\n \ub525\ub7ec\ub2dd\uc740 \uc778\uacf5\uc9c0\ub2a5 \ubc0f \uba38\uc2e0\ub7ec\ub2dd\uc758 \ud55c \ubd84\uc57c\ub85c, \uc0ac\ub78c\uc758 \ub1cc\ub97c \ubaa8\uc0ac\ud55c \uc778\uacf5\uc2e0\uacbd\ub9dd\uc744 \uae30\ubc18\uc73c\ub85c \ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.
\n \ub9ce\uc740 \ub370\uc774\ud130\uc640 \uac15\ub825\ud55c \ucef4\ud4e8\ud305 \ud30c\uc6cc\uc758 \ubc1c\uc804 \ub355\ubd84\uc5d0 \ub525\ub7ec\ub2dd \uae30\uc220\uc774 \uacfc\uac70\ubcf4\ub2e4 \ud6e8\uc52c \ub354 \ub9ce\uc740 \uc8fc\ubaa9\uc744 \ubc1b\uace0 \uc788\uc2b5\ub2c8\ub2e4.
\n \ud2b9\ud788, \ud30c\uc774\ud1a0\uce58\ub294 \ub525\ub7ec\ub2dd \ud504\ub808\uc784\uc6cc\ud06c \uc911 \ud558\ub098\ub85c, \uc0ac\uc6a9\uc758 \uc6a9\uc774\uc131\uacfc \uc720\uc5f0\uc131 \ub355\ubd84\uc5d0 \ub9ce\uc740 \uc5f0\uad6c\uc790\uc640 \uac1c\ubc1c\uc790\ub4e4\uc774 \uc120\ud638\ud558\uace0 \uc788\uc2b5\ub2c8\ub2e4.
\n \uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58\ub97c \uc774\uc6a9\ud55c \ubaa8\ub378 \uc815\uc758\uc5d0 \ub300\ud574 \uae4a\uc774 \uc788\uac8c \uc54c\uc544\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \ub525\ub7ec\ub2dd \ubaa8\ub378\uc740 \uc785\ub825 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc608\uce21\uc744 \uc218\ud589\ud558\uae30 \uc704\ud574 \uc5ec\ub7ec \uce35\uc758 \uc2e0\uacbd\ub9dd\uc744 \uc0ac\uc6a9\ud558\ub294 \uc54c\uace0\ub9ac\uc998\uc785\ub2c8\ub2e4.
\n \ubaa8\ub378\uc740 \uc785\ub825\uce35, \uc740\ub2c9\uce35, \ucd9c\ub825\uce35\uc73c\ub85c \uad6c\uc131\ub418\uba70, \uac01 \uce35\uc740 \uc2e0\uacbd\ub9dd\uc758 \ub178\ub4dc(\ub274\ub7f0)\ub85c \uc774\ub8e8\uc5b4\uc838 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n
\n \uae30\ubcf8\uc801\uc778 \uc2e0\uacbd\ub9dd\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \uc694\uc18c\ub85c \uad6c\uc131\ub429\ub2c8\ub2e4:\n <\/p>\n
\n \ud30c\uc774\ud1a0\uce58\ub97c \uc0ac\uc6a9\ud558\uae30 \uc704\ud574 \uba3c\uc800 \uc124\uce58\ud574\uc57c \ud569\ub2c8\ub2e4. \ud30c\uc774\uc36c \ud328\ud0a4\uc9c0 \uad00\ub9ac \ub3c4\uad6c\uc778 pip\ub97c \uc0ac\uc6a9\ud558\uc5ec \uc124\uce58\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.
\n \uc544\ub798 \uba85\ub839\uc5b4\ub97c \ud130\ubbf8\ub110\uc5d0 \uc785\ub825\ud558\uc5ec \uc124\uce58\ud574 \uc8fc\uc138\uc694:\n <\/p>\n
pip install torch torchvision<\/code><\/pre>\n3. \ubaa8\ub378 \uc815\uc758\ud558\uae30<\/h2>\n
\n \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c \ubaa8\ub378\uc744 \uc815\uc758\ud558\ub294 \ubc29\uc2dd\uc740 \ub9e4\uc6b0 \uc9c1\uad00\uc801\uc785\ub2c8\ub2e4. \ub124\ud2b8\uc6cc\ud06c\uc758 \uad6c\uc870\ub97c \uc815\uc758\ud560 \ub54c\ub294 \uc8fc\ub85c
\n torch.nn.Module<\/code> \ud074\ub798\uc2a4\ub97c \uc0c1\uc18d\ubc1b\uc544 Custom Module\uc744 \ub9cc\ub4dc\ub294 \ubc29\uc2dd\uc73c\ub85c \uc9c4\ud589\ub429\ub2c8\ub2e4.
