{"id":30108,"date":"2024-10-28T03:19:38","date_gmt":"2024-10-28T03:19:38","guid":{"rendered":"http:\/\/atmokpo.com\/w\/?p=30108"},"modified":"2024-11-26T06:49:55","modified_gmt":"2024-11-26T06:49:55","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-%ec%a1%b0%ea%b8%b0-%ec%a2%85%eb%a3%8c%eb%a5%bc-%ec%9d%b4%ec%9a%a9%ed%95%9c-%ec%84%b1%eb%8a%a5-%ec%b5%9c%ec%a0%81","status":"publish","type":"post","link":"https:\/\/atmokpo.com\/w\/30108\/","title":{"rendered":"\ub525\ub7ec\ub2dd \ud30c\uc774\ud1a0\uce58 \uac15\uc88c, \uc870\uae30 \uc885\ub8cc\ub97c \uc774\uc6a9\ud55c \uc131\ub2a5 \ucd5c\uc801\ud654"},"content":{"rendered":"

<\/p>\n

\n \ub525\ub7ec\ub2dd \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\ub294 \uacfc\uc815\uc5d0\uc11c \uacfc\uc801\ud569(overfitting)\uc740 \ud754\ud788 \ubc1c\uc0dd\ud558\ub294 \ubb38\uc81c \uc911 \ud558\ub098\uc785\ub2c8\ub2e4. \uacfc\uc801\ud569\uc774\ub780, \ubaa8\ub378\uc774 \ud6c8\ub828 \ub370\uc774\ud130\uc5d0 \uc9c0\ub098\uce58\uac8c \uc801\ud569\ud558\uc5ec \uc0c8\ub85c\uc6b4 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc77c\ubc18\ud654 \ub2a5\ub825\uc774 \ub5a8\uc5b4\uc9c0\ub294 \ud604\uc0c1\uc744 \ub9d0\ud569\ub2c8\ub2e4. \uadf8\ub798\uc11c \ub9ce\uc740 \uc5f0\uad6c\uc790\uc640 \uc5d4\uc9c0\ub2c8\uc5b4\uac00 \ub2e4\uc591\ud55c \ubc29\ubc95\uc744 \ud1b5\ud574 \uacfc\uc801\ud569\uc744 \ubc29\uc9c0\ud558\uace0\uc790 \ub178\ub825\ud569\ub2c8\ub2e4. \uadf8 \uc911 \ud558\ub098\uac00 \ubc14\ub85c ‘\uc870\uae30 \uc885\ub8cc(Early Stopping)’\uc785\ub2c8\ub2e4.\n <\/p>\n

\uc870\uae30 \uc885\ub8cc(Early Stopping)\uc774\ub780?<\/h2>\n

\n \uc870\uae30 \uc885\ub8cc\ub294 \ubaa8\ub378\uc758 \ud6c8\ub828 \uacfc\uc815\uc744 \ubaa8\ub2c8\ud130\ub9c1\ud558\uc5ec \uac80\uc99d \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc131\ub2a5\uc774 \uac1c\uc120\ub418\uc9c0 \uc54a\uc744 \ub54c \ud6c8\ub828\uc744 \uc911\ub2e8\ud558\ub294 \uae30\ubc95\uc785\ub2c8\ub2e4. \uc774 \ubc29\ubc95\uc740 \ud6c8\ub828 \ub370\uc774\ud130\uc5d0\uc11c \ubaa8\ub378\uc774 \uc131\uacf5\uc801\uc73c\ub85c \ud559\uc2b5\ud588\ub354\ub77c\ub3c4 \uac80\uc99d \ub370\uc774\ud130\uc5d0\uc11c \uc131\ub2a5\uc774 \ub5a8\uc5b4\uc9c0\uba74 \ud6c8\ub828\uc744 \uba48\ucda4\uc73c\ub85c\uc368 \uacfc\uc801\ud569\uc744 \ubc29\uc9c0\ud569\ub2c8\ub2e4.\n <\/p>\n

