\n \ucd5c\uadfc \uba87 \ub144\uac04 \ub525\ub7ec\ub2dd \ubd84\uc57c\ub294 \uc5ec\ub7ec \uac00\uc9c0 \uc0c8\ub85c\uc6b4 \uc5f0\uad6c \ubc0f \uae30\uc220 \ubc1c\ub2ec\ub85c \uc778\ud574 \uae09\uc18d\ud558\uac8c \ubc1c\uc804\ud574 \uc654\uc2b5\ub2c8\ub2e4. \uadf8 \uc911\uc5d0\uc11c\ub3c4 \uadf8\ub798\ud504 \uc2e0\uacbd\ub9dd(Graph Neural Networks, GNN)\uc740 \uc810\uc810 \ub354 \ub9ce\uc740 \uad00\uc2ec\uc744 \ubc1b\uace0 \uc788\uc73c\uba70, \uc5ec\ub7ec \ubd84\uc57c\uc5d0\uc11c \uc720\ub9dd\ud55c \uacb0\uacfc\ub97c \ubcf4\uc5ec\uc8fc\uace0 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 GNN\uc758 \uac1c\ub150, \uc791\ub3d9 \ubc29\uc2dd, \uadf8\ub9ac\uace0 \ud30c\uc774\ud1a0\uce58(PyTorch)\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc774\ub97c \uad6c\ud604\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc124\uba85\ud558\uaca0\uc2b5\ub2c8\ub2e4. \uad81\uadf9\uc801\uc73c\ub85c\ub294 GNN\uc774 \uc5b4\ub5a4 \ubb38\uc81c\ub97c \ud574\uacb0\ud558\ub294 \ub370 \uc801\ud569\ud55c\uc9c0\uc5d0 \ub300\ud55c \uae4a\uc740 \uc774\ud574\ub97c \uc81c\uacf5\ud558\ub294 \uac83\uc744 \ubaa9\ud45c\ub85c \ud569\ub2c8\ub2e4.\n <\/p>\n
\n \uadf8\ub798\ud504 \uc2e0\uacbd\ub9dd\uc740 \ube44\uc815\ud615 \ub370\uc774\ud130\uc5d0\uc11c \ub178\ub4dc\uc640 \uc5e3\uc9c0\ub97c \uae30\ubc18\uc73c\ub85c \ud55c \uc2e0\uacbd\ub9dd \uad6c\uc870\uc785\ub2c8\ub2e4. \uc804\ud1b5\uc801\uc778 \uc2e0\uacbd\ub9dd\uacfc\ub294 \ub2ec\ub9ac, GNN\uc740 \uadf8\ub798\ud504\uc758 \uad6c\uc870\ub97c \uace0\ub824\ud558\uc5ec \ub178\ub4dc\uc758 \ud2b9\uc131\uacfc \uc5f0\uacb0\uc131\uc744 \ub3d9\uc2dc\uc5d0 \ud559\uc2b5\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. GNN\uc740 \uc8fc\ub85c \ub178\ub4dc \ubd84\ub958, \ub9c1\ud06c \uc608\uce21, \uadf8\ub798\ud504 \ubd84\ub958\uc640 \uac19\uc740 \uc791\uc5c5\uc5d0 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.\n <\/p>\n
\n GNN\uc758 \uc8fc\uc694 \uad6c\uc131 \uc694\uc18c\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:<\/p>\n
\n GNN\uc740 \uc8fc\ub85c \ub450 \ub2e8\uacc4\ub85c \uc791\ub3d9\ud569\ub2c8\ub2e4:<\/p>\n
\n GNN\uc740 \uc5ec\ub7ec \ubd84\uc57c\uc5d0\uc11c \uc720\uc6a9\ud558\uac8c \uc0ac\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
\n \ubcf8 \uc139\uc158\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c \uadf8\ub798\ud504 \uc2e0\uacbd\ub9dd\uc744 \uad6c\ud604\ud558\ub294 \uacfc\uc815\uc744 \uc124\uba85\ud569\ub2c8\ub2e4. PyTorch Geometric \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c \uc0ac\uc6a9\ud558\uc5ec GNN\uc744 \uad6c\ud604\ud560 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n
\n \uba3c\uc800, PyTorch\uc640 PyTorch Geometric\uc744 \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. \uc544\ub798\uc758 \uba85\ub839\uc5b4\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4:<\/p>\n
pip install torch torchvision torchaudio<\/code><\/pre>\npip install torch-geometric<\/code><\/pre>\n<\/p>\n4.2 \ub370\uc774\ud130\uc14b \uc900\ube44<\/h3>\n
\n PyTorch Geometric\uc5d0\uc11c\ub294 \uc5ec\ub7ec\uac00\uc9c0 \ub370\uc774\ud130\uc14b\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4. \uc5ec\uae30\uc5d0\uc11c\ub294 Cora \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud560 \uac83\uc785\ub2c8\ub2e4. \uc774\ub294 \ub300\ud45c\uc801\uc778 \ub17c\ubb38 \ub124\ud2b8\uc6cc\ud06c \ub370\uc774\ud130\uc14b\uc785\ub2c8\ub2e4. \ub370\uc774\ud130\ub97c \ubd88\ub7ec\uc624\ub294 \ucf54\ub4dc\ub294 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:\n <\/p>\n
