(30) PyTorch\u3067MNIST\uff08GeForce RTX 5070ti\u3067\u9ad8\u901f\u5b9f\u884c\uff09<\/a> \u3067\u4f7f\u7528\u3057\u305f MNIST\u7528\u30b5\u30f3\u30d7\u30eb\u30d7\u30ed\u30b0\u30e9\u30e0<\/td><\/tr>\n<\/table> <\/p>\n\u4e0b\u8a18\u306e\u6570\u70b9\u3060\u3051\u30d7\u30ed\u30b0\u30e9\u30e0\u3092\u5909\u66f4\u3057\u305f\u3002
\n\u30fbGPU or CPU \u4f7f\u7528\u4e2d\u306e\u30c7\u30d0\u30a4\u30b9\u3092\u8868\u793a\u3059\u308b\u3002
\n\u30fb\u6240\u8981\u6642\u9593\u3092\u6e2c\u5b9a\u3001\u8868\u793a\u3059\u308b\u3002
\n\u30fb\u30df\u30cb\u30d0\u30c3\u30c1\u3054\u3068\u306e\u7d50\u679c\u8868\u793a\u3092\u6291\u5236\u3059\u308b\u3002
\n\u30fb\u4f7f\u308f\u306a\u3044\u30b3\u30de\u30f3\u30c9\u30e9\u30a4\u30f3\u5f15\u6570\u3092\u7121\u8996\u3059\u308b\u3002
\n\u30fb\u305d\u306e\u4ed6\u3001\u4e0d\u8981\u306a\u51e6\u7406\u3092\u524a\u9664\u3057\u305f\u3002<\/p>\n
\r\nimport argparse\r\nimport torch\r\nimport torch.nn as nn\r\nimport torch.nn.functional as F\r\nimport torch.optim as optim\r\nfrom torchvision import datasets, transforms\r\nfrom torch.optim.lr_scheduler import StepLR\r\nimport time\r\n\r\nclass Net(nn.Module):\r\n def __init__(self):\r\n super(Net, self).__init__()\r\n self.conv1 = nn.Conv2d(1, 32, 3, 1)\r\n self.conv2 = nn.Conv2d(32, 64, 3, 1)\r\n self.dropout1 = nn.Dropout(0.25)\r\n self.dropout2 = nn.Dropout(0.5)\r\n self.fc1 = nn.Linear(9216, 128)\r\n self.fc2 = nn.Linear(128, 10)\r\n\r\n def forward(self, x):\r\n x = self.conv1(x)\r\n x = F.relu(x)\r\n x = self.conv2(x)\r\n x = F.relu(x)\r\n x = F.max_pool2d(x, 2)\r\n x = self.dropout1(x)\r\n x = torch.flatten(x, 1)\r\n x = self.fc1(x)\r\n x = F.relu(x)\r\n x = self.dropout2(x)\r\n x = self.fc2(x)\r\n output = F.log_softmax(x, dim=1)\r\n return output\r\n\r\ndef train(args, model, device, train_loader, optimizer, epoch):\r\n model.train()\r\n for batch_idx, (data, target) in enumerate(train_loader):\r\n data, target = data.to(device), target.to(device)\r\n optimizer.zero_grad()\r\n output = model(data)\r\n loss = F.nll_loss(output, target)\r\n loss.backward()\r\n optimizer.step()\r\n\r\ndef test(model, device, test_loader):\r\n model.eval()\r\n test_loss = 0\r\n correct = 0\r\n with torch.no_grad():\r\n for data, target in test_loader:\r\n data, target = data.to(device), target.to(device)\r\n output = model(data)\r\n test_loss += F.nll_loss(output, target, reduction='sum').item() # sum up batch loss\r\n pred = output.argmax(dim=1, keepdim=True) # get the index of the max log-probability\r\n correct += pred.eq(target.view_as(pred)).sum().item()\r\n test_loss \/= len(test_loader.dataset)\r\n print('Test set: Average loss: {:.4f}, Accuracy: {}\/{} ({:.2f}%)'.format(\r\n test_loss, correct, len(test_loader.dataset), 100. * correct \/ len(test_loader.dataset)))\r\n\r\ndef main():\r\n parser = argparse.ArgumentParser(description='PyTorch MNIST Example')\r\n parser.add_argument('--epochs', type=int, default=3, metavar='N', help='number of epochs to train (default: 14)')\r\n parser.add_argument('--no-cuda', action='store_true', default=False, help='disables CUDA training')\r\n args = parser.parse_args()\r\n use_cuda = not args.no_cuda and torch.cuda.is_available()\r\n if use_cuda:\r\n device = torch.device("cuda")\r\n else:\r\n device = torch.device("cpu")\r\n print("use device is {}".format(device))\r\n train_kwargs = {'batch_size': 64}\r\n test_kwargs = {'batch_size': 1000}\r\n start_time = time.time()\r\n if use_cuda:\r\n cuda_kwargs = {'num_workers': 1, 'pin_memory': True, 'shuffle': True}\r\n train_kwargs.update(cuda_kwargs)\r\n test_kwargs.update(cuda_kwargs)\r\n transform=transforms.Compose([\r\n transforms.ToTensor(),\r\n transforms.Normalize((0.1307,), (0.3081,))\r\n ])\r\n dataset1 = datasets.MNIST('..\/data', train=True, download=True, transform=transform)\r\n dataset2 = datasets.MNIST('..\/data', train=False, transform=transform)\r\n train_loader = torch.utils.data.DataLoader(dataset1,**train_kwargs)\r\n test_loader = torch.utils.data.DataLoader(dataset2, **test_kwargs)\r\n model = Net().to(device)\r\n optimizer = optim.Adadelta(model.parameters(), lr=1.0)\r\n scheduler = StepLR(optimizer, step_size=1, gamma=0.7)\r\n for epoch in range(1, args.epochs + 1):\r\n train(args, model, device, train_loader, optimizer, epoch)\r\n test(model, device, test_loader)\r\n scheduler.step()\r\n end_time = time.time()\r\n elapsed_time = round(end_time - start_time, 1)\r\n print(f"\u51e6\u7406\u6642\u9593: {elapsed_time} \u79d2")\r\n\r\nif __name__ == '__main__':\r\n main()\r\n<\/pre>\n![\"\"](\"https:\/\/www.dogrow.net\/nnet\/wp-content\/uploads\/2025\/04\/i008.jpg\")
<\/p>\n