\u4ed4\u7ec6\u5b66\u4e60\u4e0b\u6d4b\u8bd5\u548c\u8bad\u7ec3\u4ee3\u7801\u7684\u903b\u8f91，\u8fd9\u662f\u57fa\u7840，\u8fd9\u4e2a\u4ee3\u7801\u6846\u67b6\u540e\u7eed\u4f1a\u4e00\u76f4\u6cbf\u7528，\u540e\u7eed\u7684\u91cd\u70b9\u6162\u6162\u5c31\u662f\u8f6c\u5411\u6a21\u578b\u5b9a\u4e49\u9636\u6bb5\u4e86\u3002<\/p>\n
# \u5148\u7ee7\u7eed\u4e4b\u524d\u7684\u4ee3\u7801
\nimport torch
\nimport torch.nn as nn
\nimport torch.optim as optim
\nfrom torch.utils.data import DataLoader , Dataset # DataLoader \u662f PyTorch \u4e2d\u7528\u4e8e\u52a0\u8f7d\u6570\u636e\u7684\u5de5\u5177
\nfrom torchvision import datasets, transforms # torchvision \u662f\u4e00\u4e2a\u7528\u4e8e\u8ba1\u7b97\u673a\u89c6\u89c9\u7684\u5e93，datasets \u548c transforms \u662f\u5176\u4e2d\u7684\u6a21\u5757
\nimport matplotlib.pyplot as plt
\nimport warnings
\n# \u5ffd\u7565\u8b66\u544a\u4fe1\u606f
\nwarnings.filterwarnings("ignore")
\n# \u8bbe\u7f6e\u968f\u673a\u79cd\u5b50，\u786e\u4fdd\u7ed3\u679c\u53ef\u590d\u73b0
\ntorch.manual_seed(42)
\ndevice = torch.device("cuda" if torch.cuda.is_available() else "cpu")
\nprint(f"\u4f7f\u7528\u8bbe\u5907: {device}")
\n# 1. \u6570\u636e\u9884\u5904\u7406
\ntransform = transforms.Compose([
\n transforms.ToTensor(), # \u8f6c\u6362\u4e3a\u5f20\u91cf\u5e76\u5f52\u4e00\u5316\u5230[0,1]
\n transforms.Normalize((0.1307,), (0.3081,)) # MNIST\u6570\u636e\u96c6\u7684\u5747\u503c\u548c\u6807\u51c6\u5dee
\n])<\/p>\n
# 2. \u52a0\u8f7dMNIST\u6570\u636e\u96c6
\ntrain_dataset = datasets.MNIST(
\n root='.\/data',
\n train=True,
\n download=True,
\n transform=transform
\n)<\/p>\n
test_dataset = datasets.MNIST(
\n root='.\/data',
\n train=False,
\n transform=transform
\n)<\/p>\n
# 3. \u521b\u5efa\u6570\u636e\u52a0\u8f7d\u5668
\nbatch_size = 64 # \u6bcf\u6279\u5904\u740664\u4e2a\u6837\u672c
\ntrain_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
\ntest_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)<\/p>\n
# 4. \u5b9a\u4e49\u6a21\u578b\u3001\u635f\u5931\u51fd\u6570\u548c\u4f18\u5316\u5668
\nclass MLP(nn.Module):
\n def __init__(self):
\n super(MLP, self).__init__()
\n self.flatten = nn.Flatten() # \u5c0628×28\u7684\u56fe\u50cf\u5c55\u5e73\u4e3a784\u7ef4\u5411\u91cf
\n self.layer1 = nn.Linear(784, 128) # \u7b2c\u4e00\u5c42：784\u4e2a\u8f93\u5165，128\u4e2a\u795e\u7ecf\u5143
\n self.relu = nn.ReLU() # \u6fc0\u6d3b\u51fd\u6570
\n self.layer2 = nn.Linear(128, 10) # \u7b2c\u4e8c\u5c42：128\u4e2a\u8f93\u5165，10\u4e2a\u8f93\u51fa（\u5bf9\u5e9410\u4e2a\u6570\u5b57\u7c7b\u522b）<\/p>\n
def forward(self, x):
\n x = self.flatten(x) # \u5c55\u5e73\u56fe\u50cf
\n x = self.layer1(x) # \u7b2c\u4e00\u5c42\u7ebf\u6027\u53d8\u6362
