\ub525\ub7ec\ub2dd \ubd84\uc57c\uc5d0\uc11c Residual Network, \uc904\uc5ec\uc11c ResNet\uc740 \ub9e4\uc6b0 \uc911\uc694\ud55c \uc544\ud0a4\ud14d\ucc98\ub85c \uc790\ub9ac \uc7a1\uc558\uc2b5\ub2c8\ub2e4. ResNet\uc740 2015\ub144 Kaiming He\uc5d0 \uc758\ud574 \uc81c\uc548\ub418\uc5c8\uc73c\uba70, \ud2b9\ud788 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \uae4a\uc774\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \uc99d\uac00\uc2dc\ud0ac \uc218 \uc788\ub294 \ubc29\ubc95\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. \ud604\ub300\uc758 \ub2e4\uc591\ud55c \ucef4\ud4e8\ud130 \ube44\uc804 \ubb38\uc81c\ub4e4\uc5d0\uc11c ResNet\uc740 \uc131\ub2a5 \ud5a5\uc0c1\uc758 \uc8fc\uc694 \uc6d0\uc778 \uc911 \ud558\ub098\ub85c \uaf3d\ud799\ub2c8\ub2e4.<\/p>\n
ResNet\uc740 “Residual Learning” \ud504\ub808\uc784\uc6cc\ud06c\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \uc2e0\uacbd\ub9dd\uc785\ub2c8\ub2e4. \uc804\ud1b5\uc801\uc73c\ub85c, \uc2ec\uce35 \uc2e0\uacbd\ub9dd(dnn)\uc740 \ub354 \uae4a\uc5b4\uc9c8\uc218\ub85d \uc131\ub2a5\uc758 \uc800\ud558\uac00 \ubc1c\uc0dd\ud558\ub294 \uacbd\ud5a5\uc774 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 \uc8fc\ub85c \uae30\uc6b8\uae30 \uc18c\uc2e4(vanishing gradient) \ubb38\uc81c \ub54c\ubb38\uc778\ub370, \uc774 \ubb38\uc81c\ub294 \uc2e0\uacbd\ub9dd\uc758 \uae4a\uc774\uac00 \uae4a\uc5b4\uc9c8\uc218\ub85d \uc5ed\uc804\ud30c \uacfc\uc815\uc5d0\uc11c \uae30\uc6b8\uae30\uac00 \uc18c\uc2e4\ub418\uc5b4 \uac00\ub294 \ud604\uc0c1\uc785\ub2c8\ub2e4.<\/p>\n
ResNet\uc740 \uc774\ub7ec\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\uae30 \uc704\ud574 \uc794\ucc28 \uc5f0\uacb0(residual connection)\uc744 \ub3c4\uc785\ud558\uc600\uc2b5\ub2c8\ub2e4. \uc794\ucc28 \uc5f0\uacb0\uc740 \ub124\ud2b8\uc6cc\ud06c\uc758 \uc785\ub825\uc744 \ucd9c\ub825\uc5d0 \ub354\ud568\uc73c\ub85c\uc368 \ud55c \uacc4\uce35\uc5d0\uc11c \uc774\uc804 \uacc4\uce35\uc758 \uc815\ubcf4\ub97c \uc9c1\uc811 \uc804\ub2ec\ud569\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubc29\uc2dd\uc744 \ud1b5\ud574 \ub354 \uae4a\uc740 \ub124\ud2b8\uc6cc\ud06c\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \ud559\uc2b5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
ResNet\uc740 \ub2e4\uc591\ud55c \uae4a\uc774\ub97c \uac00\uc9c4 \ubaa8\ub378\ub85c \uad6c\uc131\ub420 \uc218 \uc788\uc73c\uba70, \uc77c\ubc18\uc801\uc73c\ub85c “ResNet50”, “ResNet101”, “ResNet152″\uc640 \uac19\uc740 \uc2dd\uc73c\ub85c \ud45c\uae30\ub429\ub2c8\ub2e4. \uc774 \uc22b\uc790\ub294 \ub124\ud2b8\uc6cc\ud06c\uc758 \ucd1d \uce35 \uc218\ub97c \uc758\ubbf8\ud569\ub2c8\ub2e4.<\/p>\n
ResNet\uc758 \uae30\ubcf8 \uad6c\uc131 \uc694\uc18c\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ube14\ub85d\uc73c\ub85c \uc774\ub8e8\uc5b4\uc838 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\uc77c\ubc18\uc801\uc778 ResNet \ube14\ub85d\uc758 \uad6c\uc870\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\ndef resnet_block(input_tensor, filters, kernel_size=3, stride=1):\n x = Conv2D(filters, kernel_size=kernel_size, strides=stride, padding='same')(input_tensor)\n x = BatchNormalization()(x)\n x = ReLU()(x)\n x = Conv2D(filters, kernel_size=kernel_size, strides=stride, padding='same')(x)\n x = BatchNormalization()(x)\n \n shortcut = Conv2D(filters, kernel_size=1, strides=stride, padding='same')(input_tensor)\n x = Add()([x, shortcut])\n x = ReLU()(x)\n \n return x\n<\/code><\/pre>\n3. PyTorch\ub97c \uc774\uc6a9\ud55c ResNet \uad6c\ud604<\/h2>\n