\n \ub610\ud55c, \ubaa8\ub378\uc744 \uad6c\uc131\ud558\ub294 torch.nn<\/code> \ubaa8\ub4c8\uc744 \ud1b5\ud574 \ub2e4\uc591\ud55c \ub808\uc774\uc5b4\uc640 \uae30\ub2a5\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n3.1 \uac04\ub2e8\ud55c \ubaa8\ub378 \uc608\uc81c<\/h3>\n
\n \uc544\ub798\uc758 \uc608\uc81c\ub294 \uac04\ub2e8\ud55c \ub2e4\uce35 \ud37c\uc149\ud2b8\ub860(MLP) \ubaa8\ub378\uc744 \uc815\uc758\ud558\ub294 \ucf54\ub4dc\uc785\ub2c8\ub2e4. \uc774 \ubaa8\ub378\uc740
\n \uc785\ub825\uc73c\ub85c 784\ucc28\uc6d0 \ubca1\ud130\ub97c \ubc1b\uc544 10\ucc28\uc6d0 \ubca1\ud130(0-9 \uc22b\uc790 \ubd84\ub958)\ub97c \ucd9c\ub825\ud558\ub294 \uad6c\uc870\uc785\ub2c8\ub2e4.\n <\/p>\n
\nimport torch\nimport torch.nn as nn\nimport torch.optim as optim\n\n# \ubaa8\ub378 \uc815\uc758\nclass MLP(nn.Module):\n def __init__(self):\n super(MLP, self).__init__()\n self.fc1 = nn.Linear(784, 128) # \uc785\ub825 \ub808\uc774\uc5b4\n self.fc2 = nn.Linear(128, 64) # \uc740\ub2c9 \ub808\uc774\uc5b4\n self.fc3 = nn.Linear(64, 10) # \ucd9c\ub825 \ub808\uc774\uc5b4\n self.relu = nn.ReLU() # ReLU \ud65c\uc131\ud654 \ud568\uc218\n self.softmax = nn.Softmax(dim=1) # \uc18c\ud504\ud2b8\ub9e5\uc2a4 \ud568\uc218\n\n def forward(self, x):\n x = self.fc1(x) # \uc785\ub825\uce35 -> \uc740\ub2c9\uce351\n x = self.relu(x) # \ud65c\uc131\ud654\n x = self.fc2(x) # \uc740\ub2c9\uce351 -> \uc740\ub2c9\uce352\n x = self.relu(x) # \ud65c\uc131\ud654\n x = self.fc3(x) # \uc740\ub2c9\uce352 -> \ucd9c\ub825\uce35\n return self.softmax(x) # \uc18c\ud504\ud2b8\ub9e5\uc2a4 \uc801\uc6a9\n <\/code><\/pre>\n \uc704\uc758 \ucf54\ub4dc\uc5d0\uc11c, MLP<\/code> \ud074\ub798\uc2a4\ub294 \uc2e0\uacbd\ub9dd \ubaa8\ub378\uc744 \uc815\uc758\ud569\ub2c8\ub2e4. \uadf8 \uc548\uc5d0\ub294 \uc138 \uac1c\uc758 \uc120\ud615 \ub808\uc774\uc5b4\uc640 \ub450 \uac1c\uc758 ReLU \ud65c\uc131\ud654 \ub808\uc774\uc5b4\uac00 \ud3ec\ud568\ub429\ub2c8\ub2e4. \ubaa8\ub378\uc758 forward<\/code> \uba54\uc18c\ub4dc\ub294 \ub370\uc774\ud130\uac00 \ub124\ud2b8\uc6cc\ud06c\ub97c \ud1b5\ud574 \ud750\ub974\ub294 \ubc29\uc2dd\uc744 \uc815\uc758\ud569\ub2c8\ub2e4. <\/p>\n3.2 \ubaa8\ub378 \ud559\uc2b5\ud558\uae30<\/h3>\n
\n \ubaa8\ub378\uc744 \uc815\uc758\ud55c \ud6c4\uc5d0\ub294 \ub370\uc774\ud130\ub97c \ud1b5\ud574 \ud559\uc2b5\uc2dc\ucf1c\uc57c \ud569\ub2c8\ub2e4. \uc774\ub97c \uc704\ud574 \uc190\uc2e4 \ud568\uc218\uc640 \ucd5c\uc801\ud654 \uc54c\uace0\ub9ac\uc998\uc744 \uc124\uc815\ud574\uc57c \ud569\ub2c8\ub2e4.
