc815\uc758\ud558\uace0, \ub450 \uac1c\uc758 \ud569\uc131\uacf1\uce35\uacfc \ub450 \uac1c\uc758 \uc804\uacb0\ud569\uce35\uc73c\ub85c \uad6c\uc131\ub41c CNN \ubaa8\ub378\uc744 \uc124\uacc4\ud588\uc2b5\ub2c8\ub2e4. torch.nn.Conv2d<\/code>\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud569\uc131\uacf1\uce35\uc744 \uc815\uc758\ud558\uc600\uace0, torch.nn.MaxPool2d<\/code>\ub97c \ud1b5\ud574 \ud480\ub9c1\uce35\uc744 \uc124\uc815\ud588\uc2b5\ub2c8\ub2e4. \ud6c8\ub828 \uacfc\uc815\uc5d0\uc11c\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc600\uace0, 10 epoch \ub3d9\uc548 \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ucf30\uc2b5\ub2c8\ub2e4.<\/p>\n
4. \uacb0\ub860<\/h2>\n
\ud569\uc131\uacf1\uce35\uc740 \uc774\ubbf8\uc9c0 \ub370\uc774\ud130\uc5d0\uc11c \uc911\uc694\ud55c \ud2b9\uc9d5\uc744 \ud6a8\uacfc\uc801\uc73c\ub85c \ucd94\ucd9c\ud558\ub294 \ub370 \ub9e4\uc6b0 \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4. \uae30\ubcf8\uc801\uc778 \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd\uc758 \uad6c\uc870\uc640 \ub3d9\uc791 \uc6d0\ub9ac\ub97c \uc774\ud574\ud558\ub294 \uac83\uc740 \ub525\ub7ec\ub2dd \ubd84\uc57c\uc5d0\uc11c \uc911\uc694\ud569\ub2c8\ub2e4. \uc774 \uae00\uc744 \ud1b5\ud574 \ud569\uc131\uacf1\uce35\uc758 \ud544\uc694\uc131\uacfc \uadf8 \uae30\ub2a5, \uadf8\ub9ac\uace0 \ud30c\uc774\ud1a0\uce58\ub97c \uc774\uc6a9\ud55c \uac04\ub2e8\ud55c \uad6c\ud604 \uc608\uc81c\ub97c \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub354 \ubcf5\uc7a1\ud55c CNN \uc544\ud0a4\ud14d\ucc98\uc640 \ub2e4\uc591\ud55c \uc751\uc6a9 \ubd84\uc57c\ub97c \ud0d0\uad6c\ud574 \ub098\uac00\uae38 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n
5. \ucc38\uace0 \uc790\ub8cc<\/h2>\n
\uc774 \uac15\uc88c\uc5d0\uc11c \ucc38\uace0\ud55c \uc790\ub8cc\ub294 \uc544\ub798\uc640 \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n