\nimport torch\nfrom torch_geometric.datasets import Planetoid\n\n# \ub370\uc774\ud130\uc14b \ubd88\ub7ec\uc624\uae30\ndataset = Planetoid(root='\/tmp\/Cora', name='Cora')\ndata = dataset[0]\n <\/code><\/pre>\n4.3 \uadf8\ub798\ud504 \uc2e0\uacbd\ub9dd \ubaa8\ub378 \uc815\uc758\ud558\uae30<\/h3>\n
\n \uc774\uc81c \uac04\ub2e8\ud55c GNN \ubaa8\ub378\uc744 \uc815\uc758\ud574\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. We will use a Graph Convolutional Network (GCN) as our GNN architecture.\n <\/p>\n
\nimport torch.nn.functional as F\nfrom torch_geometric.nn import GCNConv\n\nclass GCN(torch.nn.Module):\n def __init__(self, num_node_features, num_classes):\n super(GCN, self).__init__()\n self.conv1 = GCNConv(num_node_features, 16)\n self.conv2 = GCNConv(16, num_classes)\n\n def forward(self, data):\n x, edge_index = data.x, data.edge_index\n x = self.conv1(x, edge_index)\n x = F.relu(x)\n x = F.dropout(x, training=self.training)\n x = self.conv2(x, edge_index)\n return F.log_softmax(x, dim=1)\n <\/code><\/pre>\n4.4 \ubaa8\ub378 \ud559\uc2b5\ud558\uae30<\/h3>\n
\n \ubaa8\ub378 \ud559\uc2b5\uc758 \uc8fc\uc694 \ub2e8\uacc4\ub97c \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc6b0\ub9ac\ub294 \uc190\uc2e4 \ud568\uc218\ub85c\ub294 \uad50\ucc28 \uc5d4\ud2b8\ub85c\ud53c \uc190\uc2e4\uc744 \uc0ac\uc6a9\ud558\uace0, \uc635\ud2f0\ub9c8\uc774\uc800\ub85c\ub294 Adam\uc744 \uc120\ud0dd\ud558\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\nmodel = GCN(num_node_features=dataset.num_node_features, num_classes=dataset.num_classes)\noptimizer = torch.optim.Adam(model.parameters(), lr=0.01)\nloss_fn = F.nll_loss\n\ndef train():\n model.train()\n optimizer.zero_grad()\n out = model(data)\n loss = loss_fn(out[data.train_mask], data.y[data.train_mask])\n loss.backward()\n optimizer.step()\n return loss.item()\n <\/code><\/pre>\n4.5 \ubaa8\ub378 \ud3c9\uac00\ud558\uae30<\/h3>\n
\n \ud559\uc2b5\uc774 \ub05d\ub09c \ud6c4 \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud558\ub294 \ubc29\ubc95\uc740 \ub2e4\uc74c\uacfc \uac19\uc2b5\ub2c8\ub2e4:\n <\/p>\n
\ndef test():\n model.eval()\n with torch.no_grad():\n pred = model(data).argmax(dim=1)\n correct = pred[data.test_mask].eq(data.y[data.test_mask]).sum().item()\n acc = correct \/ data.test_mask.sum().item()\n return acc\n\nfor epoch in range(200):\n loss = train()\n if epoch % 10 == 0:\n acc = test()\n print(f'Epoch: {epoch}, Loss: {loss:.4f}, Test Accuracy: {acc:.4f}')\n <\/code><\/pre>\n5. \uacb0\ub860<\/h2>\n
\n \uadf8\ub798\ud504 \uc2e0\uacbd\ub9dd\uc740 \ube44\uc815\ud615 \ub370\uc774\ud130\ub85c\ubd80\ud130 \ubcf5\uc7a1\ud55c \uad6c\uc870\uc801 \uc815\ubcf4\ub97c \ud6a8\uacfc\uc801\uc73c\ub85c \ud559\uc2b5\ud560 \uc218 \uc788\ub294 \uc720\uc6a9\ud55c \ubaa8\ub378\uc785\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 GNN\uc758 \uae30\ubcf8 \uac1c\ub150\uacfc \uc791\ub3d9 \uc6d0\ub9ac, \uadf8\ub9ac\uace0 PyTorch\ub97c \uc0ac\uc6a9\ud55c \uc2e4\uc2b5 \uc608\uc81c\ub97c \uc0b4\ud3b4\ubcf4\uc558\uc2b5\ub2c8\ub2e4. GNN\uc740 \ud2b9\ud788 \uc18c\uc15c \ub124\ud2b8\uc6cc\ud06c \ubd84\uc11d, \ud654\ud559\uc801 \ub370\uc774\ud130 \ubaa8\ub378\ub9c1 \ub4f1 \ub2e4\uc591\ud55c \ubd84\uc57c\uc5d0\uc11c \ud070 \uc7a0\uc7ac\ub825\uc744 \uac00\uc9c0\uace0 \uc788\uc2b5\ub2c8\ub2e4. \uc55e\uc73c\ub85c GNN \uad00\ub828 \uc5f0\uad6c\uac00 \ub354\uc6b1 \ud65c\uc131\ud654\ub418\uace0 \ub354\uc6b1 \ub9ce\uc740 \uc751\uc6a9 \uc0ac\ub840\uac00 \ub4f1\uc7a5\ud560 \uac83\uc73c\ub85c \uae30\ub300\ud569\ub2c8\ub2e4.\n <\/p>\n
6. \ucd94\uac00 \uc790\ub8cc<\/h2>\n