\ub525\ub7ec\ub2dd\uc740 \uc778\uacf5 \uc2e0\uacbd\ub9dd\uc744 \ud1b5\ud574 \ub370\uc774\ud130\ub97c \ud559\uc2b5\ud558\uace0 \uc608\uce21\ud558\ub294 \uae30\uc220\uc785\ub2c8\ub2e4. \uc774 \uae00\uc5d0\uc11c\ub294 PyTorch\ub97c \uc0ac\uc6a9\ud558\uc5ec \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130\ub97c \uc815\uc758\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc790\uc138\ud788 \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4. PyTorch\ub294 \ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504\ub97c \uc81c\uacf5\ud558\uc5ec \uc5f0\uad6c \ubc0f \ud504\ub85c\ud1a0\ud0c0\uc785 \uac1c\ubc1c\uc5d0 \ub9e4\uc6b0 \uc720\uc6a9\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\uc785\ub2c8\ub2e4. \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130\ub294 \ud559\uc2b5 \uacfc\uc815\uc5d0\uc11c \uc5c5\ub370\uc774\ud2b8\ub418\uba70, \uc2e0\uacbd\ub9dd\uc758 \uc131\ub2a5\uc5d0 \uc9c1\uc811\uc801\uc778 \uc601\ud5a5\uc744 \ubbf8\uce69\ub2c8\ub2e4.<\/p>\n
\ub525\ub7ec\ub2dd \ubaa8\ub378\uc740 \uc77c\ubc18\uc801\uc73c\ub85c \uc785\ub825\uce35, \uc740\ub2c9\uce35, \ucd9c\ub825\uce35\uc73c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4. \uac01 \uce35\uc740 \uc5ec\ub7ec \uac1c\uc758 \ub178\ub4dc(\ud639\uc740 \ub274\ub7f0)\ub85c \uc774\ub8e8\uc5b4\uc838 \uc788\uc73c\uba70, \uac01 \ub178\ub4dc\ub294 \uc774\uc804 \uce35\uc758 \ub178\ub4dc\uc640 \uc5f0\uacb0\ub429\ub2c8\ub2e4. \uc774\ub7ec\ud55c \uc5f0\uacb0 \uac15\ub3c4\uac00 \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130\uc785\ub2c8\ub2e4. \uc77c\ubc18\uc801\uc73c\ub85c \uc6b0\ub9ac\ub294 \ub2e4\uc74c\uacfc \uac19\uc740 \ud30c\ub77c\ubbf8\ud130\ub97c \uc815\uc758\ud569\ub2c8\ub2e4:<\/p>\n
PyTorch\uc5d0\uc11c\ub294 \ubaa8\ub378\uc744 \uc815\uc758\ud560 \ub54c \uc544\ub798 \ucf54\ub4dc\ub294 \ud30c\uc774\ud1a0\uce58\ub85c \uac04\ub2e8\ud55c \ub2e4\uce35 \ud37c\uc149\ud2b8\ub860(MLP) \ubaa8\ub378\uc744 \uc815\uc758\ud558\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 \uc785\ub825\uce35, \ub450 \uac1c\uc758 \uc740\ub2c9\uce35, \uadf8\ub9ac\uace0 \ucd9c\ub825\uce35\uc744 \uac00\uc9c4 \ubaa8\ub378\uc744 \uad6c\ud604\ud569\ub2c8\ub2e4.<\/p>\n \uc704 \ucf54\ub4dc\uc5d0\uc11c \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130\ub294 \ud559\uc2b5 \uc804\uc5d0 \ucd08\uae30\ud654\ub418\uc5b4\uc57c \ud569\ub2c8\ub2e4. \uae30\ubcf8\uc801\uc73c\ub85c \ubcf4\ub2e4 \ub098\uc740 \uc131\ub2a5\uc744 \uc5bb\uae30 \uc704\ud574 \ucd08\uae30\ud654\ub97c \uc798\ud558\ub294 \uac83\uc774 \uc911\uc694\ud569\ub2c8\ub2e4. \ucd08\uae30\ud654 \ud328\ud134\uc5d0 \ub530\ub77c \ubaa8\ub378 \ud559\uc2b5\uc5d0 \ud070 \uc601\ud5a5\uc744 \uc904 \uc218 \uc788\uc73c\uba70, \uc774\ub97c \ud1b5\ud574 \ub9e4 \uc5d0\ud3ed\ub9c8\ub2e4 \ud559\uc2b5\uc774 \ub354 \ube68\ub77c\uc9c8 \uc218\ub3c4 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \ud6c8\ub828 \uacfc\uc815\uc5d0\uc11c \ud30c\ub77c\ubbf8\ud130\ub294 \uc5ed\uc804\ud30c \uc54c\uace0\ub9ac\uc998\uc744 \ud1b5\ud574 \uc5c5\ub370\uc774\ud2b8\ub429\ub2c8\ub2e4. \uc190\uc2e4 \ud568\uc218\uc758 \uae30\uc6b8\uae30\ub97c \uacc4\uc0b0\ud55c \ud6c4, \uc635\ud2f0\ub9c8\uc774\uc800\uac00 \uc774\ub97c \uc0ac\uc6a9\ud574 \uac00\uc911\uce58\uc640 \ud3b8\ud5a5\uc744 \uc5c5\ub370\uc774\ud2b8\ud569\ub2c8\ub2e4.