{"id":30006,"date":"2024-10-28T03:19:07","date_gmt":"2024-10-28T03:19:07","guid":{"rendered":"http:\/\/atmokpo.com\/w\/?p=30006"},"modified":"2024-11-26T06:50:27","modified_gmt":"2024-11-26T06:50:27","slug":"%eb%94%a5%eb%9f%ac%eb%8b%9d-%ed%8c%8c%ec%9d%b4%ed%86%a0%ec%b9%98-%ea%b0%95%ec%a2%8c-%ea%b7%b8%eb%9e%98%ed%94%84%eb%9e%80","status":"publish","type":"post","link":"https:\/\/atmokpo.com\/w\/30006\/","title":{"rendered":"\ub525\ub7ec\ub2dd \ud30c\uc774\ud1a0\uce58 \uac15\uc88c, \uadf8\ub798\ud504\ub780"},"content":{"rendered":"

<\/p>\n

1. \uadf8\ub798\ud504\uc758 \uac1c\ub150<\/h2>\n

\uadf8\ub798\ud504\ub294 \uc810\uacfc \uc120\uc758 \uc9d1\ud569\uc73c\ub85c, \uc810\uc740 \ub178\ub4dc(node)\ub85c, \uc120\uc740 \uc5e3\uc9c0(edge)\ub85c \ud45c\ud604\ub429\ub2c8\ub2e4. \uc774 \uad6c\uc870\ub294 \ub2e4\uc591\ud55c \ub370\uc774\ud130\ub97c \uc2dc\uac01\uc801\uc73c\ub85c \ud45c\ud604\ud560 \uc218 \uc788\ub294 \uac15\ub825\ud55c \ub3c4\uad6c\uc785\ub2c8\ub2e4. \ub525\ub7ec\ub2dd\uc5d0\uc11c\ub294 \uc8fc\ub85c \ub370\uc774\ud130\uc758 \uad00\uacc4\ub97c \ubd84\uc11d\ud558\uac70\ub098 \uc2e0\uacbd\ub9dd\uc758 \uad6c\uc870\ub97c \uc815\uc758\ud558\ub294 \ub370 \uc0ac\uc6a9\ub429\ub2c8\ub2e4.<\/p>\n

2. \ub525\ub7ec\ub2dd\uc5d0\uc11c\uc758 \uadf8\ub798\ud504 \uc0ac\uc6a9<\/h2>\n

\ub525\ub7ec\ub2dd\uc5d0\uc11c\ub294 \uadf8\ub798\ud504\ub97c \ud1b5\ud574 \uacc4\uc0b0 \uacfc\uc815\uc744 \ubaa8\ub378\ub9c1\ud569\ub2c8\ub2e4. \uac01 \ub178\ub4dc\ub294 \ub370\uc774\ud130\ub098 \ubcc0\uc218\ub97c, \uac01 \uc5e3\uc9c0\ub294 \uc774\ub4e4 \uac04\uc758 \ubcc0\ud658(\uc608: \uc5f0\uc0b0)\uc744 \ud45c\ud604\ud569\ub2c8\ub2e4. \ub370\uc774\ud130\uac00 \uc785\ub825\ub418\uba74, \ub2e4\uc591\ud55c \uc5f0\uc0b0\uc744 \uac70\uccd0 \uacb0\uacfc\uc5d0 \ub3c4\ub2ec\ud569\ub2c8\ub2e4. \uc774 \uacfc\uc815\uc740 \ub2e4\uc74c\uc758 \ub450 \uac00\uc9c0 \uc8fc\uc694 \ub2e8\uacc4\ub85c \ub098\ub269\ub2c8\ub2e4:<\/p>\n

    \n
  1. \ud3ec\uc6cc\ub4dc \ud328\uc2a4(Forward Pass)<\/strong>: \uc785\ub825 \ub370\uc774\ud130\uac00 \uc2e0\uacbd\ub9dd\uc744 \ud1b5\uacfc\ud558\uba70 \uacb0\uacfc\ub97c \uc0dd\uc131\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4.<\/li>\n
  2. \ubc31\uc6cc\ub4dc \ud328\uc2a4(Backward Pass)<\/strong>: \uc190\uc2e4 \ud568\uc218\uc5d0 \ub530\ub77c \uac00\uc911\uce58\ub97c \uc5c5\ub370\uc774\ud2b8\ud558\ub294 \uacfc\uc815\uc73c\ub85c, \uc5ed\uc804\ud30c(Backpropagation)\ub77c \ubd88\ub9bd\ub2c8\ub2e4.<\/li>\n<\/ol>\n

