\ub525\ub7ec\ub2dd\uc740 \ud604\uc7ac \uc778\uacf5\uc9c0\ub2a5 \ubd84\uc57c\uc5d0\uc11c \uac00\uc7a5 \uc911\uc694\ud55c \uae30\uc220 \uc911 \ud558\ub098\ub85c \uc790\ub9ac \uc7a1\uace0 \uc788\uc73c\uba70, \uadf8\uc911\uc5d0\uc11c\ub3c4 \uc2e0\uacbd\ub9dd(Neural Networks)\uc740 \ub2e4\uc591\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud558\ub294 \ub370 \ub110\ub9ac \uc0ac\uc6a9\ub429\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 GoogLeNet\uc774\ub77c\ub294 CNN(Convolutional Neural Network)\uc744 \uc790\uc138\ud788 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. GoogLeNet\uc740 2014\ub144 ILSVRC(Imagenet Large Scale Visual Recognition Challenge)\uc5d0\uc11c \uc6b0\uc2b9\ud558\uba70 \ud070 \uc8fc\ubaa9\uc744 \ubc1b\uc558\uc2b5\ub2c8\ub2e4.<\/p>\n
GoogLeNet\uc740 ‘Inception v1’\uc774\ub77c\ub294 \uc774\ub984\uc73c\ub85c\ub3c4 \uc54c\ub824\uc838 \uc788\uc73c\uba70, \uc5ec\ub7ec \uac1c\uc758 \ud569\uc131\uacf1(Convolution) \ub808\uc774\uc5b4\uac00 \ud3ec\ud568\ub41c \ub3c5\ucc3d\uc801\uc778 \uad6c\uc870\ub97c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. GoogLeNet\uc758 \uc8fc\uc694 \ud2b9\uc9d5\uc740 \ub2e4\uc591\ud55c \ud06c\uae30\uc758 \ud544\ud130\ub97c \ub3d9\uc2dc\uc5d0 \uc0ac\uc6a9\ud558\uc5ec \uc774\ubbf8\uc9c0\ub97c \ucc98\ub9ac\ud558\ub294 ‘Inception \ubaa8\ub4c8’\uc785\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubc29\uc2dd\uc740 \ub124\ud2b8\uc6cc\ud06c\uac00 \ub354 \ub9ce\uc740 \uc815\ubcf4\ub97c \uc190\uc2e4 \uc5c6\uc774 \ud559\uc2b5\ud560 \uc218 \uc788\ub3c4\ub85d \ub3c4\uc640\uc90d\ub2c8\ub2e4.<\/p>\n
Inception \ubaa8\ub4c8\uc740 \uc5ec\ub7ec \uac1c\uc758 \ud544\ud130\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc11c\ub85c \ub2e4\ub978 \uc218\uc900\uc758 \uc138\ubd80 \uc0ac\ud56d\uc744 \ud3ec\ucc29\ud569\ub2c8\ub2e4. \uac01 \ubaa8\ub4c8\uc740 \ub2e4\uc74c\uacfc \uac19\uc774 \uad6c\uc131\ub429\ub2c8\ub2e4:<\/p>\n
\uc774 \ubaa8\ub4e0 \ucd9c\ub825\uc744 \uacb0\ud569\ud558\uc5ec \ub2e4\uc74c \ub808\uc774\uc5b4\ub85c \uc804\ub2ec\ud569\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \ub2e4\uc591\ud55c \uc2a4\ucf00\uc77c\uc758 \ud2b9\uc9d5\uc744 \ud68d\ub4dd\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc774\uc81c PyTorch\uc5d0\uc11c GoogLeNet\uc744 \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uba3c\uc800 PyTorch\uc640 \uae30\ud0c0 \ud544\uc218 \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4.<\/p>\n
pip install torch torchvision<\/code><\/pre>\n3.1 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\uc608\uc81c\uc5d0\uc11c\ub294 CIFAR-10 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. \uc774 \ub370\uc774\ud130\uc14b\uc740 10\uac1c\uc758 \ud074\ub798\uc2a4\ub85c \uc774\ub8e8\uc5b4\uc9c4 60,000\uac1c\uc758 \uc774\ubbf8\uc9c0\ub97c \ud3ec\ud568\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\nimport torch\nimport torchvision\nimport torchvision.transforms as transforms\n\n# \ub370\uc774\ud130 \ubcc0\ud658 \uc815\uc758\ntransform = transforms.Compose(\n [transforms.Resize((224, 224)),\n transforms.ToTensor()])\n\n# CIFAR-10 \ub370\uc774\ud130\uc14b \ub2e4\uc6b4\ub85c\ub4dc\ntrainset = torchvision.datasets.CIFAR10(root='.