\uc791\uc131\uc790: \uc870\uad11\ud615<\/p>\n
\uc791\uc131\uc77c: [\ub0a0\uc9dc]<\/p>\n<\/header>\n \uacf5\uac04 \ud53c\ub77c\ubbf8\ub4dc \ud480\ub9c1(SPP, Spatial Pyramid Pooling)\uc740 \uc774\ubbf8\uc9c0 \ubd84\ub958\uc640 \uac19\uc740 \ub2e4\uc591\ud55c \ube44\uc804 \ud0dc\uc2a4\ud06c\uc758 \ubaa8\ub378\uc5d0\uc11c \uc0ac\uc6a9\ub418\ub294 \uae30\ubc95\uc785\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc778 \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd(CNN)\uc740 \uace0\uc815\ub41c \ud06c\uae30\uc758 \uc785\ub825\uc744 \ud544\uc694\ub85c \ud558\uc9c0\ub9cc, SPP\ub294 \ubcc0\ub3d9\ud558\ub294 \ud06c\uae30\uc758 \uc774\ubbf8\uc9c0\ub97c \uc785\ub825\uc73c\ub85c \ubc1b\uc744 \uc218 \uc788\ub294 \ud2b9\uc131\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub294 SPP\uac00 \uc785\ub825 \uc774\ubbf8\uc9c0\uc758 \ud06c\uae30\ub97c \uc5ec\ub7ec \uacb9\uc73c\ub85c \ub098\ub204\ub294 \ud53c\ub77c\ubbf8\ub4dc \uad6c\uc870\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud2b9\uc9d5\uc744 \ucd94\ucd9c\ud558\uae30 \ub54c\ubb38\uc785\ub2c8\ub2e4.<\/p>\n \uc804\ud1b5\uc801\uc778 \ud480\ub9c1 \ubc29\uc2dd\uc740 \uace0\uc815\ub41c \ud06c\uae30\uc758 \uc601\uc5ed\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud2b9\uc131\uc744 \uc9d1\uacc4\ud558\ub294 \ubc18\uba74, SPP\ub294 \ub2e4\uc591\ud55c \ud06c\uae30\uc758 \uc601\uc5ed\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud480\ub9c1\uc744 \uc218\ud589\ud569\ub2c8\ub2e4. \uc774 \ubc29\uc2dd\uc740 \uac1d\uccb4\uac00 \ub2e4\uc591\ud55c \ud06c\uae30\ub85c \uc874\uc7ac\ud558\ub294 \uc2e4\uc81c \ud658\uacbd\uc5d0\uc11c \ub354 \ub098\uc740 \uc131\ub2a5\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.<\/p>\n<\/section>\n SPP\ub294 \uc5ec\ub7ec \uc218\uc900\uc758 \ud480\ub9c1 \ub808\uc774\uc5b4\ub97c \ud1b5\ud574 \uc785\ub825 \uc774\ubbf8\uc9c0\ub97c \ucc98\ub9ac\ud569\ub2c8\ub2e4. \ud53c\ub77c\ubbf8\ub4dc \uad6c\uc870\uac00 \uc0ac\uc6a9\ub418\ubbc0\ub85c, \uac01 \uc218\uc900(level)\uc5d0\uc11c \ub2e4\ub978 \ud06c\uae30\uc758 \uc601\uc5ed\uc744 \uc124\uc815\ud558\uc5ec \ud574\ub2f9 \uc601\uc5ed \ub0b4\uc758 \ud2b9\uc9d5\uc744 \ucd94\ucd9c\ud569\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, 1×1, 2×2, 4×4 \uc601\uc5ed\uc744 \uc0ac\uc6a9\ud558\uc5ec \uac01\uae30 \ub2e4\ub978 \uac1c\uc218\uc758 \ud2b9\uc131\uc744 \ucd94\ucd9c\ud569\ub2c8\ub2e4.<\/p>\n \uc774\ub807\uac8c \ucd94\ucd9c\ub41c \ud2b9\uc9d5\ub4e4\uc740 \ucd5c\uc885\uc801\uc73c\ub85c \ud558\ub098\uc758 \ubca1\ud130\ub85c \uacb0\ud569\ub418\uc5b4 \ubd84\ub958\uae30\ub85c \uc804\ub2ec\ub429\ub2c8\ub2e4. SPP\ub294 \uc774\ubbf8\uc9c0\uc758 \ub2e4\uc591\ud55c \uacf5\uac04 \uc815\ubcf4\uc640 \ud2b9\uc9d5\uc744 \ud6a8\uacfc\uc801\uc73c\ub85c \ucea1\ucc98\ud558\uc5ec \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ub370 \uae30\uc5ec\ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n