\uc870\uae30 \uc885\ub8cc\uc758 \uc791\ub3d9 \uc6d0\ub9ac<\/h2>\n

\n \uc870\uae30 \uc885\ub8cc\ub294 \uae30\ubcf8\uc801\uc73c\ub85c \ubaa8\ub378 \ud6c8\ub828 \uc911 \uac80\uc99d \uc190\uc2e4(validation loss) \ub610\ub294 \uac80\uc99d \uc815\ud655\ub3c4(validation accuracy)\ub97c \uad00\ucc30\ud558\uba70, \uc77c\uc815 \uc5d0\ud3ed(epoch) \ub3d9\uc548 \uc131\ub2a5 \ud5a5\uc0c1\uc774 \uc5c6\uc744 \uacbd\uc6b0 \ud6c8\ub828\uc744 \uc911\uc9c0\ud569\ub2c8\ub2e4. \uc774\ub54c, \ucd5c\uc0c1\uc758 \ubaa8\ub378 \ud30c\ub77c\ubbf8\ud130\ub97c \uc800\uc7a5\ud574\ub450\uc5b4, \ud6c8\ub828\uc774 \ub05d\ub09c \ud6c4 \uc774 \ubaa8\ub378\uc744 \uc0ac\uc6a9\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4.\n <\/p>\n

\uc870\uae30 \uc885\ub8cc\uc758 \uad6c\ud604<\/h2>\n

\n \uc5ec\uae30\uc11c\ub294 PyTorch\ub97c \ud65c\uc6a9\ud558\uc5ec \uac04\ub2e8\ud55c \uc774\ubbf8\uc9c0 \ubd84\ub958 \ubaa8\ub378\uc744 \ud6c8\ub828\uc2dc\ud0a4\ub294 \uc608\uc81c\ub97c \ud1b5\ud574 \uc870\uae30 \uc885\ub8cc\ub97c \uad6c\ud604\ud574 \ubcf4\uaca0\uc2b5\ub2c8\ub2e4. \uc774 \uc608\uc81c\uc5d0\uc11c\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\uc5ec \uc190\uae00\uc528 \uc22b\uc790\ub97c \uc778\uc2dd\ud558\ub294 \ubaa8\ub378\uc744 \ud559\uc2b5\ud569\ub2c8\ub2e4.\n <\/p>\n

\ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac \uc124\uce58<\/h3>\n
pip install torch torchvision matplotlib numpy<\/code><\/pre>\n
\ucf54\ub4dc \uc608\uc81c<\/h3>\n
\n        \uc544\ub798\ub294 \uc870\uae30 \uc885\ub8cc\ub97c \uc801\uc6a9\ud55c PyTorch \ucf54\ub4dc \uc608\uc81c\uc785\ub2c8\ub2e4.\n    <\/p>\n
import torch\nimport torch.nn as nn\nimport torch.optim as optim\nimport torchvision\nimport torchvision.transforms as transforms\nfrom torch.utils.data import DataLoader\n\n# \ud558\uc774\ud37c\ud30c\ub77c\ubbf8\ud130 \uc124\uc815\ninput_size = 28 * 28  # MNIST \uc774\ubbf8\uc9c0 \ud06c\uae30\nnum_classes = 10  # \ubd84\ub958\ud560 \ud074\ub798\uc2a4 \uc218\nnum_epochs = 20  # \uc804\uccb4 \ud559\uc2b5 \uc5d0\ud3ed\nbatch_size = 100  # \ubc30\uce58 \ud06c\uae30\nlearning_rate = 0.001  # \ud559\uc2b5\ub960\n\n# MNIST \ub370\uc774\ud130\uc14b \ub85c\ub4dc\ntransform = transforms.Compose([\n    transforms.ToTensor(),\n    transforms.Normalize((0.5,), (0.5,))\n])\n\ntrain_dataset = torchvision.datasets.MNIST(root='.\/data', train=True, transform=transform, download=True)\ntrain_loader = DataLoader(dataset=train_dataset, batch_size=batch_size, shuffle=True)\n\ntest_dataset = torchvision.datasets.MNIST(root='.\/data', train=False, transform=transform, download=True)\ntest_loader = DataLoader(dataset=test_dataset, batch_size=batch_size, shuffle=False)\n\n# \ub2e8\uc21c\ud55c \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(input_size, 128)\n        self.fc2 = nn.Linear(128, num_classes)\n\n    def forward(self, x):\n        x = x.view(-1, input_size)  # \uc774\ubbf8\uc9c0\uc758 \ucc28\uc6d0 \ubcc0\uacbd\n        x = torch.relu(self.fc1(x))  # \ud65c\uc131\ud654 \ud568\uc218\n        x = self.fc2(x)\n        return x\n\n# \ubaa8\ub378, \uc190\uc2e4 \ud568\uc218 \ubc0f \uc635\ud2f0\ub9c8\uc774\uc800 \ucd08\uae30\ud654\nmodel = SimpleNN()\ncriterion = nn.CrossEntropyLoss()\noptimizer = optim.Adam(model.parameters(), lr=learning_rate)\n\n# \uc870\uae30 \uc885\ub8cc\ub97c \uc704\ud55c \ubcc0\uc218 \ucd08\uae30\ud654\nbest_loss = float('inf')\npatience, trials = 5, 0  # \ucd5c\ub300 5\ud68c \uc131\ub2a5 \ud5a5\uc0c1 \uc5c6\uc744 \uc2dc \ud6c8\ub828 \uc911\uc9c0\ntrain_losses, val_losses = [], []\n\n# \ud6c8\ub828 \ub8e8\ud504\nfor epoch in range(num_epochs):\n    model.train()  # \ubaa8\ub378\uc744 \ud6c8\ub828 \ubaa8\ub4dc\ub85c \uc804\ud658\n    running_loss = 0.0\n\n    for images, labels in train_loader:\n        optimizer.zero_grad()  # \uae30\uc6b8\uae30 \ucd08\uae30\ud654\n        outputs = model(images)  # \ubaa8\ub378 \uc608\uce21\n        loss = criterion(outputs, labels)  # \uc190\uc2e4 \uacc4\uc0b0\n        loss.backward()  # \uae30\uc6b8\uae30 \uacc4\uc0b0\n        optimizer.step()  # \uac00\uc911\uce58 \uc5c5\ub370\uc774\ud2b8\n\n        running_loss += loss.item()\n\n    avg_train_loss = running_loss \/ len(train_loader)\n    train_losses.append(avg_train_loss)\n\n    # \uac80\uc99d \ub2e8\uacc4\n    model.eval()  # \ubaa8\ub378\uc744 \ud3c9\uac00 \ubaa8\ub4dc\ub85c \uc804\ud658\n    val_loss = 0.0\n\n    with torch.no_grad():  # \uae30\uc6b8\uae30 \uacc4\uc0b0 \ube44\ud65c\uc131\ud654\n        for images, labels in test_loader:\n            outputs = model(images)\n            loss = criterion(outputs, labels)\n            val_loss += loss.item()\n\n    avg_val_loss = val_loss \/ len(test_loader)\n    val_losses.append(avg_val_loss)\n\n    print(f'Epoch [{epoch+1}\/{num_epochs}], Train Loss: {avg_train_loss:.4f}, Valid Loss: {avg_val_loss:.4f}')\n\n    # \uc870\uae30 \uc885\ub8cc \ub85c\uc9c1\n    if avg_val_loss < best_loss:\n        best_loss = avg_val_loss\n        trials = 0  # \uc131\ub2a5 \ud5a5\uc0c1 \uae30\ub85d \ub9ac\uc14b\n        torch.save(model.state_dict(), 'best_model.pth')  # \ucd5c\uc801 \ubaa8\ub378 \uc800\uc7a5\n    else:\n        trials += 1\n        if trials >= patience:  # patience \ub9cc\ud07c \uc131\ub2a5 \ud5a5\uc0c1\uc774 \uc5c6\uc73c\uba74 \ud6c8\ub828 \uc911\uc9c0\n            print(\"Early stopping...\")\n            break\n\n# \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc131\ub2a5 \ud3c9\uac00\nmodel.load_state_dict(torch.load('best_model.pth'))  # \ucd5c\uc801 \ubaa8\ub378 \ub85c\ub4dc\nmodel.eval()  # \ubaa8\ub378\uc744 \ud3c9\uac00 \ubaa8\ub4dc\ub85c \uc804\ud658\ncorrect, total = 0, 0\n\nwith torch.no_grad():\n    for images, labels in test_loader:\n        outputs = model(images)\n        _, predicted = torch.max(outputs.data, 1)  # \ucd5c\ub300 \ud655\ub960 \ud074\ub798\uc2a4 \uc120\ud0dd\n        total += labels.size(0)\n        correct += (predicted == labels).sum().item()\n\nprint(f'Accuracy of the model on the test images: {100 * correct \/ total:.2f}%')<\/code><\/pre>\n
\ucf54\ub4dc \uc124\uba85<\/h2>\n
\n        \uc704 \ucf54\ub4dc\ub294 MNIST \ub370\uc774\ud130\uc14b\uc744 \uc0ac\uc6a9\ud558\ub294 \uac04\ub2e8\ud55c \uc2e0\uacbd\ub9dd \ubaa8\ub378\uc744 \ud6c8\ub828\ud558\ub294 \uacfc\uc815\uc785\ub2c8\ub2e4. \uba3c\uc800 \ud544\uc694\ud55c \ub77c\uc774\ube0c\ub7ec\ub9ac\ub97c import\ud558\uace0, MNIST \ub370\uc774\ud130\uc14b\uc744 \ub85c\ub4dc\ud569\ub2c8\ub2e4. \uadf8\ub9ac\uace0 \ub2e8\uc21c\ud55c \ub450 \uac1c\uc758 Fully Connected Layer\ub85c \uad6c\uc131\ub41c \uc2e0\uacbd\ub9dd\uc744 \uc815\uc758\ud569\ub2c8\ub2e4.\n    <\/p>\n
\n        \uc774\ud6c4, \uc5d0\ud3ed\ub9c8\ub2e4 \ud6c8\ub828 \uc190\uc2e4\uacfc \uac80\uc99d \uc190\uc2e4\uc744 \uacc4\uc0b0\ud558\uace0, \uc870\uae30 \uc885\ub8cc \ub85c\uc9c1\uc744 \ud1b5\ud574 \uac80\uc99d \uc190\uc2e4\uc774 \uac1c\uc120\ub418\uc9c0 \uc54a\uc73c\uba74 \ud6c8\ub828\uc744 \uc911\ub2e8\ud569\ub2c8\ub2e4. \ub9c8\uc9c0\ub9c9\uc73c\ub85c \ud14c\uc2a4\ud2b8 \ub370\uc774\ud130\uc5d0 \ub300\ud55c \uc815\ud655\ub3c4\ub97c \uacc4\uc0b0\ud558\uc5ec \ubaa8\ub378\uc758 \uc131\ub2a5\uc744 \ud3c9\uac00\ud569\ub2c8\ub2e4.\n    <\/p>\n
\uacb0\ub860<\/h2>\n
\n        \uc870\uae30 \uc885\ub8cc\ub294 \ub525\ub7ec\ub2dd \ubaa8\ub378\uc758 \uc131\ub2a5 \ucd5c\uc801\ud654\ub97c \uc704\ud55c \uc720\uc6a9\ud55c \uae30\uc220\uc785\ub2c8\ub2e4. \uc774\ub97c \ud1b5\ud574 \uacfc\uc801\ud569\uc744 \ubc29\uc9c0\ud558\uace0 \ucd5c\uc801\uc758 \ubaa8\ub378\uc744 \ub3c4\ucd9c\ud560 \uc218 \uc788\uc2b5\ub2c8\ub2e4. \ubcf8 \uac15\uc88c\uc5d0\uc11c\ub294 PyTorch\ub97c \uc774\uc6a9\ud558\uc5ec \uc870\uae30 \uc885\ub8cc\ub97c \uad6c\ud604\ud558\uc5ec MNIST \ubd84\ub958 \ubb38\uc81c\ub97c \ud574\uacb0\ud574\ubcf4\uc558\uc2b5\ub2c8\ub2e4. \uc774\ub97c \ubc14\ud0d5\uc73c\ub85c \ub2e4\uc591\ud55c \ub525\ub7ec\ub2dd \ubb38\uc81c\uc5d0 \uc870\uae30 \uc885\ub8cc \uae30\uc220\uc744 \uc801\uc6a9\ud574 \ubcf4\uc2dc\uae30\ub97c \uad8c\uc7a5\ud569\ub2c8\ub2e4.\n    <\/p>\n
\ucc38\uace0\ubb38\ud5cc<\/h2>\n