\uc774\uc81c \ud30c\uc774\ud1a0\uce58(Pytorch)\ub97c \uc0ac\uc6a9\ud558\uc5ec ResNet\uc744 \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud569\ub2c8\ub2e4:<\/p>\n
pip install torch torchvision<\/code><\/pre>\n\uc774\ud6c4, \ub2e4\uc74c\uc73c\ub85c \uae30\ubcf8 ResNet \ubaa8\ub378\uc744 \uad6c\ud604\ud569\ub2c8\ub2e4:<\/p>\n
\nimport torch\nimport torch.nn as nn\nimport torchvision.transforms as transforms\nimport torchvision.datasets as datasets\nimport torchvision.models as models\n\nclass BasicBlock(nn.Module):\n expansion = 1\n\n def __init__(self, in_channels, out_channels, stride=1, downsample=None):\n super(BasicBlock, self).__init__()\n self.conv1 = nn.Conv2d(in_channels, out_channels, kernel_size=3, stride=stride, padding=1, bias=False)\n self.bn1 = nn.BatchNorm2d(out_channels)\n self.relu = nn.ReLU(inplace=True)\n self.conv2 = nn.Conv2d(out_channels, out_channels, kernel_size=3, stride=1, padding=1, bias=False)\n self.bn2 = nn.BatchNorm2d(out_channels)\n self.downsample = downsample\n\n def forward(self, x):\n identity = x\n out = self.conv1(x)\n out = self.bn1(out)\n out = self.relu(out)\n out = self.conv2(out)\n out = self.bn2(out)\n\n if self.downsample is not None:\n identity = self.downsample(x)\n\n out += identity\n out = self.relu(out)\n\n return out\n\nclass ResNet(nn.Module):\n def __init__(self, block, layers, num_classes=1000):\n super(ResNet, self).__init__()\n self.in_channels = 64\n self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3, bias=False)\n self.bn1 = nn.BatchNorm2d(64)\n self.relu = nn.ReLU(inplace=True)\n self.maxpool = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)\n self.layer1 = self._make_layer(block, 64, layers[0])\n self.layer2 = self._make_layer(block, 128, layers[1], stride=2)\n self.layer3 = self._make_layer(block, 256, layers[2], stride=2)\n self.layer4 = self._make_layer(block, 512, layers[3], stride=2)\n self.avgpool = nn.AdaptiveAvgPool2d((1, 1))\n self.fc = nn.Linear(512 * block.expansion, num_classes)\n\n def _make_layer(self, block, out_channels, blocks, stride=1):\n downsample = None\n if stride != 1 or self.in_channels != out_channels * block.expansion:\n downsample = nn.Sequential(\n nn.Conv2d(self.in_channels, out_channels * block.expansion, kernel_size=1, stride=stride, bias=False),\n nn.BatchNorm2d(out_channels * block.expansion),\n )\n layers = []\n layers.append(block(self.in_channels, out_channels, stride, downsample))\n self.in_channels = out_channels * block.expansion\n for _ in range(1, blocks):\n layers.append(block(self.in_channels, out_channels))\n\n return nn.Sequential(*layers)\n\n def forward(self, x):\n x = self.conv1(x)\n x = self.bn1(x)\n x = self.relu(x)\n x = self.maxpool(x)\n\n x = self.layer1(x)\n x = self.layer2(x)\n x = self.layer3(x)\n x = self.layer4(x)\n\n x = self.avgpool(x)\n x = torch.flatten(x, 1)\n x = self.fc(x)\n\n return x\n\ndef resnet18(num_classes=1000):\n return ResNet(BasicBlock, [2, 2, 2, 2], num_classes)\n<\/code><\/pre>\n3.1 \ubaa8\ub378 \ud6c8\ub828 \uc900\ube44\ud558\uae30<\/h3>\n