SPP\ub294 CNN\uacfc \ud1b5\ud569\ub418\uc5b4 \ub2e4\uc74c\uacfc \uac19\uc740 \ubc29\uc2dd\uc73c\ub85c \uc791\ub3d9\ud569\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc778 CNN \uad6c\uc870\ub97c \uac00\uc9c4 \ub124\ud2b8\uc6cc\ud06c\uc758 \ucd9c\ub825\uc5d0 SPP \ub808\uc774\uc5b4\ub97c \ucd94\uac00\ud558\uc5ec, \ucd9c\ub825\ub41c \ud2b9\uc9d5 \ub9f5\uc744 SPP\ub97c \ud1b5\ud574 \ud480\ub9c1\ud558\uace0 \uc774\ub97c \ubd84\ub958\uae30\ub85c \uc804\ub2ec\ud569\ub2c8\ub2e4. SPP \ub808\uc774\uc5b4\ub294 \uc77c\ubc18\uc801\uc73c\ub85c CNN\uc758 \ub9c8\uc9c0\ub9c9 \ud3b8\uc9d1 \ub808\uc774\uc5b4\uc5d0 \uc704\uce58\ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n \uc774\uc81c \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c SPP \ub808\uc774\uc5b4\ub97c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc544\ub798 \ucf54\ub4dc\ub294 SPP \ub808\uc774\uc5b4\ub97c \uc815\uc758\ud558\ub294 \uac04\ub2e8\ud55c \uc608\uc81c\uc785\ub2c8\ub2e4:<\/p>\n \uc704 \ucf54\ub4dc\ub294 SPP \ub808\uc774\uc5b4\uc758 \uae30\ubcf8 \uad6c\ud604\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uc5ec\ub7ec \ub808\ubca8\uc5d0\uc11c\uc758 \ud480\ub9c1\uc744 \uc9c0\uc6d0\ud558\uba70, \uc785\ub825\ub41c \ud2b9\uc9d5 \ub9f5\uc5d0\uc11c SPP\ub97c \ud1b5\ud574 \ucd5c\uc885 \ucd9c\ub825\uc744 \uc0dd\uc131\ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n \uc774\uc81c SPP \ub808\uc774\uc5b4\ub97c CNN \ub124\ud2b8\uc6cc\ud06c\uc5d0 \ud1b5\ud569\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc544\ub798 \uc608\uc81c \ucf54\ub4dc\ub294 CNN \uad6c\uc870\uc5d0 SPP \ub808\uc774\uc5b4\ub97c \ud1b5\ud569\ud558\ub294 \ubc29\ubc95\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4:<\/p>\n \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 2\uac1c\uc758 \ud569\uc131\uacf1 \ub808\uc774\uc5b4\uc640 2\uac1c\uc758 \uc644\uc804 \uc5f0\uacb0 \ub808\uc774\uc5b4\ub97c \uac00\uc9c4 \uac04\ub2e8\ud55c CNN \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud588\uc2b5\ub2c8\ub2e4. SPP \ub808\uc774\uc5b4\ub294 \ud569\uc131\uacf1 \ub808\uc774\uc5b4 \ub2e4\uc74c\uc5d0 \uc704\uce58\ud558\uc5ec \uc785\ub825 \uc774\ubbf8\uc9c0\ub97c \ucc98\ub9ac\ud569\ub2c8\ub2e4.<\/p>\n<\/section>\n \uba3c\uc800 \ubaa8\ub378\uc744 \ud559\uc2b5\ud558\uae30 \uc704\ud55c \ub370\uc774\ud130\uc14b\uc744 \uc124\uc815\ud558\uace0 \uc635\ud2f0\ub9c8\uc774\uc800 \ubc0f \uc190\uc2e4 \ud568\uc218\ub97c \uc815\uc758\ud558\uaca0\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \ubaa8\ub378 \ud559\uc2b5\uc758 \uc804\uccb4 \uacfc\uc815\uc785\ub2c8\ub2e4:<\/p>\n \uc704 \ucf54\ub4dc\ub294 CIFAR-10 \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378\uc744 \ud559\uc2b5\ud558\ub294 \uacfc\uc815\uc744 \ubcf4\uc5ec\uc90d\ub2c8\ub2e4. \uac01 \uc5d0\ud3ed\uc758 \uc190\uc2e4\uc744 \ucd9c\ub825\ud558\uc5ec \ud559\uc2b5 \uacfc\uc815\uc744 \ubaa8\ub2c8\ud130\ub9c1\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n \ubaa8\ub378 \ud559\uc2b5\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\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ub2e4\uc74c\uc740 \ubaa8\ub378 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\ub294 \ucf54\ub4dc\uc785\ub2c8\ub2e4:<\/p>\n \uc704 \ucf54\ub4dc\ub294 \ubaa8\ub378\uc758 \uc815\ud655\ub3c4\ub97c \ud3c9\uac00\ud558\uace0 \uacb0\uacfc\ub97c \ucd9c\ub825\ud569\ub2c8\ub2e4. \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc5d0\uc11c \ubaa8\ub378\uc774 \uc5bc\ub9c8\ub098 \uc815\ud655\ud55c\uc9c0 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n<\/section>\n \uc774\ubc88 \uac15\uc88c\uc5d0\uc11c\ub294 SPP(Space Pyramid Pooling)\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \uc6d0\ub9ac, \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c\uc758 \uad6c\ud604 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. SPP\ub294 \ub2e4\uc591\ud55c \ud06c\uae30\uc758 \uc774\ubbf8\uc9c0\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \ucc98\ub9ac\ud560 \uc218 \uc788\ub294 \uac15\ub825\ud55c \uae30\uc220\ub85c, \ub525\ub7ec\ub2dd \ube44\uc804 \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ub370 \ud070 \ub3c4\uc6c0\uc744 \uc90d\ub2c8\ub2e4.<\/p>\n \ub354 \uae4a\uc774 \uc788\ub294 \ud559\uc2b5\uc744 \uc6d0\ud558\uc2e0\ub2e4\uba74 \ub2e4\uc74c \uc790\ub8cc\ub4e4\uc744 \ucc38\uace0\ud558\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4:<\/p>\n1. \uacf5\uac04 \ud53c\ub77c\ubbf8\ub4dc \ud480\ub9c1(SPP)\uc774\ub780?<\/h2>\n
2. SPP\uc758 \uc791\ub3d9 \uc6d0\ub9ac<\/h2>\n
3. SPP\uc758 \uc7a5\uc810<\/h2>\n
\n
4. SPP\uc640 CNN \ud1b5\ud569<\/h2>\n
5. \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c SPP \ub808\uc774\uc5b4 \uad6c\ud604\ud558\uae30<\/h2>\n
\nimport torch\nimport torch.nn as nn\nimporttorch.nn.functional as F\n\nclass SpatialPyramidPooling(nn.Module):\n def __init__(self, levels):\n super(SpatialPyramidPooling, self).__init__()\n # \uac01 \uc218\uc900(level)\uc5d0\uc11c\uc758 \ud480\ub9c1 \ud06c\uae30 \uc815\uc758\n self.levels = levels\n self.pooling_layers = []\n\n for level in levels:\n self.pooling_layers.append(nn.AdaptiveAvgPool2d((level, level)))\n\n def forward(self, x):\n # \ud2b9\uc9d5 \ub9f5\uc744 \ucc98\ub9ac\ud558\uc5ec \ud2b9\uc131\uc744 \ucd94\ucd9c\n batch_size = x.size(0)\n pooled_outputs = []\n\n for pooling_layer in self.pooling_layers:\n pooled_output = pooling_layer(x)\n pooled_output = pooled_output.view(batch_size, -1)\n pooled_outputs.append(pooled_output)\n\n # \ubaa8\ub4e0 \ud480\ub9c1\ub41c \ucd9c\ub825 \uacb0\ud569\n final_output = torch.cat(pooled_outputs, 1)\n return final_output\n <\/code><\/pre>\n6. SPP \ub808\uc774\uc5b4\ub97c CNN\uc5d0 \ud1b5\ud569\ud558\uae30<\/h2>\n