\n x = self.relu(x) # \u5e94\u7528ReLU\u6fc0\u6d3b\u51fd\u6570
\n x = self.layer2(x) # \u7b2c\u4e8c\u5c42\u7ebf\u6027\u53d8\u6362，\u8f93\u51falogits
\n return x<\/p>\n
# \u521d\u59cb\u5316\u6a21\u578b
\nmodel = MLP()
\nmodel = model.to(device) # \u5c06\u6a21\u578b\u79fb\u81f3GPU（\u5982\u679c\u53ef\u7528）<\/p>\n
# from torchsummary import summary # \u5bfc\u5165torchsummary\u5e93
\n# print("\\\\n\u6a21\u578b\u7ed3\u6784\u4fe1\u606f：")
\n# summary(model, input_size=(1, 28, 28)) # \u8f93\u5165\u5c3a\u5bf8\u4e3aMNIST\u56fe\u50cf\u5c3a\u5bf8<\/p>\n
criterion = nn.CrossEntropyLoss() # \u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570，\u9002\u7528\u4e8e\u591a\u5206\u7c7b\u95ee\u9898
\noptimizer = optim.Adam(model.parameters(), lr=0.001) # Adam\u4f18\u5316\u5668<\/p>\n
5. \u8bad\u7ec3\u6a21\u578b（\u8bb0\u5f55\u6bcf\u4e2a iteration \u7684\u635f\u5931）
\ndef train(model, train_loader, test_loader, criterion, optimizer, device, epochs):
\n model.train() # \u8bbe\u7f6e\u4e3a\u8bad\u7ec3\u6a21\u5f0f<\/p>\n
# \u65b0\u589e：\u8bb0\u5f55\u6bcf\u4e2a iteration \u7684\u635f\u5931
\n all_iter_losses = [] # \u5b58\u50a8\u6240\u6709 batch \u7684\u635f\u5931
\n iter_indices = [] # \u5b58\u50a8 iteration \u5e8f\u53f7（\u4ece1\u5f00\u59cb）<\/p>\n
for epoch in range(epochs):
\n running_loss = 0.0
\n correct = 0
\n total = 0<\/p>\n
for batch_idx, (data, target) in enumerate(train_loader):
\n # enumerate() \u662f Python \u5185\u7f6e\u51fd\u6570，\u7528\u4e8e\u904d\u5386\u53ef\u8fed\u4ee3\u5bf9\u8c61（\u5982\u5217\u8868\u3001\u5143\u7ec4）\u5e76\u540c\u65f6\u83b7\u53d6\u7d22\u5f15\u548c\u503c\u3002
\n # batch_idx：\u5f53\u524d\u6279\u6b21\u7684\u7d22\u5f15（\u4ece 0 \u5f00\u59cb）
\n # (data, target)：\u5f53\u524d\u6279\u6b21\u7684\u6837\u672c\u6570\u636e\u548c\u5bf9\u5e94\u7684\u6807\u7b7e，\u662f\u4e00\u4e2a\u5143\u7ec4，\u8fd9\u662f\u56e0\u4e3adataloader\u5185\u7f6e\u7684getitem\u65b9\u6cd5\u8fd4\u56de\u7684\u662f\u4e00\u4e2a\u5143\u7ec4，\u5305\u542b\u6570\u636e\u548c\u6807\u7b7e\u3002
\n # \u53ea\u9700\u8981\u8bb0\u4f4f\u8fd9\u79cd\u56fa\u5b9a\u5199\u6cd5\u5373\u53ef
\n data, target = data.to(device), target.to(device) # \u79fb\u81f3GPU(\u5982\u679c\u53ef\u7528)<\/p>\n
optimizer.zero_grad() # \u68af\u5ea6\u6e05\u96f6
\n output = model(data) # \u524d\u5411\u4f20\u64ad
\n loss = criterion(output, target) # \u8ba1\u7b97\u635f\u5931
\n loss.backward() # \u53cd\u5411\u4f20\u64ad
\n optimizer.step() # \u66f4\u65b0\u53c2\u6570<\/p>\n
# \u8bb0\u5f55\u5f53\u524d iteration \u7684\u635f\u5931（\u6ce8\u610f：\u8fd9\u91cc\u76f4\u63a5\u4f7f\u7528\u5355 batch \u635f\u5931，\u800c\u975e\u7d2f\u52a0\u5e73\u5747）
\n iter_loss = loss.item()
