14 | 构建网络：一站式实现模型搭建与训练

class MyCNN(nn.Module): def __init__(self): super().__init__() self.conv1 = nn.Conv2d(3, 16, kernel_size=3) # conv1输出的特征图为222x222大小 self.fc = nn.Linear(16 * 222 * 222, 10) def forward(self, input): x = self.conv1(input) # 进去全连接层之前，先将特征图铺平 x = x.view(x.shape[0], -1) x = self.fc(x) return x # 尽量使用gpu进行训练,如果没有cpu则使用gpu来训练 device = torch.device("cuda:0" if torch.cuda.is_available() else "cpu") cnn = MyCNN().to(device) transform = transforms.Compose([ # 修改裁剪图片的尺寸 transforms.RandomResizedCrop((224, 224)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) cifar10_dataset = torchvision.datasets.CIFAR10(root='./data', train=False, transform=transform, target_transform=None, download=True) dataloader = DataLoader(dataset=cifar10_dataset, # 传入的数据集, 必须参数 batch_size=32, # 输出的batch大小 shuffle=True, # 数据是否打乱 num_workers=2) # 进程数, 0表示只有主进程 # 定义优化器 optimizer = torch.optim.SGD(cnn.parameters(), lr=1e-4, weight_decay=1e-2, momentum=0.9) steps = 0 for epoch in range(16): for item in dataloader: steps += 1 output = cnn(item[0].to(device)) target = item[1].to(device) # 使用交叉熵损失函数 loss = nn.CrossEntropyLoss()(output, target) # 每100步打印一次loss if steps % 100 == 0: print('Epoch {}, Loss {}'.format(epoch + 1, loss)) cnn.zero_grad() loss.backward() optimizer.step() # 测试分类结果 im = Image.open('data/img.png') input_tensor = transform(im).unsqueeze(0) result = cnn(input_tensor.to(device)).argmax() print(result) # tensor(3, device='cuda:0')

我在使用预训练好的alexnet-owt-7be5be79.pth模型，反复执行下面这段代码的时候 alexnet(input_tensor).argmax() 得到的结果并不总是263，而有时候会得到151和264或者其他的数值，请问我最终应该相信哪一个预测结果呢，是进行多次预测取预测次数最多的那个吗？还是有别的科学的方法呢？

调整之后，可以正常训练了，但是测试数据的时候我发现有些图片分类会有问题，比如狗和猫，鹿和马就容易分类错误，这是因为欠拟合吗？应该如何优化？

老师请教个问题，为什么改全连接后，训练出来的模型，总是输出的是5。而未做改动前，预测出来的是263.源码如下： alexnet = models.alexnet() alexnet.load_state_dict(torch.load('./model/alexnet-owt-7be5be79.pth')) fc_in_features = alexnet.classifier[6].in_features alexnet.classifier[6] = torch.nn.Linear(fc_in_features, 10) print(alexnet) transform = transforms.Compose([ transforms.RandomResizedCrop((224,224)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) cifar10_dataset = torchvision.datasets.CIFAR10(root='./data', train=False, transform=transform, target_transform=None, download=False) dataloader = DataLoader(dataset=cifar10_dataset, batch_size=32, shuffle=True, num_workers=4) optimizer = torch.optim.SGD(alexnet.parameters(), lr=1e-4, weight_decay=1e-2, momentum=0.9) for epoch in range(3): for item in dataloader: output = alexnet(item[0]) target = item[1] loss = nn.CrossEntropyLoss()(output, target) print('Epoch {}, Loss {}'.format(epoch + 1 , loss)) alexnet.zero_grad() loss.backward() optimizer.step() im = Image.open('dog.jpg') transform = transforms.Compose([ transforms.Resize((224,224)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) input_tensor = transform(im).unsqueeze(0) alexnet.eval() print(alexnet(input_tensor).argmax())

hi，老师，微调的时候，如果这样写，参数都不更新了，那还有哪部分参数会被训练呢？是需要再接一层全连接层么？ alexnet = models.alexnet() alexnet.load_state_dict(torch.load('./model/alexnet-owt-4df8aa71.pth')) for param in alexnet.parameters(): param.requires_grad = False

