一、K折交叉验证
将训练集分成K份，一份做验证集，其他做测试集。这K份都有机会做验证集
二、代码
1 import torch 2 import torch.nn as nn 3 import torchvision 4 from torchvision import datasets,transforms 5 from torch.nn import functional as F 6 import torch.optim as optim 7 8 9 batch_size = 200 10 learning_rate = 1e-2 11 epochs = 10 12 train_db = datasets.MNIST('datasets/mnist_data', 13 train=True, 14 download=True, 15 transform=torchvision.transforms.Compose([ 16 torchvision.transforms.ToTensor(), # 数据类型转化 17 torchvision.transforms.Normalize((0.1307, ), (0.3081, )) # 数据归一化处理 18 ])) 19 20 train_loader = torch.utils.data.DataLoader( 21 train_db, 22 batch_size = batch_size, 23 shuffle = True) 24 25 test_db = datasets.MNIST('datasets/mnist_data/', 26 train=False, 27 download=True, 28 transform=torchvision.transforms.Compose([ 29 torchvision.transforms.ToTensor(), 30 torchvision.transforms.Normalize((0.1307, ), (0.3081, )) 31 ])) 32 33 test_loader = torch.utils.data.DataLoader( 34 test_db, 35 batch_size = batch_size, 36 shuffle = True 37 ) 38 39 print('train:', len(train_db), 'test:', len(test_db)) 40 train_db, val_db = torch.utils.data.random_split(train_db, [50000, 10000]) 41 print('db1:', len(train_db), 'db2:', len(val_db)) 42 train_loader = torch.utils.data.DataLoader( 43 train_db, 44 batch_size=batch_size, shuffle=True) 45 val_loader = torch.utils.data.DataLoader( 46 val_db, 47 batch_size=batch_size, shuffle=True) 48 49 class MLP(nn.Module): 50 51 def __init__(self): 52 super(MLP, self).__init__() 53 54 self.model = nn.Sequential( 55 nn.Linear(784, 200), 56 nn.LeakyReLU(inplace=True), 57 nn.Linear(200, 200), 58 nn.LeakyReLU(inplace=True), 59 nn.Linear(200, 10), 60 nn.LeakyReLU(inplace=True), 61 ) 62 63 def forward(self, x): 64 x = self.model(x) 65 66 return x 67 68 device = torch.device('cuda:0') 69 net = MLP().to(device) 70 optimizer = optim.SGD(net.parameters(), lr=learning_rate) 71 criteon = nn.CrossEntropyLoss().to(device) 72 73 for epoch in range(epochs): 74 75 for batch_idx, (data, target) in enumerate(train_loader): 76 data = data.view(-1, 28*28) 77 data, target = data.to(device), target.cuda() 78 79 logits = net(data) 80 loss = criteon(logits, target) 81 82 optimizer.zero_grad() 83 loss.backward() 84 # print(w1.grad.norm(), w2.grad.norm()) 85 optimizer.step() 86 87 if batch_idx % 100 == 0: 88 print('Train Epoch: {} [{}/{} ({:.0f}%)] Loss: {:.6f}'.format( 89 epoch, batch_idx * len(data), len(train_loader.dataset), 90 100. * batch_idx / len(train_loader), loss.item())) 91 92 93 test_loss = 0 94 correct = 0 95 for data, target in val_loader: 96 data = data.view(-1, 28 * 28) 97 data, target = data.to(device), target.cuda() 98 logits = net(data) 99 test_loss += criteon(logits, target).item() 100 101 pred = logits.data.max(1)[1] 102 correct += pred.eq(target.data).sum() 103 104 test_loss /= len(val_loader.dataset) 105 print(' VAL set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%) '.format( 106 test_loss, correct, len(val_loader.dataset), 107 100. * correct / len(val_loader.dataset))) 108 109 110 111 test_loss = 0 112 correct = 0 113 for data, target in test_loader: 114 data = data.view(-1, 28 * 28) 115 data, target = data.to(device), target.cuda() 116 logits = net(data) 117 test_loss += criteon(logits, target).item() 118 119 pred = logits.data.max(1)[1] 120 correct += pred.eq(target.data).sum() 121 122 test_loss /= len(test_loader.dataset) 123 print(' Test set: Average loss: {:.4f}, Accuracy: {}/{} ({:.0f}%) '.format( 124 test_loss, correct, len(test_loader.dataset), 125 100. * correct / len(test_loader.dataset)))