\n all_iter_losses.append(iter_loss)
\n iter_indices.append(epoch * len(train_loader) + batch_idx + 1) # iteration \u5e8f\u53f7\u4ece1\u5f00\u59cb<\/p>\n
# \u7edf\u8ba1\u51c6\u786e\u7387\u548c\u635f\u5931
\n running_loss += loss.item() #\u5c06loss\u8f6c\u5316\u4e3a\u6807\u91cf\u503c\u5e76\u4e14\u7d2f\u52a0\u5230running_loss\u4e2d，\u8ba1\u7b97\u603b\u635f\u5931
\n _, predicted = output.max(1) # output：\u662f\u6a21\u578b\u7684\u8f93\u51fa（logits），\u5f62\u72b6\u4e3a [batch_size, 10]（MNIST \u6709 10 \u4e2a\u7c7b\u522b）
\n # \u83b7\u53d6\u9884\u6d4b\u7ed3\u679c，max(1) \u8fd4\u56de\u6bcf\u884c（\u5373\u6bcf\u4e2a\u6837\u672c）\u7684\u6700\u5927\u503c\u548c\u5bf9\u5e94\u7684\u7d22\u5f15，\u8fd9\u91cc\u6211\u4eec\u53ea\u9700\u8981\u7d22\u5f15
\n total += target.size(0) # target.size(0) \u8fd4\u56de\u5f53\u524d\u6279\u6b21\u7684\u6837\u672c\u6570\u91cf，\u5373 batch_size，\u7d2f\u52a0\u6240\u6709\u6279\u6b21\u7684\u6837\u672c\u6570，\u6700\u7ec8\u7b49\u4e8e\u8bad\u7ec3\u96c6\u7684\u603b\u6837\u672c\u6570
\n correct += predicted.eq(target).sum().item() # \u8fd4\u56de\u4e00\u4e2a\u5e03\u5c14\u5f20\u91cf，\u8868\u793a\u9884\u6d4b\u662f\u5426\u6b63\u786e，sum() \u8ba1\u7b97\u6b63\u786e\u9884\u6d4b\u7684\u6570\u91cf，item() \u5c06\u7ed3\u679c\u8f6c\u6362\u4e3a Python \u6570\u5b57<\/p>\n
# \u6bcf100\u4e2a\u6279\u6b21\u6253\u5370\u4e00\u6b21\u8bad\u7ec3\u4fe1\u606f（\u53ef\u9009：\u540c\u65f6\u6253\u5370\u5355 batch \u635f\u5931）
\n if (batch_idx + 1) % 100 == 0:
\n print(f'Epoch: {epoch+1}\/{epochs} | Batch: {batch_idx+1}\/{len(train_loader)} '
\n f'| \u5355Batch\u635f\u5931: {iter_loss:.4f} | \u7d2f\u8ba1\u5e73\u5747\u635f\u5931: {running_loss\/(batch_idx+1):.4f}')<\/p>\n
# \u6d4b\u8bd5\u3001\u6253\u5370 epoch \u7ed3\u679c
\n epoch_train_loss = running_loss \/ len(train_loader)
\n epoch_train_acc = 100. * correct \/ total
\n epoch_test_loss, epoch_test_acc = test(model, test_loader, criterion, device)<\/p>\n
print(f'Epoch {epoch+1}\/{epochs} \u5b8c\u6210 | \u8bad\u7ec3\u51c6\u786e\u7387: {epoch_train_acc:.2f}% | \u6d4b\u8bd5\u51c6\u786e\u7387: {epoch_test_acc:.2f}%')<\/p>\n
# \u7ed8\u5236\u6240\u6709 iteration \u7684\u635f\u5931\u66f2\u7ebf
\n plot_iter_losses(all_iter_losses, iter_indices)
\n # \u4fdd\u7559\u539f epoch \u7ea7\u66f2\u7ebf（\u53ef\u9009）
\n # plot_metrics(train_losses, test_losses, train_accuracies, test_accuracies, epochs)<\/p>\n
return epoch_test_acc # \u8fd4\u56de\u6700\u7ec8\u6d4b\u8bd5\u51c6\u786e\u7387 <\/p>\n
\u5f69\u8272\u56fe\u7247\u5199\u6cd5<\/p>\n
import torch
\nimport torch.nn as nn
\nimport torch.optim as optim
\nfrom torchvision import datasets, transforms