思考： 使用 `nn.CrossEntropyLoss` 作为 loss function 时，会自动在网络最后添加 `nn.LogSoftmax` 和 `nn.nLLLos`，因此不用再在 fc 层后面手动添加 Softmax 层； 问题： transform 中，标准化的 mean 和 std 是如何确定的（我们需要使用均值为[0.485, 0.456, 0.406]，标准差为[0.229, 0.224, 0.225]对数据进行正规化）？

老师请教一个问题，在梯度清零代码中 13 节课用的是优化器 optimizer.zero_grad() 这节课用的是模型 alexnet.zero_grad() 这两行代码有什么区别吗？

import torchvision.models as models from PIL import Image import torchvision import torchvision.transforms as transforms alexnet = models.alexnet(pretrained=True) im = Image.open('./data/dog.jpg') transform = transforms.Compose([ transforms.RandomResizedCrop((224,224)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) input_tensor = transform(im).unsqueeze(0) print(alexnet(input_tensor).argmax()) 以上代码输出的不一定是264，这个是啥情况？

老师您好：我一步一步按照咱们那个步骤来的，为什么代码运行不了那？谢谢您！ import torch import torch.nn as nn import torch.optim as optim import torchvision.models as models import torchvision.models as models alexnet = models.alexnet(pretrained=True) import torchvision.transforms as transforms transform = transforms.Compose([ transforms.RandomResizedCrop((224,224)), transforms.ToTensor(), transforms.Normalize(mean=[0.485,0.456,0.406],std=[0.229,0.224,0.225]) ]) import torchvision cifar10_dataset = torchvision.datasets.CIFAR10(root='./data',#注：这里是存在 pycharm文件当中的data文件夹里 train=False, transform=transform, target_transform=None, download=True) from torch.utils.data import DataLoader dataloader = DataLoader(dataset=cifar10_dataset, batch_size=32, shuffle=True, num_workers=2) fc_in_features = alexnet.classifier[6].in_features alexnet.classifier[6] = torch.nn.Linear(fc_in_features,10) optimizer = torch.optim.SGD(alexnet.parameters(), lr=1e-4, weight_decay=1e-2, momentum=0.9) for epoch in range(3): # ↓定义比对的元素 for item in dataloader: output = alexnet(item[0]) target = item[1] # ↓使用损失函数 loss = nn.CrossEntropyLoss()(output, target) print('Epoch{},Loss{}'.format(epoch + 1, loss)) # ↓更新损失函数和优化函数 alexnet.zero_grad() loss.backward() optimizer.step()

巨人肩膀@马克图库 import torch import torch.nn as nn import torchvision import torchvision.transforms as transforms from PIL import Image transform = transforms.Compose([ transforms.RandomResizedCrop((224, 224)), transforms.ToTensor(), transforms.Normalize(mean=[0.485, 0.456, 0.406], std=[0.229, 0.224, 0.225]) ]) cifar10_train_dataset = torchvision.datasets.CIFAR10(root='./data',train=True,transform=transform,target_transform=None) train_loader = torch.utils.data.DataLoader(dataset=cifar10_train_dataset,batch_size=32,shuffle=True,num_workers=2) classes = ('plane', 'car', 'bird', 'cat', 'deer', 'dog', 'frog', 'horse', 'ship', 'truck') n_total_steps = len(train_loader) class MyCNN(nn.Module): def __init__(self): super().__init__() self.conv1 = nn.Conv2d(3, 16, kernel_size=3) self.fc = nn.Linear(16 * 222 * 222, 10) def forward(self, input): x = self.conv1(input) x = x.view(x.shape[0], -1) x = self.fc(x) return x device = torch.device('cuda:0' if torch.cuda.is_available() else 'cpu') model = MyCNN().to(device) criterion = nn.CrossEntropyLoss() optimizer = torch.optim.SGD(model.parameters(), lr=1e-4, weight_decay=1e-2, momentum=0.9) print('Start training...') for epoch in range(4): for i, (images, labels) in enumerate(train_loader): images = images.to(device) labels = labels.to(device) labels_pred = model(images) loss = criterion(labels_pred, labels) optimizer.zero_grad() loss.backward() optimizer.step if (i+1) % 100 == 0: print(f'Epoch [{epoch+1}/{4}], Step [{i+1:5d}/{n_total_steps}], Loss={loss.item():.4f}') print('Training complete!') with torch.no_grad(): im = Image.open('./images/dog.jpg') input_tensor = transform(im).unsqueeze(0).to(device) label_pred = model(input_tensor).argmax() print(f'Your label predicted: {classes[label_pred]}')