\uc804\uc774 \ud559\uc2b5(Transfer Learning)\uc740 \ud2b9\uc815 \ubd84\uc57c\uc5d0\uc11c \ud559\uc2b5\ub41c \uc2e0\uacbd\ub9dd\uc758 \uc77c\ubd80 \ub2a5\ub825\uc744 \uc720\uc0ac\ud558\uac70\ub098 \uc804\ud600 \uc0c8\ub85c\uc6b4 \ubd84\uc57c\uc5d0\uc11c \uc0ac\uc6a9\ub418\ub294 \uc2e0\uacbd\ub9dd\uc758 \ud559\uc2b5\uc5d0 \uc774\uc6a9\ud558\ub294 \uac83\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4.<\/p>\n
\uc774\ubbf8\uc9c0 \ubd84\ub958\ub97c \uc608\ub85c \ub4e4\uc5b4 Resnet\uc774\ub098 VGG \ub4f1\uacfc \uac19\uc740 \uc2e0\uacbd\ub9dd\uc758 \uad6c\uc131 \uc911 \uc55e\ub2e8\uc740 CNN \ub808\uc774\uc5b4\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. \uc774 CNN \ub808\uc774\uc5b4\ub294 \uc774\ubbf8\uc9c0\uc758 \ud2b9\uc9d5\uc744 \ucd94\ucd9c\ud558\ub294 \ub2a5\ub825\uc744 \uac16\ub294\ub370\uc694. \ucc98\uc74c\uc5d0\ub294 \uc2e0\ud615\uc131\uc744 \ucd94\ucd9c\ud558\uace0 \ub2e4\uc74c\uc5d0\ub294 \ud328\ud134\uc744, \ub9c8\uc9c0\ub9c9\uc5d0\ub294 \ud615\uc0c1 \ub4f1\uc744 \ucd94\ucd9c\ud55c\ub2e4\uace0 \uc54c\ub824\uc838 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc774\ubbf8\uc9c0\uc758 \ud2b9\uc9d5\uc744 \ucd94\ucd9c\ud558\ub294 \uc2e0\uacbd\ub9dd\uc758 \ub2a5\ub825\uc740 \ub2e4\ub978 \ubd84\uc57c\uc5d0\uc11c\ub3c4 \ud65c\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc989, \uc218\ub9cc\uc5d0\uc11c \uc218\ucc9c\ub9cc\uc7a5\uc758 \uc774\ubbf8\uc9c0\ub97c \ud1b5\ud574 \ud559\uc2b5\ub41c \ub192\uc740 \uc131\ub2a5\uc744 \uac16\ub294 Resnet\uc774\ub098 VGG \uc2e0\uacbd\ub9dd\uc758 \ud2b9\uc9d5 \ucd94\ucd9c \ub2a5\ub825\uc744 \uadf8\ub300\ub85c \uc774\uc6a9\ud558\uace0, \ub9c8\uc9c0\ub9c9 \ucd9c\ub825 \uacc4\uce35\uc73c\ub85c\uc368.. \uc8fc\ub85c \uc120\ud615(Affine; \uac00\uc911\uce58\uc640 \ud3b8\ud5a5\uc5d0 \ub300\ud55c \ud589\ub82c \uc5f0\uc0b0) \ub808\uc774\uc5b4\ub9cc\uc744 \ubcc0\uacbd\ud558\uc5ec \uc774 \ubcc0\uacbd\ub41c \ub808\uc774\uc5b4\ub9cc\uc744 \uc7ac\ud559\uc2b5\uc2dc\ud0a4\ub294 \uac83\uc774 \uc804\uc774 \ud559\uc2b5\uc785\ub2c8\ub2e4.<\/p>\n
\uc804\uc774 \ud559\uc2b5\uc740 \ud559\uc2b5 \ub370\uc774\ud130\uc758 \uc218\uac00 \uc801\uc744\ub54c\ub3c4 \ud6a8\uacfc\uc801\uc774\uba70, \ud559\uc2b5 \uc18d\ub3c4\ub3c4 \ube60\ub985\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \uc804\uc774\ud559\uc2b5 \uc5c6\uc774 \ud559\uc2b5\ud558\ub294 \uac83\ubcf4\ub2e4 \ud6e8\uc52c \ub192\uc740 \uc815\ud655\ub3c4\ub97c \uc81c\uacf5\ud55c\ub2e4\ub294 \uc7a5\uc810\uc774 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc774 \uae00\uc740 Resnet\uacfc VGG \uc2e0\uacbd\ub9dd\uc5d0 \ub300\ud55c \uc804\uc774\ud559\uc2b5 \ucf54\ub4dc \uc911 \uc804\uc774\ud559\uc2b5\uc744 \uc704\ud55c \uc804\ucc98\ub9ac \ucf54\ub4dc\ub97c \uc815\ub9ac\ud569\ub2c8\ub2e4. \ub098\uba38\uc9c0 \ud559\uc2b5 \ub4f1\uc758 \ucf54\ub4dc\ub294 \uc5ec\ud0c0 \ub2e4\ub978 \uc2e0\uacbd\ub9dd\uacfc \ub3d9\uc77c\ud569\ub2c8\ub2e4. \uba3c\uc800 \uc804\uc774\ud559\uc2b5\uc744 \uc704\ud55c Resnet \uc2e0\uacbd\ub9dd\uc758 \uc804\ucc98\ub9ac \ucf54\ub4dc\uc785\ub2c8\ub2e4.<\/p>\n