\nfrom torch.utils.data import DataLoader
\nimport matplotlib.pyplot as plt
\nimport numpy as np<\/p>\n
# \u8bbe\u7f6e\u4e2d\u6587\u5b57\u4f53\u652f\u6301
\nplt.rcParams["font.family"] = ["SimHei"]
\nplt.rcParams['axes.unicode_minus'] = False # \u89e3\u51b3\u8d1f\u53f7\u663e\u793a\u95ee\u9898<\/p>\n
# 1. \u6570\u636e\u9884\u5904\u7406
\ntransform = transforms.Compose([
\n transforms.ToTensor(), # \u8f6c\u6362\u4e3a\u5f20\u91cf
\n transforms.Normalize((0.5, 0.5, 0.5), (0.5, 0.5, 0.5)) # \u6807\u51c6\u5316\u5904\u7406
\n])<\/p>\n
# 2. \u52a0\u8f7dCIFAR-10\u6570\u636e\u96c6
\ntrain_dataset = datasets.CIFAR10(
\n root='.\/data',
\n train=True,
\n download=True,
\n transform=transform
\n)<\/p>\n
test_dataset = datasets.CIFAR10(
\n root='.\/data',
\n train=False,
\n transform=transform
\n)<\/p>\n
# 3. \u521b\u5efa\u6570\u636e\u52a0\u8f7d\u5668
\nbatch_size = 64
\ntrain_loader = DataLoader(train_dataset, batch_size=batch_size, shuffle=True)
\ntest_loader = DataLoader(test_dataset, batch_size=batch_size, shuffle=False)<\/p>\n
# 4. \u5b9a\u4e49MLP\u6a21\u578b（\u9002\u5e94CIFAR-10\u7684\u8f93\u5165\u5c3a\u5bf8）
\nclass MLP(nn.Module):
\n def __init__(self):
\n super(MLP, self).__init__()
\n self.flatten = nn.Flatten() # \u5c063x32x32\u7684\u56fe\u50cf\u5c55\u5e73\u4e3a3072\u7ef4\u5411\u91cf
\n self.layer1 = nn.Linear(3072, 512) # \u7b2c\u4e00\u5c42：3072\u4e2a\u8f93\u5165，512\u4e2a\u795e\u7ecf\u5143
\n self.relu1 = nn.ReLU()
\n self.dropout1 = nn.Dropout(0.2) # \u6dfb\u52a0Dropout\u9632\u6b62\u8fc7\u62df\u5408
\n self.layer2 = nn.Linear(512, 256) # \u7b2c\u4e8c\u5c42：512\u4e2a\u8f93\u5165，256\u4e2a\u795e\u7ecf\u5143
\n self.relu2 = nn.ReLU()
\n self.dropout2 = nn.Dropout(0.2)
\n self.layer3 = nn.Linear(256, 10) # \u8f93\u51fa\u5c42：10\u4e2a\u7c7b\u522b<\/p>\n
def forward(self, x):
\n # \u7b2c\u4e00\u6b65：\u5c06\u8f93\u5165\u56fe\u50cf\u5c55\u5e73\u4e3a\u4e00\u7ef4\u5411\u91cf
\n x = self.flatten(x) # \u8f93\u5165\u5c3a\u5bf8: [batch_size, 3, 32, 32] \u2192 [batch_size, 3072]<\/p>\n
# \u7b2c\u4e00\u5c42\u5168\u8fde\u63a5 + \u6fc0\u6d3b + Dropout
\n x = self.layer1(x) # \u7ebf\u6027\u53d8\u6362: [batch_size, 3072] \u2192 [batch_size, 512]
\n x = self.relu1(x) # \u5e94\u7528ReLU\u6fc0\u6d3b\u51fd\u6570
\n x = self.dropout1(x) # \u8bad\u7ec3\u65f6\u968f\u673a\u4e22\u5f03\u90e8\u5206\u795e\u7ecf\u5143\u8f93\u51fa<\/p>\n
# \u7b2c\u4e8c\u5c42\u5168\u8fde\u63a5 + \u6fc0\u6d3b + Dropout
\n x = self.layer2(x) # \u7ebf\u6027\u53d8\u6362: [batch_size, 512] \u2192 [batch_size, 256]
\n x = self.relu2(x) # \u5e94\u7528ReLU\u6fc0\u6d3b\u51fd\u6570
\n x = self.dropout2(x) # \u8bad\u7ec3\u65f6\u968f\u673a\u4e22\u5f03\u90e8\u5206\u795e\u7ecf\u5143\u8f93\u51fa<\/p>\n