\n \uc77c\ubc18\uc801\uc73c\ub85c \ub2e4\uc911 \ud074\ub798\uc2a4 \ubd84\ub958\uc758 \uacbd\uc6b0, \uad50\ucc28 \uc5d4\ud2b8\ub85c\ud53c \uc190\uc2e4 \ud568\uc218\uc640 Adam \uc635\ud2f0\ub9c8\uc774\uc800\ub97c \ub9ce\uc774 \uc0ac\uc6a9\ud569\ub2c8\ub2e4.\n <\/p>\n
\n# \uc190\uc2e4 \ud568\uc218\uc640 \uc635\ud2f0\ub9c8\uc774\uc800 \uc124\uc815\nmodel = MLP()\ncriterion = nn.CrossEntropyLoss() # \uc190\uc2e4 \ud568\uc218\noptimizer = optim.Adam(model.parameters(), lr=0.001) # Adam \uc635\ud2f0\ub9c8\uc774\uc800\n <\/code><\/pre>\n\n \uc774\ud6c4, \ubaa8\ub378\uc744 \ud559\uc2b5\uc2dc\ud0a4\uae30 \uc704\ud574 \ub9ce\uc740 \ubc18\ubcf5\uc744 \uac70\uce58\uac8c \ub429\ub2c8\ub2e4. \uc544\ub798 \ucf54\ub4dc\ub294 \uc804\uccb4 \ud559\uc2b5 \uacfc\uc815\uc758 \uc608\ub97c \ubcf4\uc5ec\uc90d\ub2c8\ub2e4.
\n MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec 10,000\ud68c \ubc18\ubcf5 \ud559\uc2b5\uc744 \uc218\ud589\ud569\ub2c8\ub2e4.
\n \uc8fc\uc758\ud574\uc57c \ud560 \uc810\uc740 \ub370\uc774\ud130\ub97c \ubbf8\ub2c8 \ubc30\uce58 \ub2e8\uc704\ub85c \ub098\ub204\uc5b4 \ud559\uc2b5\uc2dc\ud0a4\ub294 \uac83\uc785\ub2c8\ub2e4.\n <\/p>\n
\nfrom torchvision import datasets, transforms\nfrom torch.utils.data import DataLoader\n\n# \ub370\uc774\ud130\uc14b \ubc0f \ub85c\ub354 \uc124\uc815\ntransform = transforms.Compose([transforms.ToTensor(), transforms.Normalize((0.5,), (0.5,))])\ntrain_dataset = datasets.MNIST(root='.\/data', train=True, transform=transform, download=True)\ntrain_loader = DataLoader(dataset=train_dataset, batch_size=64, shuffle=True)\n\n# \ubaa8\ub378 \ud559\uc2b5\nfor epoch in range(5): # 5 \uc5d0\ud3ec\ud06c \ub3d9\uc548 \ud559\uc2b5\n for i, (images, labels) in enumerate(train_loader):\n optimizer.zero_grad() # \uae30\uc6b8\uae30 \ucd08\uae30\ud654\n outputs = model(images.view(-1, 784)) # \uc774\ubbf8\uc9c0\ub97c 784 \ucc28\uc6d0 \ubca1\ud130\ub85c \ud3c9\ud0c4\ud654\n loss = criterion(outputs, labels) # \uc190\uc2e4 \uacc4\uc0b0\n loss.backward() # \uc5ed\uc804\ud30c\n optimizer.step() # \uac00\uc911\uce58 \uc5c5\ub370\uc774\ud2b8\n \n if (i+1) % 100 == 0: # 100\ud130 \ubc30\uce58\ub9c8\ub2e4 \uae30\ub85d\n print(f'\uc5d0\ud3ec\ud06c [{epoch+1}\/5], \ubc30\uce58 [{i+1}\/{len(train_loader)}], \uc190\uc2e4: {loss.item():.4f}')\n <\/code><\/pre>\n4. \ub2e4\uc591\ud55c \ubaa8\ub378 \uad6c\uc870<\/h2>\n
\n \uc704\uc758 \uc608\uc2dc\ub294 \uac04\ub2e8\ud55c \ub2e4\uce35 \ud37c\uc149\ud2b8\ub860 \ubaa8\ub378\uc744 \uc815\uc758\ud588\uc9c0\ub9cc, \uc2e4\uc81c \ub525\ub7ec\ub2dd\uc5d0\uc11c\ub294 \ub2e4\uc591\ud55c \ubaa8\ub378 \uad6c\uc870\uac00 \ud544\uc694\ud569\ub2c8\ub2e4.