    3. \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c\uc758 \uadf8\ub798\ud504<\/h2>\n

    \ud30c\uc774\ud1a0\uce58(PyTorch)\ub294 \ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504(Dynamic Computation Graph)\ub97c \uc9c0\uc6d0\ud569\ub2c8\ub2e4. \uc774\ub294 \uc2e4\ud589 \uc2dc\uc810\uc5d0 \uadf8\ub798\ud504\uac00 \uc0dd\uc131\ub418\uc5b4, \ub2e4\uc774\ub098\ubbf9\ud55c \ub370\uc774\ud130 \ud750\ub984\uc744 \ucc98\ub9ac\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \uadf8\ub7ec\ubbc0\ub85c, \ubaa8\ub378\uc744 \uc815\uc758\ud558\uace0 \ud559\uc2b5\ud558\ub294 \uacfc\uc815\uc774 \uc9c1\uad00\uc801\uc774\uace0 \uc720\uc5f0\ud558\uac8c \uc9c4\ud589\ub429\ub2c8\ub2e4.<\/p>\n

    3.1. Static vs Dynamic<\/h3>\n

    \uc804\ud1b5\uc801\uc73c\ub85c \uc0ac\uc6a9\ub41c \uc815\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504(Static Computation Graph)\uc5d0\uc11c\ub294 \uadf8\ub798\ud504\uac00 \uad6c\ucd95\ub41c \ud6c4 \uace0\uc815\ub41c \ud615\ud0dc\ub85c \uacc4\uc0b0\uc774 \uc9c4\ud589\ub418\uc5c8\uc2b5\ub2c8\ub2e4. \ubc18\uba74, \ud30c\uc774\ud1a0\uce58\uc758 \ub3d9\uc801 \uacc4\uc0b0 \uadf8\ub798\ud504\ub294 \ud544\uc694\uc5d0 \ub530\ub77c \uadf8\ub798\ud504\ub97c \uc0dd\uc131\ud558\uace0 \uc218\uc815\ud560 \uc218 \uc788\uc5b4 \ub354 \ud070 \uc720\uc5f0\uc131\uc744 \uc81c\uacf5\ud569\ub2c8\ub2e4.<\/p>\n

    4. \ud30c\uc774\ud1a0\uce58 \uc608\uc81c<\/h2>\n

    4.1. \uae30\ubcf8\uc801\uc778 \uacc4\uc0b0 \uadf8\ub798\ud504 \ub9cc\ub4e4\uae30<\/h3>\n

    \uc774\ubc88 \uc139\uc158\uc5d0\uc11c\ub294 \uae30\ubcf8\uc801\uc778 \uacc4\uc0b0 \uadf8\ub798\ud504\ub97c \uc0dd\uc131\ud558\uc5ec \uac04\ub2e8\ud55c \ud150\uc11c \uc5f0\uc0b0\uc744 \uc218\ud589\ud574\ubcfc \uac83\uc785\ub2c8\ub2e4.<\/p>\n

