\n\ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \uc131\ub2a5 \ucd5c\uc801\ud654\ub294 \ud56d\uc0c1 \uc911\uc694\ud55c \uc8fc\uc81c\uc785\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 \ubc30\uce58 \uc815\uaddc\ud654(Batch Normalization)\ub97c \ud65c\uc6a9\ud558\uc5ec \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \uac1c\uc120\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \ubc30\uce58 \uc815\uaddc\ud654\ub294 \ud6c8\ub828 \uacfc\uc815\uc744 \uc548\uc815\ud654\ud558\uace0 \ud559\uc2b5 \uc18d\ub3c4\ub97c \uc99d\uac00\uc2dc\ud0a4\ub294 \ub370 \ub3c4\uc6c0\uc744 \uc90d\ub2c8\ub2e4. \uc774\uc5b4\uc11c \ubc30\uce58 \uc815\uaddc\ud654\ub97c \uc0ac\uc6a9\ud558\ub294 \uc774\uc720, \uadf8 \uc791\ub3d9 \uc6d0\ub9ac, \uadf8\ub9ac\uace0 \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c \uc774\ub97c \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n<\/p>\n
\n\ubc30\uce58 \uc815\uaddc\ud654\ub294 \ub0b4\ubd80 \uacf5\ubcc0\ub7c9 \uc804\uc774(Internal Covariate Shift) \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \uc81c\uc548\ub41c \uae30\ubc95\uc785\ub2c8\ub2e4. \ub0b4\ubd80 \uacf5\ubcc0\ub7c9 \uc804\uc774\ub294 \ud6c8\ub828 \uacfc\uc815\uc5d0\uc11c \ub124\ud2b8\uc6cc\ud06c\uc758 \uac01 \uce35\uc758 \ubd84\ud3ec\uac00 \ubcc0\ud574\ubc84\ub9ac\ub294 \ud604\uc0c1\uc744 \ub9d0\ud569\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubcc0\ud654\ub294 \uac01 \uce35\uc758 \uae30\uc6b8\uae30(Graident)\uac00 \ub2ec\ub77c\uc9c0\ub3c4\ub85d \ub9cc\ub4e4\uace0, \uc774\ub294 \ud6c8\ub828 \uc18d\ub3c4\ub97c \uc800\ud558\uc2dc\ud0ac \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n<\/p>\n
\n\ubc30\uce58 \uc815\uaddc\ud654\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \uacfc\uc815\uc73c\ub85c \uc774\ub8e8\uc5b4\uc9d1\ub2c8\ub2e4:<\/p>\n
\n\ubc30\uce58 \uc815\uaddc\ud654\ub294 \uc5ec\ub7ec \uac00\uc9c0 \uc774\uc810\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\n\ud30c\uc774\ud1a0\uce58\ub294 \ubc30\uce58 \uc815\uaddc\ud654\ub97c \uc27d\uac8c \uad6c\ud604\ud560 \uc218 \uc788\ub294 \ud568\uc218\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4. \ub2e4\uc74c \ucf54\ub4dc\ub294 \uae30\ubcf8 \uc2e0\uacbd\ub9dd \ubaa8\ub378\uc5d0\uc11c \ubc30\uce58 \uc815\uaddc\ud654\ub97c \uc801\uc6a9\ud558\ub294 \uc608\uc2dc\uc785\ub2c8\ub2e4.\n<\/p>\n
import torch\nimport torch.nn as nn\nimport torch.optim as optim\nimport torchvision.datasets as datasets\nimport torchvision.transforms as transforms\n\n# \uc2e0\uacbd\ub9dd \ubaa8\ub378\nclass SimpleCNN(nn.Module):\n def __init__(self):\n super(SimpleCNN, self).