\n \ub450 \ubc88\uc9f8 \uc608\uc2dc\ub85c \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd(CNN)\uc744 \uc815\uc758\ud558\ub294 \ubc29\ubc95\uc744 \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.\n <\/p>\n
4.1 \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd(CNN) \uc815\uc758\ud558\uae30<\/h3>\n
\n \uc774\ubbf8\uc9c0 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\ub294 \ub370 \ub110\ub9ac \uc0ac\uc6a9\ub418\ub294 \ud569\uc131\uacf1 \uc2e0\uacbd\ub9dd\uc740 \ub2e4\uc74c\uacfc \uac19\uc740 \uad6c\uc870\ub85c \uc815\uc758\ub429\ub2c8\ub2e4.\n <\/p>\n
\nclass CNN(nn.Module):\n def __init__(self):\n super(CNN, self).__init__()\n self.conv1 = nn.Conv2d(1, 32, kernel_size=3, stride=1, padding=1) # \ud569\uc131\uacf1 \ub808\uc774\uc5b4\n self.conv2 = nn.Conv2d(32, 64, kernel_size=3, stride=1, padding=1)\n self.pool = nn.MaxPool2d(kernel_size=2, stride=2) # \ub9e5\uc2a4\ud480\ub9c1 \ub808\uc774\uc5b4\n self.fc1 = nn.Linear(64 * 7 * 7, 128) # \uc120\ud615 \ub808\uc774\uc5b4\n self.fc2 = nn.Linear(128, 10)\n\n def forward(self, x):\n x = self.pool(F.relu(self.conv1(x))) # \uccab \ubc88\uc9f8 \ud569\uc131\uacf1 + \ud65c\uc131\ud654 + \ud480\ub9c1\n x = self.pool(F.relu(self.conv2(x))) # \ub450 \ubc88\uc9f8 \ud569\uc131\uacf1 + \ud65c\uc131\ud654 + \ud480\ub9c1\n x = x.view(-1, 64 * 7 * 7) # \ud3c9\ud0c4\ud654\n x = F.relu(self.fc1(x)) # \uc120\ud615 \ub808\uc774\uc5b4 + \ud65c\uc131\ud654\n x = self.fc2(x) # \ucd9c\ub825 \ub808\uc774\uc5b4\n return x\n <\/code><\/pre>\n5. \uacb0\ub860<\/h2>\n
\n \uc774\ubc88 \ud3ec\uc2a4\ud2b8\uc5d0\uc11c\ub294 \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c\uc758 \ubaa8\ub378 \uc815\uc758 \ubc29\ubc95\uc5d0 \ub300\ud574 \uc54c\uc544\ubcf4\uc558\uc2b5\ub2c8\ub2e4.
\n \uc704\uc5d0\uc11c \ub2e4\ub8ec \ucf54\ub4dc\ub294 \uae30\ubcf8\uc801\uc778 \uc2e0\uacbd\ub9dd\ubd80\ud130 CNN\uc5d0 \uc774\ub974\uae30\uae4c\uc9c0 \ub2e4\uc591\ud558\uac8c \uc751\uc6a9\ub420 \uc218 \uc788\uc2b5\ub2c8\ub2e4.
\n \ub525\ub7ec\ub2dd\uc758 \uae30\ucd08\ub97c \uc774\ud574\ud558\uace0, \ubaa8\ub378 \uc815\uc758\uc5d0 \ub300\ud55c \uc774\ud574\uac00 \uae4a\uc5b4\uc9c0\uba74 \ub354 \ubcf5\uc7a1\ud55c \ubb38\uc81c\ub97c \ud574\uacb0\ud560 \uc218 \uc788\ub294 \uae30\ucd08\uac00 \ub420 \uac83\uc785\ub2c8\ub2e4.
\n \uc55e\uc73c\ub85c\ub3c4 \ub525\ub7ec\ub2dd\uc758 \ub2e4\uc591\ud55c \uae30\uc220\ub4e4\uc744 \ubc30\uc6b0\uace0 \uc2e4\uc2b5\ud558\uc5ec \ub9ce\uc740 \uacbd\ud5d8\uc744 \uc313\uc544\ubcf4\uc2dc\uae30 \ubc14\ub78d\ub2c8\ub2e4.\n <\/p>\n
6. \ucc38\uace0 \uc790\ub8cc<\/h2>\n
\n – PYTorch \uacf5\uc2dd \ubb38\uc11c: https:\/\/pytorch.org\/docs\/stable\/index.html<\/a>
\n – \ub525\ub7ec\ub2dd \uc785\ubb38\uc11c: “Deep Learning” by Ian Goodfellow, Yoshua Bengio, Aaron Courville\n <\/p>\n<\/body><\/p>\n","protected":false},"excerpt":{"rendered":"