    \n
    import torch\n\n# \ud150\uc11c \uc0dd\uc131\nx = torch.tensor([2.0], requires_grad=True)  # requires_grad=True\ub85c \uc124\uc815\ud558\uc5ec \ubbf8\ubd84 \uac00\ub2a5\ud558\uac8c \ub9cc\ub4ed\ub2c8\ub2e4.\ny = x**2 + 3*x + 1  # \uc5f0\uc0b0 \uc218\ud589\n\n# \ud3ec\uc6cc\ub4dc \ud328\uc2a4\nprint(f'y: {y.item()}')  # \uacb0\uacfc \ucd9c\ub825\n\n# \ubc31\uc6cc\ub4dc \ud328\uc2a4\ny.backward()  # y\uc5d0 \ub300\ud55c x\uc758 \ubbf8\ubd84 \uacc4\uc0b0\n\n# \uae30\uc6b8\uae30 \ucd9c\ub825\nprint(f'\uae30\uc6b8\uae30: {x.grad.item()}')  # x\uc5d0 \ub300\ud55c \uae30\uc6b8\uae30 \ucd9c\ub825\n        <\/code><\/pre>\n<\/div>\n
    \uc704 \ucf54\ub4dc\ub294 \uac04\ub2e8\ud55c \ub2e4\ud56d\uc2dd y = x^2 + 3x + 1<\/code>\uc744 \uc815\uc758\ud558\uace0, \uc774\uc758 \ubbf8\ubd84\uc744 \ud1b5\ud574 \uae30\uc6b8\uae30\ub97c \uacc4\uc0b0\ud558\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4. requires_grad=True<\/code>\ub85c \uc124\uc815\ud55c \ud150\uc11c x<\/code>\ub97c \ud1b5\ud574, \uacc4\uc0b0 \uadf8\ub798\ud504\uac00 \uc0dd\uc131\ub418\uace0, backward()<\/code> \uba54\uc11c\ub4dc\ub97c \ud638\ucd9c\ud558\uc5ec \ubbf8\ubd84\uc744 \uc218\ud589\ud569\ub2c8\ub2e4.<\/p>\n
    4.2. \uc2e0\uacbd\ub9dd \uadf8\ub798\ud504 \uc608\uc81c<\/h3>\n
    \uc774\ubc88\uc5d0\ub294 \uc2e0\uacbd\ub9dd\uc744 \uad6c\uc131\ud558\uc5ec MNIST \uc190\uae00\uc528 \uc22b\uc790 \uc778\uc2dd \ub370\uc774\ud130\uc14b\uc744 \ud559\uc2b5\ud558\ub294 \uc608\uc81c\ub97c \uc0b4\ud3b4\ubcf4\uaca0\uc2b5\ub2c8\ub2e4.<\/p>\n
    \n
    import torch\nimport torch.nn as nn\nimport torch.optim as optim\nfrom torchvision import datasets, transforms\n\n# \uc2e0\uacbd\ub9dd \ubaa8\ub378 \uc815\uc758\nclass SimpleNN(nn.Module):\n    def __init__(self):\n        super(SimpleNN, self).__init__()\n        self.fc1 = nn.Linear(28 * 28, 128)  # \uccab \ubc88\uc9f8 \ub808\uc774\uc5b4\n        self.fc2 = nn.Linear(128, 64)  # \ub450 \ubc88\uc9f8 \ub808\uc774\uc5b4\n        self.fc3 = nn.Linear(64, 10)  # \ucd9c\ub825 \ub808\uc774\uc5b4\n\n    def forward(self, x):\n        x = x.view(-1, 28 * 28)  # 2D \uc774\ubbf8\uc9c0\ub97c 1D\ub85c \ubcc0\ud658\n        x = torch.relu(self.fc1(x))  # ReLU \ud65c\uc131\ud654 \ud568\uc218\n        x = torch.relu(self.fc2(x))\n        x = self.fc3(x)  # \ucd5c\uc885 \ucd9c\ub825\n        return x\n\n# \ub370\uc774\ud130 \ub85c\ub4dc \ubc0f \uc804\ucc98\ub9ac\ntransform = transforms.Compose([\n    transforms.ToTensor(),\n    transforms.Normalize((0.5,), (0.5,))\n])\ntrain_data = datasets.MNIST(root='.