# \u7b2c\u4e09\u5c42（\u8f93\u51fa\u5c42）\u5168\u8fde\u63a5
\n x = self.layer3(x) # \u7ebf\u6027\u53d8\u6362: [batch_size, 256] \u2192 [batch_size, 10]<\/p>\n
return x # \u8fd4\u56de\u672a\u7ecf\u8fc7Softmax\u7684logits<\/p>\n
# \u68c0\u67e5GPU\u662f\u5426\u53ef\u7528
\ndevice = torch.device("cuda" if torch.cuda.is_available() else "cpu")<\/p>\n
# \u521d\u59cb\u5316\u6a21\u578b
\nmodel = MLP()
\nmodel = model.to(device) # \u5c06\u6a21\u578b\u79fb\u81f3GPU（\u5982\u679c\u53ef\u7528）<\/p>\n
criterion = nn.CrossEntropyLoss() # \u4ea4\u53c9\u71b5\u635f\u5931\u51fd\u6570
\noptimizer = optim.Adam(model.parameters(), lr=0.001) # Adam\u4f18\u5316\u5668<\/p>\n
# 5. \u8bad\u7ec3\u6a21\u578b（\u8bb0\u5f55\u6bcf\u4e2a iteration \u7684\u635f\u5931）
\ndef train(model, train_loader, test_loader, criterion, optimizer, device, epochs):
\n model.train() # \u8bbe\u7f6e\u4e3a\u8bad\u7ec3\u6a21\u5f0f<\/p>\n
# \u8bb0\u5f55\u6bcf\u4e2a iteration \u7684\u635f\u5931
\n all_iter_losses = [] # \u5b58\u50a8\u6240\u6709 batch \u7684\u635f\u5931
\n iter_indices = [] # \u5b58\u50a8 iteration \u5e8f\u53f7<\/p>\n
for epoch in range(epochs):
\n running_loss = 0.0
\n correct = 0
\n total = 0<\/p>\n
for batch_idx, (data, target) in enumerate(train_loader):
\n data, target = data.to(device), target.to(device) # \u79fb\u81f3GPU<\/p>\n
optimizer.zero_grad() # \u68af\u5ea6\u6e05\u96f6
\n output = model(data) # \u524d\u5411\u4f20\u64ad
\n loss = criterion(output, target) # \u8ba1\u7b97\u635f\u5931
\n loss.backward() # \u53cd\u5411\u4f20\u64ad
\n optimizer.step() # \u66f4\u65b0\u53c2\u6570<\/p>\n
# \u8bb0\u5f55\u5f53\u524d iteration \u7684\u635f\u5931
\n iter_loss = loss.item()
\n all_iter_losses.append(iter_loss)
\n iter_indices.append(epoch * len(train_loader) + batch_idx + 1)<\/p>\n
# \u7edf\u8ba1\u51c6\u786e\u7387\u548c\u635f\u5931
\n running_loss += iter_loss
\n _, predicted = output.max(1)
\n total += target.size(0)
\n correct += predicted.eq(target).sum().item()<\/p>\n
# \u6bcf100\u4e2a\u6279\u6b21\u6253\u5370\u4e00\u6b21\u8bad\u7ec3\u4fe1\u606f
\n if (batch_idx + 1) % 100 == 0:
\n print(f'Epoch: {epoch+1}\/{epochs} | Batch: {batch_idx+1}\/{len(train_loader)} '
\n f'| \u5355Batch\u635f\u5931: {iter_loss:.4f} | \u7d2f\u8ba1\u5e73\u5747\u635f\u5931: {running_loss\/(batch_idx+1):.4f}')<\/p>\n
# \u8ba1\u7b97\u5f53\u524depoch\u7684\u5e73\u5747\u8bad\u7ec3\u635f\u5931\u548c\u51c6\u786e\u7387
\n epoch_train_loss = running_loss \/ len(train_loader)
\n epoch_train_acc = 100. * correct \/ total<\/p>\n
# \u6d4b\u8bd5\u9636\u6bb5
\n model.eval() # \u8bbe\u7f6e\u4e3a\u8bc4\u4f30\u6a21\u5f0f