\r\nimport torch.nn as nn\r\nfrom torchvision import models\r\n\r\nnet = models.resnet18(pretrained=True)\r\n\r\nfor p in net.parameters():\r\n p.requires_grad = False\r\n\r\nfc_input_dim = net.fc.in_features\r\nnet.fc = nn.Linear(fc_input_dim, 2)\r\n<\/pre>\n\uba3c\uc800 \uc774\ubbf8 \ud559\uc2b5\ub41c resnet18 \uc2e0\uacbd\ub9dd\uc744 \ubd88\ub7ec\uc624\uace0, \uc774 \uc2e0\uacbd\ub9dd\uc758 \uac00\uc911\uce58\uac00 \ud559\uc2b5\ub418\uc9c0 \uc54a\ub3c4\ub85d \ud569\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \uc774 \uc2e0\uacbd\ub9dd\uc758 \ub9c8\uc9c0\ub9c9 \uad6c\uc131 \ub808\uc774\uc5b4(fully connected layer\ub85c\uc368 Affine Layer, Dense layer\ub77c\uace0\ub3c4 \ud568)\uc758 \uc785\ub825 \ub370\uc774\ud130 \uc218\ub97c \uc5bb\uace0, \uc774\ub807\uac8c \uc5bb\ub294 \uc785\ub825 \ub370\uc774\ud130\uc758 \uc218\uc640 \ucd9c\ub825\ud558\uace0\uc790 \ud558\ub294, \uc989 \ubd84\ub958 \uac1c\uc218\uc778 2\uc5d0 \ub300\ud55c \uc120\ud615 \ub808\uc774\uc5b4\ub97c \uc0dd\uc131\ud558\uc5ec \uc2e0\uacbd\ub9dd\uc744 \uad6c\uc131\ud558\ub294 \ub9c8\uc9c0\ub9c9 \ub808\uc774\uc5b4\ub97c \uad50\uccb4\ud569\ub2c8\ub2e4. \uacb0\uacfc\uc801\uc73c\ub85c \uc774 \uc2e0\uacbd\ub9dd\uc758 \ub9c8\uc9c0\ub9c9 \ub808\uc774\uc5b4\ub97c \uc81c\uc678\ud55c \ud2b9\uc9d5 \ucd94\ucd9c \ub808\uc774\uc5b4\ub4e4\uc740 \ud559\uc2b5\ub418\uc9c0 \uc54a\uace0, \ub9c8\uc9c0\ub9c9 \ub808\uc774\uc5b4\ub9cc\uc774 \ud559\uc2b5\ub420 \uac83\uc785\ub2c8\ub2e4.<\/p>\n
\ucc38\uace0\ub85c \uc704\uc758 \uc2e0\uacbd\ub9dd\uc758 \uad6c\uc131 \ub808\uc774\uc5b4\ub97c \ucd9c\ub825\ud558\ub294 \ucf54\ub4dc\uc640 \uadf8 \uacb0\uacfc\ub294 \ub2e4\uc74c\uacfc \uac19\uc740\ub370, \uad6c\uc131 \ub808\uc774\uc5b4\uc758 \ub9c8\uc9c0\ub9c9\uc774 fc\ub77c\ub294 \uac83\uc744 \uc54c \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\r\nfor name,module in net.named_children():\r\n print(name)\r\n\r\n''' output:\r\nconv1\r\nbn1\r\nrelu\r\nmaxpool\r\nlayer1\r\nlayer2\r\nlayer3\r\nlayer4\r\navgpool\r\nfc\r\n'''\r\n<\/pre>\n\ub2e4\uc74c\uc740 VGG \uc2e0\uacbd\ub9dd\uc5d0 \ub300\ud55c \uc804\uc774\ud559\uc2b5 \uc804\ucc98\ub9ac \ucf54\ub4dc\uc785\ub2c8\ub2e4.<\/p>\n
\r\nfrom torchvision import models\r\n\r\nnet = models.vgg16(pretrained=True)\r\n \r\nfeatures = net.features\r\nfor params in vgg.features.parameters():\r\n param.requires_grad = False\r\n\r\nnet.classifier[6].out_features = 2\r\n<\/pre>\n\uac1d\uccb4 net\uc744 \uc0dd\uc131\ud55c \ud6c4 \ubc14\ub85c print(net)\uc744 \uc2e4\ud589\ud574 \ubcf4\uba74 \ub2e4\uc74c\uacfc \uac19\uc740 \ucd9c\ub825\uc744 \ubcfc \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\r\nVGG(\r\n (features): Sequential(\r\n (0): Conv2d(3, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (1): ReLU(inplace=True)\r\n (2): Conv2d(64, 64, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (3): ReLU(inplace=True)\r\n (4): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\r\n (5): Conv2d(64, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (6): ReLU(inplace=True)\r\n (7): Conv2d(128, 128, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (8): ReLU(inplace=True)\r\n (9): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\r\n (10): Conv2d(128, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (11): ReLU(inplace=True)\r\n (12): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (13): ReLU(inplace=True)\r\n (14): Conv2d(256, 256, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (15): ReLU(inplace=True)\r\n (16): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\r\n (17): Conv2d(256, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (18): ReLU(inplace=True)\r\n (19): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (20): ReLU(inplace=True)\r\n (21): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (22): ReLU(inplace=True)\r\n (23): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\r\n (24): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (25): ReLU(inplace=True)\r\n (26): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (27): ReLU(inplace=True)\r\n (28): Conv2d(512, 512, kernel_size=(3, 3), stride=(1, 1), padding=(1, 1))\r\n (29): ReLU(inplace=True)\r\n (30): MaxPool2d(kernel_size=2, stride=2, padding=0, dilation=1, ceil_mode=False)\r\n )\r\n (avgpool): AdaptiveAvgPool2d(output_size=(7, 7))\r\n (classifier): Sequential(\r\n (0): Linear(in_features=25088, out_features=4096, bias=True)\r\n (1): ReLU(inplace=True)\r\n (2): Dropout(p=0.5, inplace=False)\r\n (3): Linear(in_features=4096, out_features=4096, bias=True)\r\n (4): ReLU(inplace=True)\r\n (5): Dropout(p=0.5, inplace=False)\r\n (6): Linear(in_features=4096, out_features=1000, bias=True)\r\n )\r\n)\r\n<\/pre>\n(classifier)\uc758 \ub9c8\uc9c0\ub9c9 \uad6c\uc131\uc694\uc18c[6]\uc744 \ubcf4\uba74 out_features\uac00 1000\uc73c\ub85c \ub418\uc5b4 \uc788\ub294 \uac83\uc744 \ubcfc \uc218 \uc788\uace0, \uc774\ub97c \ubd84\ub958\ud558\uace0\uc790 \ud558\ub294 \uac1c\uc218\uc778 2\ub85c \ubcc0\uacbd\ud558\ub294 \uc804\ucc98\ub9ac \ucf54\ub4dc\uc600\uc2b5\ub2c8\ub2e4.<\/p>\n","protected":false},"excerpt":{"rendered":"
\uc804\uc774 \ud559\uc2b5(Transfer Learning)\uc740 \ud2b9\uc815 \ubd84\uc57c\uc5d0\uc11c \ud559\uc2b5\ub41c \uc2e0\uacbd\ub9dd\uc758 \uc77c\ubd80 \ub2a5\ub825\uc744 \uc720\uc0ac\ud558\uac70\ub098 \uc804\ud600 \uc0c8\ub85c\uc6b4 \ubd84\uc57c\uc5d0\uc11c \uc0ac\uc6a9\ub418\ub294 \uc2e0\uacbd\ub9dd\uc758 \ud559\uc2b5\uc5d0 \uc774\uc6a9\ud558\ub294 \uac83\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4. \uc774\ubbf8\uc9c0 \ubd84\ub958\ub97c \uc608\ub85c \ub4e4\uc5b4 Resnet\uc774\ub098 VGG \ub4f1\uacfc \uac19\uc740 \uc2e0\uacbd\ub9dd\uc758 \uad6c\uc131 \uc911 \uc55e\ub2e8\uc740 CNN \ub808\uc774\uc5b4\ub85c \uad6c\uc131\ub418\uc5b4 \uc788\uc2b5\ub2c8\ub2e4. \uc774 CNN \ub808\uc774\uc5b4\ub294 \uc774\ubbf8\uc9c0\uc758 \ud2b9\uc9d5\uc744 \ucd94\ucd9c\ud558\ub294 \ub2a5\ub825\uc744 \uac16\ub294\ub370\uc694. \ucc98\uc74c\uc5d0\ub294 \uc2e0\ud615\uc131\uc744 \ucd94\ucd9c\ud558\uace0 \ub2e4\uc74c\uc5d0\ub294 \ud328\ud134\uc744, \ub9c8\uc9c0\ub9c9\uc5d0\ub294 \ud615\uc0c1 \ub4f1\uc744 \ucd94\ucd9c\ud55c\ub2e4\uace0 \uc54c\ub824\uc838 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub7ec\ud55c … <\/p>\n