\/data', train=True, download=True, transform=transform)\ntrain_loader = torch.utils.data.DataLoader(dataset=train_data, batch_size=64, shuffle=True)\n\n# \ubaa8\ub378, \uc190\uc2e4 \ud568\uc218, \uc635\ud2f0\ub9c8\uc774\uc800 \uc815\uc758\nmodel = SimpleNN()\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.Adam(model.parameters(), lr=0.001)\n\n# \ud6c8\ub828\nnum_epochs = 5\nfor epoch in range(num_epochs):\n    for images, labels in train_loader:\n        optimizer.zero_grad()  # \uae30\uc6b8\uae30 \ucd08\uae30\ud654\n        outputs = model(images)  # \ubaa8\ub378\uc5d0 \uc774\ubbf8\uc9c0 \ub370\uc774\ud130 \uc785\ub825\n        loss = criterion(outputs, labels)  # \uc190\uc2e4 \uacc4\uc0b0\n        loss.backward()  # \uc5ed\uc804\ud30c\n        optimizer.step()  # \ud30c\ub77c\ubbf8\ud130 \uc5c5\ub370\uc774\ud2b8\n\n    print(f'Epoch [{epoch+1}\/{num_epochs}], Loss: {loss.item():.4f}')\n        <\/code><\/pre>\n<\/div>\n
    \uc774 \ucf54\ub4dc\ub294 \uac04\ub2e8\ud55c \uc2e0\uacbd\ub9dd\uc744 \uad6c\uc131\ud558\uace0 MNIST \ub370\uc774\ud130\uc14b\uc5d0 \ub300\ud574 \ud559\uc2b5\ud558\ub294 \uc608\uc81c\uc785\ub2c8\ub2e4. forward()<\/code> \uba54\uc11c\ub4dc\ub97c \ud1b5\ud574 \uc815\uc758\ud55c \uc2e0\uacbd\ub9dd\uc758 \uad6c\uc870\uc5d0 \ub530\ub77c \uc785\ub825\ub41c \uc774\ubbf8\uc9c0 \ub370\uc774\ud130\uac00 \ubaa8\ub378\uc744 \ud1b5\uacfc\ud558\uc5ec \uc608\uce21\uac12\uc774 \uc0dd\uc131\ub429\ub2c8\ub2e4. \ube44\uc2b7\ud55c \ubc29\uc2dd\uc73c\ub85c \ubc31\uc6cc\ub4dc \ud328\uc2a4\ub97c \ud1b5\ud574 \uac00\uc911\uce58\uac00 \uc5c5\ub370\uc774\ud2b8\ub429\ub2c8\ub2e4.<\/p>\n
    5. \uacb0\ub860<\/h2>\n
    \ub525\ub7ec\ub2dd\uc5d0\uc11c \uadf8\ub798\ud504\ub294 \ub370\uc774\ud130\ub97c \ucc98\ub9ac\ud558\uace0 \ubaa8\ub378\uc758 \uad6c\uc870\ub97c \uc815\uc758\ud558\ub294 \uc911\uc694\ud55c \uc694\uc18c\uc785\ub2c8\ub2e4. \ud30c\uc774\ud1a0\uce58\ub294 \uc774\uc640 \uac19\uc740 \uadf8\ub798\ud504 \uae30\ubc18\uc758 \uc811\uadfc \ubc29\uc2dd\uc744 \ub3d9\uc801\uc73c\ub85c \ucc98\ub9ac\ud560 \uc218 \uc788\uc5b4 \ub300\uaddc\ubaa8\uc758 \ubcf5\uc7a1\ud55c \ubaa8\ub378\uc744 \ud559\uc2b5\ud560 \ub54c \uc720\uc6a9\ud569\ub2c8\ub2e4.<\/p>\n
    \uc704\uc758 \uc608\uc81c\ub97c \ud1b5\ud574 \uadf8\ub798\ud504\uc758 \uac1c\ub150\uacfc \ud30c\uc774\ud1a0\uce58\uc5d0\uc11c\uc758 \ud65c\uc6a9\ubc95\uc5d0 \ub300\ud55c \uc774\ud574\ub97c \ub192\uc77c \uc218 \uc788\uae30\ub97c \ubc14\ub78d\ub2c8\ub2e4. \uc55e\uc73c\ub85c \ub354 \uc2ec\ud654\ub41c \uc8fc\uc81c\ub85c \uc774\uc5b4\uc9c0\ub294 \ub525\ub7ec\ub2dd \uac15\uc88c\uc5d0\uc11c\ub3c4 \ub9ce\uc740 \ub3c4\uc6c0\uc774 \ub418\uae38 \ubc14\ub78d\ub2c8\ub2e4.<\/p>\n