\/data', train=True,\n download=True, transform=transform)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=32,\n shuffle=True, num_workers=2)\ntestset = torchvision.datasets.CIFAR10(root='.\/data', train=False,\n download=True, transform=transform)\ntestloader = torch.utils.data.DataLoader(testset, batch_size=32,\n shuffle=False, num_workers=2)\n <\/code><\/pre>\n3.2 GoogLeNet \ubaa8\ub378 \uc815\uc758<\/h3>\n
\ub2e4\uc74c\uc73c\ub85c GoogLeNet \ubaa8\ub378\uc744 \uc815\uc758\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc774\uc5d0 \uc0ac\uc6a9\ub420 Inception \ubaa8\ub4c8\uc744 \uc791\uc131\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\nimport torch.nn as nn\nimport torch.nn.functional as F\n\nclass Inception(nn.Module):\n def __init__(self, in_channels):\n super(Inception, self).__init__()\n self.branch1x1 = nn.Sequential(\n nn.Conv2d(in_channels, 64, kernel_size=1),\n nn.BatchNorm2d(64),\n nn.ReLU(inplace=True)\n )\n\n self.branch3x3 = nn.Sequential(\n nn.Conv2d(in_channels, 128, kernel_size=1),\n nn.BatchNorm2d(128),\n nn.ReLU(inplace=True),\n nn.Conv2d(128, 128, kernel_size=3, padding=1),\n nn.BatchNorm2d(128),\n nn.ReLU(inplace=True)\n )\n\n self.branch5x5 = nn.Sequential(\n nn.Conv2d(in_channels, 32, kernel_size=1),\n nn.BatchNorm2d(32),\n nn.ReLU(inplace=True),\n nn.Conv2d(32, 64, kernel_size=5, padding=2),\n nn.BatchNorm2d(64),\n nn.ReLU(inplace=True)\n )\n\n self.branch_pool = nn.Sequential(\n nn.MaxPool2d(kernel_size=3, stride=1, padding=1),\n nn.Conv2d(in_channels, 32, kernel_size=1),\n nn.BatchNorm2d(32),\n nn.ReLU(inplace=True)\n )\n\n def forward(self, x):\n branch1 = self.branch1x1(x)\n branch3 = self.branch3x3(x)\n branch5 = self.branch5x5(x)\n branch_pool = self.branch_pool(x)\n\n outputs = [branch1, branch3, branch5, branch_pool]\n return torch.cat(outputs, 1)\n <\/code><\/pre>\n3.3 \uc804\uccb4 GoogLeNet \uc815\uc758<\/h3>\n\nclass GoogLeNet(nn.Module):\n def __init__(self, num_classes=10):\n super(GoogLeNet, self).__init__()\n self.conv1 = nn.Conv2d(3, 64, kernel_size=7, stride=2, padding=3)\n self.pool1 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)\n\n self.conv2 = nn.Conv2d(64, 192, kernel_size=3, padding=1)\n self.pool2 = nn.MaxPool2d(kernel_size=3, stride=2, padding=1)\n\n self.inception1 = Inception(192)\n self.inception2 = Inception(256)\n self.inception3 = Inception(480)\n\n self.pool3 = nn.AvgPool2d(kernel_size=7)\n self.fc = nn.Linear(480, num_classes)\n\n def forward(self, x):\n x = F.relu(self.conv1(x))\n x = self.pool1(x)\n x = F.relu(self.conv2(x))\n x = self.pool2(x)\n\n x = self.inception1(x)\n x = self.inception2(x)\n x = self.inception3(x)\n\n x = self.pool3(x)\n x = x.view(x.size(0), -1)\n x = self.fc(x)\n return x\n\nmodel = GoogLeNet()\n <\/code><\/pre>\n3.4 \uc190\uc2e4 \ud568\uc218 \ubc0f \uc635\ud2f0\ub9c8\uc774\uc800 \uc815\uc758<\/h3>\n