\n test_loss = 0
\n correct_test = 0
\n total_test = 0<\/p>\n
with torch.no_grad():
\n for data, target in test_loader:
\n data, target = data.to(device), target.to(device)
\n output = model(data)
\n test_loss += criterion(output, target).item()
\n _, predicted = output.max(1)
\n total_test += target.size(0)
\n correct_test += predicted.eq(target).sum().item()<\/p>\n
epoch_test_loss = test_loss \/ len(test_loader)
\n epoch_test_acc = 100. * correct_test \/ total_test<\/p>\n
print(f'Epoch {epoch+1}\/{epochs} \u5b8c\u6210 | \u8bad\u7ec3\u51c6\u786e\u7387: {epoch_train_acc:.2f}% | \u6d4b\u8bd5\u51c6\u786e\u7387: {epoch_test_acc:.2f}%')<\/p>\n
# \u7ed8\u5236\u6240\u6709 iteration \u7684\u635f\u5931\u66f2\u7ebf
\n plot_iter_losses(all_iter_losses, iter_indices)<\/p>\n
return epoch_test_acc # \u8fd4\u56de\u6700\u7ec8\u6d4b\u8bd5\u51c6\u786e\u7387<\/p>\n
# 6. \u7ed8\u5236\u6bcf\u4e2a iteration \u7684\u635f\u5931\u66f2\u7ebf
\ndef plot_iter_losses(losses, indices):
\n plt.figure(figsize=(10, 4))
\n plt.plot(indices, losses, 'b-', alpha=0.7, label='Iteration Loss')
\n plt.xlabel('Iteration（Batch\u5e8f\u53f7）')
\n plt.ylabel('\u635f\u5931\u503c')
\n plt.title('\u6bcf\u4e2a Iteration \u7684\u8bad\u7ec3\u635f\u5931')
\n plt.legend()
\n plt.grid(True)
\n plt.tight_layout()
\n plt.show()<\/p>\n
# 7. \u6267\u884c\u8bad\u7ec3\u548c\u6d4b\u8bd5
\nepochs = 20 # \u589e\u52a0\u8bad\u7ec3\u8f6e\u6b21\u4ee5\u83b7\u5f97\u66f4\u597d\u6548\u679c
\nprint("\u5f00\u59cb\u8bad\u7ec3\u6a21\u578b…")
\nfinal_accuracy = train(model, train_loader, test_loader, criterion, optimizer, device, epochs)
\nprint(f"\u8bad\u7ec3\u5b8c\u6210！\u6700\u7ec8\u6d4b\u8bd5\u51c6\u786e\u7387: {final_accuracy:.2f}%")<\/p>\n
# # \u4fdd\u5b58\u6a21\u578b
\n# torch.save(model.state_dict(), 'cifar10_mlp_model.pth')
\n# # print("\u6a21\u578b\u5df2\u4fdd\u5b58\u4e3a: cifar10_mlp_model.pth") <\/p>\n","protected":false},"excerpt":{"rendered":"
\u6587\u7ae0\u6d4f\u89c8\u9605\u8bfb231\u6b21\u3002\u4ed4\u7ec6\u5b66\u4e60\u4e0b\u6d4b\u8bd5\u548c\u8bad\u7ec3\u4ee3\u7801\u7684\u903b\u8f91\uff0c\u8fd9\u662f\u57fa\u7840\uff0c\u8fd9\u4e2a\u4ee3\u7801\u6846\u67b6\u540e\u7eed\u4f1a\u4e00\u76f4\u6cbf\u7528\uff0c\u540e\u7eed\u7684\u91cd\u70b9\u6162\u6162\u5c31\u662f\u8f6c\u5411\u6a21\u578b\u5b9a\u4e49\u9636\u6bb5\u4e86\u3002<\/p>\n","protected":false},"author":2,"featured_media":0,"comment_status":"open","ping_status":"open","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[1],"tags":[50,86],"topic":[],"class_list":["post-41734","post","type-post","status-publish","format-standard","hentry","category-server","tag-50","tag-86"],"yoast_head":"\n