__init__()\n self.conv1 = nn.Conv2d(1, 32, kernel_size=3, stride=1, padding=1)\n self.bn1 = nn.BatchNorm2d(32) # \ubc30\uce58 \uc815\uaddc\ud654 \ucd94\uac00\n self.conv2 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)\n self.bn2 = nn.BatchNorm2d(64) # \ubc30\uce58 \uc815\uaddc\ud654 \ucd94\uac00\n self.fc1 = nn.Linear(64 * 7 * 7, 128)\n self.fc2 = nn.Linear(128, 10)\n\n def forward(self, x):\n x = self.conv1(x)\n x = self.bn1(x) # \ubc30\uce58 \uc815\uaddc\ud654 \uc801\uc6a9\n x = nn.ReLU()(x)\n x = self.conv2(x)\n x = self.bn2(x) # \ubc30\uce58 \uc815\uaddc\ud654 \uc801\uc6a9\n x = nn.ReLU()(x)\n x = x.view(-1, 64 * 7 * 7) # Flatten\n x = self.fc1(x)\n x = self.fc2(x)\n return x<\/code><\/pre>\n3.2 \ub370\uc774\ud130 \ub85c\ub529 \ubc0f \ubaa8\ub378 \ud6c8\ub828<\/h3>\n\n# \ub370\uc774\ud130\uc14b \ub85c\ub529\ntransform = transforms.Compose([\n transforms.ToTensor(),\n transforms.Normalize((0.5,), (0.5,))\n])\n\ntrain_dataset = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=64, shuffle=True)\n\n# \ubaa8\ub378 \ubc0f optimizer \ucd08\uae30\ud654\nmodel = SimpleCNN()\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.Adam(model.parameters(), lr=0.001)\n\n# \ubaa8\ub378 \ud6c8\ub828\nnum_epochs = 5\nfor epoch in range(num_epochs):\n for images, labels in train_loader:\n outputs = model(images)\n loss = criterion(outputs, labels)\n\n optimizer.zero_grad()\n loss.backward()\n optimizer.step()\n print(f'Epoch [{epoch+1}\/{num_epochs}], Loss: {loss.item():.4f}')<\/code><\/pre>\n\n\uc704 \ucf54\ub4dc\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c CNN \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4. \uc5ec\uae30\uc11c \ubc30\uce58 \uc815\uaddc\ud654\uac00 \uc5b4\ub5bb\uac8c \uc0ac\uc6a9\ub418\ub294\uc9c0 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n<\/p>\n
4. \uacb0\ub860<\/h2>\n
\n\ubc30\uce58 \uc815\uaddc\ud654\ub294 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \ud6c8\ub828\uc744 \uc548\uc815\ud654\ud558\uace0 \uac00\uc18d\ud654\ud558\ub294 \ub370 \ub9e4\uc6b0 \uc720\uc6a9\ud55c \uae30\ubc95\uc785\ub2c8\ub2e4. \ub2e4\uc591\ud55c \ubaa8\ub378 \uad6c\uc870\uc5d0\uc11c \uc801\uc6a9\ud560 \uc218 \uc788\uc73c\uba70, \ud2b9\ud788 \uae4a\uc740 \ub124\ud2b8\uc6cc\ud06c\uc5d0\uc11c \uadf8 \ud6a8\uacfc\uac00 \ub450\ub4dc\ub7ec\uc9d1\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c \ubc30\uce58 \uc815\uaddc\ud654\uc758 \uac1c\ub150\uacfc \uc774\ub97c \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c \uad6c\ud604\ud558\ub294 \uacfc\uc815\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub354 \ub098\uc740 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \ub9cc\ub4e4\uae30 \uc704\ud574 \ubc30\uce58 \uc815\uaddc\ud654\ub97c \uc801\uadf9\uc801\uc73c\ub85c \ud65c\uc6a9\ud574 \ubcf4\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4.\n<\/p>\n
\ub354 \ub9ce\uc740 \ub525\ub7ec\ub2dd \uac15\uc88c\uc640 \ud30c\uc774\ud1a0\uce58 \uad00\ub828 \uc790\ub8cc\ub97c \uc6d0\ud558\uc2e0\ub2e4\uba74, \uc800\ud76c \ube14\ub85c\uadf8\ub97c \ud1b5\ud574 \ucd5c\uc2e0 \uc815\ubcf4\ub97c \ud655\uc778\ud574\ubcf4\uc138\uc694!<\/p>\n
\ucc38\uace0 \ubb38\ud5cc<\/h2>\n