ResNet \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uae30 \uc704\ud574 \ub370\uc774\ud130\uc14b\uc744 \uc900\ube44\ud558\uace0, \uc635\ud2f0\ub9c8\uc774\uc800\uc640 \uc190\uc2e4 \ud568\uc218\ub97c \uc124\uc815\ud569\ub2c8\ub2e4.<\/p>\n
\n# \ub370\uc774\ud130\uc14b \uc900\ube44\ntransform = transforms.Compose([\n transforms.Resize((224, 224)),\n transforms.ToTensor(),\n])\n\ntrain_dataset = datasets.CIFAR10(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = torch.utils.data.DataLoader(train_dataset, batch_size=64, shuffle=True)\n\n# \ubaa8\ub378 \ucd08\uae30\ud654\nmodel = resnet18(num_classes=10)\ncriterion = nn.CrossEntropyLoss()\noptimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n<\/code><\/pre>\n3.2 \ud6c8\ub828 \ub2e8\uacc4<\/h3>\n
\uc774\uc81c \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0ac \uc900\ube44\uac00 \ub418\uc5c8\uc2b5\ub2c8\ub2e4:<\/p>\n
\nfor epoch in range(10): # epochs \uc124\uc815\n model.train() # \ubaa8\ub378\uc744 \ud6c8\ub828 \ubaa8\ub4dc\ub85c \uc804\ud658\n for images, labels in train_loader:\n optimizer.zero_grad() # \uae30\uc6b8\uae30 \ucd08\uae30\ud654\n outputs = model(images) # \ubaa8\ub378 \uc608\uce21\n loss = criterion(outputs, labels) # \uc190\uc2e4 \uacc4\uc0b0\n loss.backward() # \uc5ed\uc804\ud30c\n optimizer.step() # \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8\n\n print(f'Epoch [{epoch+1}\/10], Loss: {loss.item():.4f}')\n<\/code><\/pre>\n4. ResNet\uc758 \ud65c\uc6a9<\/h2>\n
ResNet\uc740 \ub2e4\uc591\ud55c \ucef4\ud4e8\ud130 \ube44\uc804 \ud0dc\uc2a4\ud06c\uc5d0\uc11c \uc0ac\uc6a9 \uac00\ub2a5\ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uc774\ubbf8\uc9c0 \ubd84\ub958, \uac1d\uccb4 \ud0d0\uc9c0, \uc138\ubd84\ud654, \uadf8\ub9ac\uace0 \ub354 \ubcf5\uc7a1\ud55c \ube44\uc804 \ubb38\uc81c\uc5d0 \uc774\ub974\uae30\uae4c\uc9c0 \ud3ed\ub113\uac8c \ud65c\uc6a9\ub429\ub2c8\ub2e4. Google, Facebook \ub4f1\uc774 \uc0ac\uc6a9\ud558\ub294 \uc5ec\ub7ec \uc774\ubbf8\uc9c0 \ubc0f \ube44\ub514\uc624 \ud0dc\uc2a4\ud06c\uc5d0 ResNet \uc544\ud0a4\ud14d\ucc98\uac00 \ud3ec\ud568\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
5. \uacb0\ub860<\/h2>\n
\uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 ResNet\uc758 \uae30\ubcf8 \uac1c\ub150 \ubc0f \uc544\ud0a4\ud14d\ucc98\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uace0, \ud30c\uc774\ud1a0\uce58\ub97c \uc774\uc6a9\ud558\uc5ec \uae30\ubcf8 ResNet \ubaa8\ub378\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \ubc30\uc6cc\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \ub354 \uae4a\uac8c \uc313\uc744 \uc218 \uc788\ub294 \uc720\uc5f0\ud55c \ubc29\ubc95\uacfc \uc794\ucc28 \ud559\uc2b5\uc744 \ud65c\uc6a9\ud558\uc5ec \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \ub0bc \uc218 \uc788\ub294 \uae30\ud68c\ub97c \uc81c\uacf5\ud558\ub294 ResNet\uc740 \ub9ce\uc740 \uc5f0\uad6c\uc790\uc640 \uac1c\ubc1c\uc790\uc5d0\uac8c \uc601\uac10\uc744 \uc8fc\ub294 \uc8fc\uc694 \uc544\ud0a4\ud14d\ucc98\uc785\ub2c8\ub2e4.<\/p>\n
\uc774\uc81c \ub354 \uc2ec\ud654\ub41c ResNet \uad6c\uc870 \ubc0f \ub2e4\uc591\ud55c \ud30c\ub77c\ubbf8\ud130 \uc870\uc815, \ub370\uc774\ud130 \uc99d\uac15 \uae30\ubc95 \ub4f1\uc744 \ud1b5\ud55c \ubaa8\ub378 \uac1c\uc120\uc744 \uacf5\ubd80\ud574 \ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
6. \ucc38\uace0 \uc790\ub8cc<\/h2>\n