\ubaa8\ub378\uc744 \ud6c8\ub828\ud560 \uc900\ube44\uac00 \ub418\uc5c8\uc73c\ubbc0\ub85c \uc190\uc2e4 \ud568\uc218\uc640 \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \uc815\uc758\ud569\ub2c8\ub2e4.<\/p>\n
\nimport torch.optim as optim\n\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.Adam(model.parameters(), lr=0.001)\n <\/code><\/pre>\n3.5 \ubaa8\ub378 \ud6c8\ub828<\/h3>\n
\uc774\uc81c \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uc8fc\uc5b4\uc9c4 Epoch \ub3d9\uc548 \uc190\uc2e4\uacfc \uc815\ud655\ub3c4\ub97c \ucd94\uc801\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
\nnum_epochs = 10\n\nfor epoch in range(num_epochs):\n model.train()\n running_loss = 0.0\n for i, data in enumerate(trainloader, 0):\n inputs, labels = data\n optimizer.zero_grad()\n outputs = model(inputs)\n loss = criterion(outputs, labels)\n loss.backward()\n optimizer.step()\n\n running_loss += loss.item()\n if i % 100 == 99: # \ub9e4 100 \ubc30\uce58\ub9c8\ub2e4 \ucd9c\ub825\n print(f'Epoch [{epoch+1}\/{num_epochs}], Step [{i+1}\/{len(trainloader)}], Loss: {running_loss \/ 100:.4f}')\n running_loss = 0.0\n print('\ud6c8\ub828 \uc644\ub8cc')\n\nprint('\ubaa8\ub378 \ud6c8\ub828 \ub05d!')\n <\/code><\/pre>\n3.6 \ubaa8\ub378 \ud3c9\uac00<\/h3>\n
\ud6c8\ub828\uc774 \uc644\ub8cc\ub418\uba74 \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud569\ub2c8\ub2e4.<\/p>\n
\ncorrect = 0\ntotal = 0\nmodel.eval()\nwith torch.no_grad():\n for data in testloader:\n images, labels = data\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'\uc815\ud655\ub3c4: {100 * correct \/ total:.2f}%')\n <\/code><\/pre>\n4. \uacb0\ub860<\/h2>\n
GoogLeNet\uc740 \ub2e4\uc591\ud55c \uc2a4\ucf00\uc77c\uc758 \ud2b9\uc9d5\uc744 \ud65c\uc6a9\ud560 \uc218 \uc788\ub294 \uac15\ub825\ud55c \ub124\ud2b8\uc6cc\ud06c \uad6c\uc870\ub97c \uc81c\uacf5\ud569\ub2c8\ub2e4. \uc774\ubc88 \uac15\uc88c\ub97c \ud1b5\ud574 GoogLeNet\uc758 \uae30\ubcf8 \uac1c\ub150 \ubc0f PyTorch\uc5d0\uc11c\uc758 \uad6c\ud604 \ubc29\ubc95\uc744 \ubc30\uc6e0\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc774\ud574\ub97c \ud1b5\ud574 \ubcf4\ub2e4 \ubcf5\uc7a1\ud55c \ubaa8\ub378\uc5d0\uc11c\ub3c4 \uc720\uc0ac\ud55c \ubc29\uc2dd\uc744 \uc801\uc6a9\ud560 \uc218 \uc788\uc744 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
\ucd94\uac00\uc801\uc73c\ub85c, GoogLeNet\uc740 \ub9ce\uc740 \ubcc0\ud615\uc774 \uc874\uc7ac\ud569\ub2c8\ub2e4. Inception v2, Inception v3 \ub4f1 \ubaa8\ub378\uc758 \uae4a\uc774\ub098 \uad6c\uc870\ub97c \uc870\uc815\ud558\uc5ec \uc131\ub2a5\uc744 \uac1c\uc120\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \ubcc0\ud615 \ubaa8\ub378\ub4e4\uc740 \ub354\uc6b1 \uc815\ud655\ud55c \uc608\uce21\uc744 \ub3c4\uc640\uc904 \uac83\uc785\ub2c8\ub2e4. \ub2e4\uc74c \uac15\uc88c\uc5d0\uc11c\ub294 \uc774\ub7ec\ud55c \ubcc0\ud615 \ubaa8\ub378\ub4e4\uc5d0 \ub300\ud574\uc11c\ub3c4 \ub2e4\ub8f0 \uc608\uc815\uc785\ub2c8\ub2e4.<\/p>\n
\uc774\uc0c1\uc73c\ub85c GoogLeNet\uc5d0 \ub300\ud55c \uc124\uba85\uc744 \ub9c8\uce58\uaca0\uc2b5\ub2c8\ub2e4. \uac10\uc0ac\ud569\ub2c8\ub2e4!<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"