\ub525\ub7ec\ub2dd\uc740 \ub2e4\uc591\ud55c \ub370\uc774\ud130\ub85c\ubd80\ud130 \uc720\uc6a9\ud55c \ud2b9\uc131\uc744 \uc790\ub3d9\uc73c\ub85c \ud559\uc2b5\ud558\uc5ec \ubb38\uc81c \ud574\uacb0\uc744 \uac00\ub2a5\ud558\uac8c \ud558\ub294 \uac15\ub825\ud55c \uae30\uc220\uc785\ub2c8\ub2e4. \uc624\ub298\uc740 \ud30c\uc774\ud1a0\uce58(PyTorch) \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uc5ec \ud2b9\uc131 \ucd94\ucd9c \uae30\ubc95\uc744 \ub2e4\ub8e8\uace0\uc790 \ud569\ub2c8\ub2e4. \uc774\ub294 \uc774\ubbf8\uc9c0, \ud14d\uc2a4\ud2b8, \uc624\ub514\uc624 \ub4f1 \uc5ec\ub7ec \ud615\ud0dc\uc758 \ub370\uc774\ud130\uc5d0\uc11c \uc18d\uc131\uc744 \ucd94\ucd9c\ud558\uc5ec \uba38\uc2e0\ub7ec\ub2dd \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud5a5\uc0c1\uc2dc\ud0a4\ub294 \ub370 \uc911\uc694\ud55c \uc5ed\ud560\uc744 \ud569\ub2c8\ub2e4.<\/p>\n
\ud2b9\uc131 \ucd94\ucd9c(features extraction)\uc740 \uc6d0\ubcf8 \ub370\uc774\ud130\ub97c \ubcf4\ub2e4 \ub0ae\uc740 \ucc28\uc6d0\uc73c\ub85c \ubcc0\ud658\ud558\uc5ec \uc720\uc6a9\ud55c \uc815\ubcf4\ub97c \ucd94\ucd9c\ud558\ub294 \uacfc\uc815\uc744 \uc758\ubbf8\ud569\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc740 \ub370\uc774\ud130\uc758 \ub178\uc774\uc988\ub97c \uc904\uc774\uace0, \ubaa8\ub378\uc774 \ud559\uc2b5\ud558\ub294 \ub370 \uc788\ub294 \uc5b4\ub824\uc6c0\uc744 \ud574\uacb0\ud558\ub294 \ub370 \ub3c4\uc6c0\uc744 \uc90d\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, \uc774\ubbf8\uc9c0 \ubd84\ub958 \ubb38\uc81c\uc5d0\uc11c, \uc774\ubbf8\uc9c0\uc758 \ud53d\uc140 \uac12\uc744 \uc9c1\uc811 \uc0ac\uc6a9\ud558\ub294 \ub300\uc2e0, CNN(\ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd)\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc911\uc694\ud55c \ud2b9\uc131\ub9cc\uc744 \ucd94\ucd9c\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
\uc6b0\ub9ac\ub294 \uc774\ubbf8\uc9c0 \ucc98\ub9ac \ubd84\uc57c\uc5d0\uc11c CNN\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud2b9\uc131\uc744 \ucd94\ucd9c\ud558\ub294 \uc608\uc81c\ub97c \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. CNN\uc740 \uc774\ubbf8\uc9c0\uc758 \uc9c0\uc5ed \uc815\ubcf4\ub97c \uc798 \uc7a1\uc544\ub0b4\ub294 \ub370 \uc720\ub9ac\ud55c \uad6c\uc870\ub97c \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n
import torchvision.transforms as transforms\nimport torchvision.datasets as datasets\nimport torch\nimport torch.nn as nn\nimport torchvision.models as models\n\n# \ub370\uc774\ud130\uc14b \ub2e4\uc6b4\ub85c\ub4dc \ubc0f \uc804\ucc98\ub9ac\ntransform = transforms.Compose([\n transforms.Resize((224, 224)),\n transforms.ToTensor(),\n])\n\n# CIFAR10 \ub370\uc774\ud130\uc14b \uc608\uc2dc\ntrain_dataset = datasets.CIFAR10(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = torch.utils.data.DataLoader(dataset=train_dataset, batch_size=32, shuffle=True)\n<\/code><\/pre>\n1.2 CNN \ubaa8\ub378 \uc815\uc758<\/h3>\n
ResNet\uc744 \uae30\ubc18\uc73c\ub85c \ud55c \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud2b9\uc131\uc744 \ucd94\ucd9c\ud558\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
# ResNet \ubaa8\ub378 \ubd88\ub7ec\uc624\uae30\nmodel = models.resnet18(pretrained=True) # \uc0ac\uc804 \ud559\uc2b5\ub41c \ubaa8\ub378\nmodel.fc = nn.Identity() # \ub9c8\uc9c0\ub9c9 \ub808\uc774\uc5b4 \uc81c\uac70\ud558\uc5ec \ud2b9\uc131\ub9cc \ucd9c\ub825\n\n# GPU \uc0ac\uc6a9 \uc124\uc815\ndevice = torch.device('cuda' if torch.cuda.is_available() else 'cpu')\nmodel = model.to(device)\n<\/code><\/pre>\n1.3 \ud2b9\uc131 \ucd94\ucd9c<\/h3>\n
\ucd94\ucd9c\ud55c \ud2b9\uc131\uc758 \ud615\ud0dc\ub97c \ud655\uc778\ud558\uae30 \uc704\ud574, \ub370\uc774\ud130\ub97c \ubaa8\ub378\uc5d0 \ud1b5\uacfc\uc2dc\ucf1c \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