\nclass CNNWithSPP(nn.Module):\n def __init__(self, num_classes):\n super(CNNWithSPP, self).__init__()\n self.conv1 = nn.Conv2d(3, 16, kernel_size=3, stride=1, padding=1)\n self.conv2 = nn.Conv2d(16, 32, kernel_size=3, stride=1, padding=1)\n self.fc1 = nn.Linear(32 * 8 * 8, 128) # \ucd5c\uc885 \ud30c\ub77c\ubbf8\ud130\ub294 SPP\uc758 \ucd9c\ub825\uc5d0 \ub530\ub77c \uc870\uc815\n self.fc2 = nn.Linear(128, num_classes)\n self.spp = SpatialPyramidPooling(levels=[1, 2, 4]) # SPP \ub808\uc774\uc5b4 \ucd94\uac00\n\n def forward(self, x):\n x = F.relu(self.conv1(x))\n x = F.relu(self.conv2(x))\n x = self.spp(x) # SPP\ub97c \ud1b5\ud574 \ud2b9\uc9d5 \ucd94\ucd9c\n x = F.relu(self.fc1(x))\n x = self.fc2(x)\n return x\n <\/code><\/pre>\n7. \ubaa8\ub378 \ud559\uc2b5 \ubc0f \ud3c9\uac00<\/h2>\n
\nimport torchvision\nimport torchvision.transforms as transforms\n\n# \ub370\uc774\ud130\uc14b \ubd88\ub7ec\uc624\uae30\ntransform = transforms.Compose(\n [transforms.Resize((32, 32)),\n transforms.ToTensor()])\ntrainset = torchvision.datasets.CIFAR10(root='.\/data', train=True,\n download=True, transform=transform)\ntrainloader = torch.utils.data.DataLoader(trainset, batch_size=64,\n shuffle=True, num_workers=2)\n\n# \ubaa8\ub378 \ubc0f \uc635\ud2f0\ub9c8\uc774\uc800 \uc124\uc815\nmodel = CNNWithSPP(num_classes=10)\ncriterion = nn.CrossEntropyLoss()\noptimizer = torch.optim.Adam(model.parameters(), lr=0.001)\n\n# \ubaa8\ub378 \ud559\uc2b5\nfor epoch in range(10): # 10 epochs\n for inputs, labels in trainloader:\n optimizer.zero_grad() # \uadf8\ub798\ub514\uc5b8\ud2b8 \ucd08\uae30\ud654\n outputs = model(inputs) # \ubaa8\ub378 \uc608\uce21\n loss = criterion(outputs, labels) # \uc190\uc2e4 \uacc4\uc0b0\n loss.backward() # \uadf8\ub798\ub514\uc5b8\ud2b8 \uacc4\uc0b0\n optimizer.step() # \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8\n\n print(f'Epoch {epoch + 1}, Loss: {loss.item()}') # \uac01 \uc5d0\ud3ed\uc758 \uc190\uc2e4 \ucd9c\ub825\n <\/code><\/pre>\n8. \ubaa8\ub378 \ud3c9\uac00 \ubc0f \uc131\ub2a5 \ubd84\uc11d<\/h2>\n
\n# \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc14b \ubd88\ub7ec\uc624\uae30\ntestset = torchvision.datasets.CIFAR10(root='.\/data', train=False,\n download=True, transform=transform)\ntestloader = torch.utils.data.DataLoader(testset, batch_size=64,\n shuffle=False, num_workers=2)\n\n# \ubaa8\ub378 \ud3c9\uac00\nmodel.eval() # \ud3c9\uac00 \ubaa8\ub4dc\ub85c \uc804\ud658\ncorrect = 0\ntotal = 0\n\nwith torch.no_grad():\n for inputs, labels in testloader:\n outputs = model(inputs) # \ubaa8\ub378 \uc608\uce21\n _, predicted = torch.max(outputs.data, 1)\n total += labels.size(0)\n correct += (predicted == labels).sum().item()\n\nprint(f'Accuracy: {100 * correct \/ total:.2f}%') # \uc815\ud655\ub3c4 \ucd9c\ub825\n <\/code><\/pre>\n9. \ub9c8\ubb34\ub9ac \ubc0f \ucd94\uac00 \ub9ac\uc18c\uc2a4<\/h2>\n