<\/p>\n \ud6c8\ub828\uc744 \uc9c4\ud589\ud558\uba74\uc11c \uc190\uc2e4 \ud568\uc218 \uac12\uc774 \uac10\uc18c\ud558\ub294 \uac83\uc744 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \ubaa8\ub378\uc774 \uc8fc\uc5b4\uc9c4 \ub370\uc774\ud130\uc5d0 \ub9de\ucdb0 \ud30c\ub77c\ubbf8\ud130\ub97c \ud559\uc2b5\ud558\uace0 \uc788\ub2e4\ub294 \uac83\uc744 \uc54c \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n \uc774 \uae00\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc2e0\uacbd\ub9dd \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130\ub97c \uc815\uc758\ud558\ub294 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \ubaa8\ub378 \uad6c\uc870\ub97c \uc815\uc758\ud558\uace0, \uac00\uc911\uce58\uc640 \ud3b8\ud5a5\uc744 \uc124\uc815\ud558\ub294 \ubc29\ubc95\uc744 \ubc30\uc6e0\uc2b5\ub2c8\ub2e4. \ub610\ud55c \ucd08\uae30\ud654 \ubc29\ubc95\uacfc \ud6c8\ub828 \uacfc\uc815\uc5d0\uc11c \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8\uc758 \uc911\uc694\uc131\uc5d0 \ub300\ud574\uc11c\ub3c4 \ub17c\uc758\ud588\uc2b5\ub2c8\ub2e4. \ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \uadf9\ub300\ud654\ud558\uae30 \uc704\ud574\uc11c\ub294 \uc774\ub7ec\ud55c \ud30c\ub77c\ubbf8\ud130 \uc815\uc758\uc640 \uc5c5\ub370\uc774\ud2b8 \uacfc\uc815\uc774 \ud544\uc218\uc801\uc785\ub2c8\ub2e4. \ud30c\uc774\uc36c\uacfc \ud30c\uc774\ud1a0\uce58\ub97c \ud65c\uc6a9\ud558\uc5ec \uc2e4\uc2b5\ud558\uba74\uc11c \uc774\ud574\ub3c4\ub97c \ub192\uc774\uace0, \ub2e4\uc591\ud55c \ubaa8\ub378\uc744 \uc2dc\ub3c4\ud574\ubcf4\uae30\ub97c \uad8c\uc7a5\ud569\ub2c8\ub2e4.<\/p>\n <\/body><\/p>\n","protected":false},"excerpt":{"rendered":"torch.nn.Module<\/code> \ud074\ub798\uc2a4\ub97c \uc0c1\uc18d\ubc1b\uc544\uc57c \ud569\ub2c8\ub2e4. \uc774 \ud074\ub798\uc2a4\ub97c \uc0c1\uc18d\ubc1b\uc544 \uc0ac\uc6a9\uc790 \uc815\uc758 \ubaa8\ub378\uc744 \ub9cc\ub4e4\uba74, forward<\/code> \uba54\uc18c\ub4dc\ub97c \uc815\uc758\ud558\uc5ec \ubaa8\ub378\uc758 \uc21c\uc804\ud30c \uacfc\uc815\uc744 \uad6c\ud604\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n\uc608\uc81c: \uac04\ub2e8\ud55c \uc2e0\uacbd\ub9dd \ubaa8\ub378 \uad6c\ud604<\/h3>\n
\n
\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\n\nclass SimpleNN(nn.Module):\n def __init__(self, input_size, hidden_size1, hidden_size2, output_size):\n super(SimpleNN, self).__init__()\n # \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130 \uc815\uc758\n self.fc1 = nn.Linear(input_size, hidden_size1) # \uccab \ubc88\uc9f8 \uc740\ub2c9\uce35\n self.fc2 = nn.Linear(hidden_size1, hidden_size2) # \ub450 \ubc88\uc9f8 \uc740\ub2c9\uce35\n self.fc3 = nn.Linear(hidden_size2, output_size) # \ucd9c\ub825\uce35\n\n def forward(self, x):\n x = torch.relu(self.fc1(x)) # \uccab \ubc88\uc9f8 \uc740\ub2c9\uce35 \ud65c\uc131\ud654 \ud568\uc218\n x = torch.relu(self.fc2(x)) # \ub450 \ubc88\uc9f8 \uc740\ub2c9\uce35 \ud65c\uc131\ud654 \ud568\uc218\n x = self.fc3(x) # \ucd9c\ub825\uce35\n return x\n\n# \ubaa8\ub378 \uc0dd\uc131\ninput_size = 10\nhidden_size1 = 20\nhidden_size2 = 10\noutput_size = 1\nmodel = SimpleNN(input_size, hidden_size1, hidden_size2, output_size)\n\n# \ubaa8\ub378\uc758 \ud30c\ub77c\ubbf8\ud130 \ud655\uc778\nprint(\"\ubaa8\ub378 \ud30c\ub77c\ubbf8\ud130:\")\nfor param in model.parameters():\n print(param.shape)\n <\/code>\n <\/pre>\nnn.Linear<\/code>\ub97c \uc0ac\uc6a9\ud558\uc5ec \uac01 \uce35\uc758 \uac00\uc911\uce58\uc640 \ud3b8\ud5a5\uc744 \uc790\ub3d9\uc73c\ub85c \ucd08\uae30\ud654\ud569\ub2c8\ub2e4. model.parameters()<\/code> \uba54\uc18c\ub4dc\ub97c \ud1b5\ud574 \ubaa8\ub378\uc758 \ubaa8\ub4e0 \ud30c\ub77c\ubbf8\ud130\ub97c \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uac01 \ud30c\ub77c\ubbf8\ud130\uc758 \ud615\ud0dc\ub294 torch.Size<\/code> \uac1d\uccb4\ub85c \ubc18\ud658\ub418\uc5b4, \uac00\uc911\uce58\uc640 \ud3b8\ud5a5\uc758 \ucc28\uc6d0\uc744 \ud655\uc778\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n\ubaa8\ub378 \ud30c\ub77c\ubbf8\ud130 \ucd08\uae30\ud654<\/h2>\n
nn.Linear<\/code>\ub294 \uac00\uc911\uce58\ub97c \uc815\uaddc \ubd84\ud3ec\ub97c \uc0ac\uc6a9\ud558\uc5ec \ucd08\uae30\ud654\ud558\uc9c0\ub9cc, \ub2e4\ub978 \ucd08\uae30\ud654 \ubc29\ubc95\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uc608\ub97c \ub4e4\uc5b4, He \ucd08\uae30\ud654\ub098 Xavier \ucd08\uae30\ud654 \ubc29\ubc95\uc774 \uc788\uc2b5\ub2c8\ub2e4.<\/p>\n\ucd08\uae30\ud654 \uc608\uc81c<\/h3>\n
\n
\ndef initialize_weights(model):\n for m in model.modules():\n if isinstance(m, nn.Linear):\n nn.init.kaiming_normal_(m.weight) # He \ucd08\uae30\ud654\n nn.init.zeros_(m.bias) # \ud3b8\ud5a5 0\uc73c\ub85c \ucd08\uae30\ud654\n\ninitialize_weights(model)\n <\/code>\n <\/pre>\n\ubaa8\ub378 \ud6c8\ub828 \uacfc\uc815\uc5d0\uc11c\uc758 \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8<\/h2>\n
\ud6c8\ub828 \ucf54\ub4dc \uc608\uc81c<\/h3>\n
\n
\n# \uc190\uc2e4 \ud568\uc218\uc640 \uc635\ud2f0\ub9c8\uc774\uc800 \uc815\uc758\ncriterion = nn.MSELoss() # \ud3c9\uade0 \uc81c\uacf1 \uc624\ucc28 \uc190\uc2e4\noptimizer = optim.Adam(model.parameters(), lr=0.001) # Adam \uc635\ud2f0\ub9c8\uc774\uc800\n\n# \uac00\uc0c1\uc758 \ub370\uc774\ud130 \uc0dd\uc131\nx_train = torch.randn(100, input_size) # \uc785\ub825 \ub370\uc774\ud130\ny_train = torch.randn(100, output_size) # \ubaa9\ud45c \ucd9c\ub825\n\n# \ubaa8\ub378 \ud6c8\ub828\nnum_epochs = 100\nfor epoch in range(num_epochs):\n model.train() # \ubaa8\ub378\uc744 \ud6c8\ub828 \ubaa8\ub4dc\ub85c \uc804\ud658\n\n # \uc21c\uc804\ud30c\n outputs = model(x_train)\n loss = criterion(outputs, y_train)\n\n # \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8\n optimizer.zero_grad() # \uae30\uc6b8\uae30 \ucd08\uae30\ud654\n loss.backward() # \uc5ed\uc804\ud30c\n optimizer.step() # \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8\n\n if (epoch+1) % 10 == 0:\n print(f'Epoch [{epoch+1}\/{num_epochs}], Loss: {loss.item():.4f}')\n <\/code>\n <\/pre>\n\uacb0\ub860<\/h2>\n