def extract_features(data_loader):\n features = []\n \n model.eval() # \ud3c9\uac00 \ubaa8\ub4dc\ub85c \uc804\ud658\n\n with torch.no_grad(): # \uae30\uc6b8\uae30 \uacc4\uc0b0 \ube44\ud65c\uc131\ud654\n for images, labels in data_loader:\n images = images.to(device)\n feature = model(images)\n features.append(feature.cpu())\n\n return torch.cat(features)\n\n# \ud2b9\uc131 \ucd94\ucd9c \uc2e4\ud589\nfeatures = extract_features(train_loader)\nprint(\"\ucd94\ucd9c\ub41c \ud2b9\uc131\uc758 \ud06c\uae30:\", features.size())\n<\/code><\/pre>\n2. \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\uc5d0\uc11c \ud2b9\uc131 \ucd94\ucd9c<\/h2>\n
\ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\ub97c \ub2e4\ub8e8\uae30 \uc704\ud574, \uc6b0\ub9ac\ub294 RNN(Recurrent Neural Network)\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud2b9\uc131\uc744 \ucd94\ucd9c\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc790\uc5f0\uc5b4 \ucc98\ub9ac(NLP) \ubd84\uc57c\uc5d0\uc11c \ud754\ud788 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n
2.1 \ub370\uc774\ud130 \uc900\ube44<\/h3>\nfrom torchtext.datasets import AG_NEWS\nfrom torchtext.data import Field, BucketIterator\n\nTEXT = Field(tokenize='spacy', lower=True)\nLABEL = Field(sequential=False)\n\n# AG News \ub370\uc774\ud130\uc14b \ub85c\ub4dc\ntrain_data, test_data = AG_NEWS splits=(TEXT, LABEL))\nTEXT.build_vocab(train_data)\nLABEL.build_vocab(train_data)\n\n# \ub370\uc774\ud130 \ub85c\ub354 \uad6c\ucd95\ntrain_iterator, test_iterator = BucketIterator.splits((train_data, test_data), batch_size=64, device=device)\n<\/code><\/pre>\n2.2 RNN \ubaa8\ub378 \uc815\uc758<\/h3>\nclass RNN(nn.Module):\n def __init__(self, input_dim, embed_dim, hidden_dim, output_dim):\n super().__init__()\n self.embedding = nn.Embedding(input_dim, embed_dim)\n self.rnn = nn.RNN(embed_dim, hidden_dim)\n self.fc = nn.Linear(hidden_dim, output_dim)\n\n def forward(self, text):\n embedded = self.embedding(text)\n output, hidden = self.rnn(embedded)\n return hidden\n\n# \ubaa8\ub378 \uc778\uc2a4\ud134\uc2a4\ud654\ninput_dim = len(TEXT.vocab)\nembed_dim = 100\nhidden_dim = 256\noutput_dim = len(LABEL.vocab)\n\nmodel = RNN(input_dim, embed_dim, hidden_dim, output_dim).to(device)\n<\/code><\/pre>\n2.3 \ud2b9\uc131 \ucd94\ucd9c<\/h3>\n
\uc774\uc81c RNN \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud558\uc5ec \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\uc758 \ud2b9\uc131\uc744 \ucd94\ucd9c\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
def extract_text_features(data_loader):\n text_features = []\n\n model.eval()\n\n with torch.no_grad():\n for batch in data_loader:\n text, labels = batch.text\n text = text.to(device)\n hidden = model(text)\n text_features.append(hidden.cpu())\n\n return torch.cat(text_features)\n\n# \ud2b9\uc131 \ucd94\ucd9c \uc2e4\ud589\ntext_features = extract_text_features(train_iterator)\nprint(\"\ucd94\ucd9c\ub41c \ud14d\uc2a4\ud2b8 \ud2b9\uc131\uc758 \ud06c\uae30:\", text_features.size())\n<\/code><\/pre>\n\uacb0\ub860<\/h2>\n
\uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58\ub97c \uc774\uc6a9\ud558\uc5ec \uc774\ubbf8\uc9c0\uc640 \ud14d\uc2a4\ud2b8 \ub370\uc774\ud130\uc5d0\uc11c \ud2b9\uc131\uc744 \ucd94\ucd9c\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. CNN\uacfc RNN \uac19\uc740 \uad6c\uc870\ub97c \ud65c\uc6a9\ud558\uc5ec \uac01 \ub370\uc774\ud130 \uc720\ud615\uc5d0 \uc801\ud569\ud55c \ud2b9\uc131 \ucd94\ucd9c \ubc29\ubc95\uc744 \uad6c\ud604\ud560 \uc218 \uc788\uc74c\uc744 \ud655\uc778\ud588\uc2b5\ub2c8\ub2e4. \ud2b9\uc131 \ucd94\ucd9c\uc740 \uba38\uc2e0\ub7ec\ub2dd \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ub192\uc774\uace0 \uc6d0\ud65c\ud55c \ub370\uc774\ud130 \ubd84\uc11d\uc744 \uac00\ub2a5\ud558\uac8c \ud558\ub294 \uc911\uc694\ud55c \ub2e8\uacc4\uc785\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub354 \ub098\uc544\uac00 \ub2e4\uc591\ud55c \ubaa8\ub378\uacfc \uae30\ubc95\uc744 \uc5f0\uad6c\ud574 \ubcf4\uae30\ub97c \uad8c\uc7a5\ud569\ub2c8\ub2e4!<\/p>\n
\ud2b9\uc131 \ucd94\ucd9c\uacfc \uad00\ub828\ub41c \uc9c8\ubb38\uc774\ub098 \ub354 \uad81\uae08\ud55c \uc0ac\ud56d\uc774 \uc788\ub2e4\uba74 \ub313\uae00\ub85c \ub0a8\uaca8 \uc8fc\uc138\uc694. \uac10\uc0ac\ud569\ub2c8\ub2e4!<\/p>\n
<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"