我想在PyTorch中尝试一些简单的例子，但在训练过程中训练损失并未下降。
这里提供了一些信息：
- 模型是VGG16，由13个卷积层和3个全连接层组成。
- 数据是PyTorch中的CIFAR100数据集。
- 我选择了交叉熵作为损失函数。
代码如下
# encoding: utf-8import torchimport torch.optim as optimimport torch.nn as nnimport torch.nn.functional as Fimport torchvision.transforms as transformsimport torchvisionimport numpy as npclass VGG16(torch.nn.Module): def __init__(self, n_classes): super(VGG16, self).__init__() # construct model self.conv1_1 = nn.Conv2d(3, 64, 3, padding=1) self.conv1_2 = nn.Conv2d(64, 64, 3, padding=1) self.conv2_1 = nn.Conv2d(64, 128, 3, padding=1) self.conv2_2 = nn.Conv2d(128, 128, 3, padding=1) self.conv3_1 = nn.Conv2d(128, 256, 3, padding=1) self.conv3_2 = nn.Conv2d(256, 256, 3, padding=1) self.conv3_3 = nn.Conv2d(256, 256, 3, padding=1) self.conv4_1 = nn.Conv2d(256, 512, 3, padding=1) self.conv4_2 = nn.Conv2d(512, 512, 3, padding=1) self.conv4_3 = nn.Conv2d(512, 512, 3, padding=1) self.conv5_1 = nn.Conv2d(512, 512, 3, padding=1) self.conv5_2 = nn.Conv2d(512, 512, 3, padding=1) self.conv5_3 = nn.Conv2d(512, 512, 3, padding=1) self.fc6 = nn.Linear(512, 512) self.fc7 = nn.Linear(512, 512) self.fc8 = nn.Linear(512, n_classes) def forward(self, x): x = F.relu(self.conv1_1(x)) x = F.relu(self.conv1_2(x)) x = F.max_pool2d(x, (2, 2)) x = F.relu(self.conv2_1(x)) x = F.relu(self.conv2_2(x)) x = F.max_pool2d(x, (2, 2)) x = F.relu(self.conv3_1(x)) x = F.relu(self.conv3_2(x)) x = F.relu(self.conv3_3(x)) x = F.max_pool2d(x, (2, 2)) x = F.relu(self.conv4_1(x)) x = F.relu(self.conv4_2(x)) x = F.relu(self.conv4_3(x)) x = F.max_pool2d(x, (2, 2)) x = F.relu(self.conv5_1(x)) x = F.relu(self.conv5_2(x)) x = F.relu(self.conv5_3(x)) x = F.max_pool2d(x, (2, 2)) x = x.view(-1, self.num_flat_features(x)) x = F.relu(self.fc6(x)) x = F.relu(self.fc7(x)) x = self.fc8(x) return x def num_flat_features(self, x): size = x.size()[1:] num_features = 1 for s in size: num_features *= s return num_featuresif __name__ == '__main__': BATCH_SIZE = 128 LOG_INTERVAL = 5 # data transform = transforms.Compose([ transforms.ToTensor() ]) trainset = torchvision.datasets.CIFAR100( root='./data', train=True, download=True, transform=transform ) testset = torchvision.datasets.CIFAR100( root='./data', train=False, download=True, transform=transform ) trainloader = torch.utils.data.DataLoader(trainset, batch_size=BATCH_SIZE, shuffle=True) testloader = torch.utils.data.DataLoader(testset, batch_size=BATCH_SIZE, shuffle=False) # model vgg16 = VGG16(100) vgg16.cuda() # optimizer optimizer = optim.SGD(vgg16.parameters(), lr=0.01) # loss criterion = nn.CrossEntropyLoss() print('———— Train Start —————') for epoch in range(20): running_loss = 0. for step, (batch_x, batch_y) in enumerate(trainloader): batch_x, batch_y = batch_x.cuda(), batch_y.cuda() # optimizer.zero_grad() output = vgg16(batch_x) loss = criterion(output, batch_y) loss.backward() optimizer.step() running_loss += loss.item() if step % LOG_INTERVAL == 0: print('[%d, %4d] loss: %.4f' % (epoch, step, running_loss / LOG_INTERVAL)) running_loss = 0. def test(): print('———— Test Start ————') correct = 0 total = 0 # with torch.no_grad(): for test_x, test_y in testloader: images, labels = test_x.cuda(), test_y.cuda() output = vgg16(images) _, predicted = torch.max(output.data, 1) total += labels.size(0) correct += (predicted == labels).sum().item() accuracy = 100 * correct / total print('Accuracy of the network is: %.4f %%' % accuracy) print('———— Test Finish ————') test() print('———— Train Finish —————')损失一直保持在4.6060左右，并且从未下降。我尝试了不同的学习率但没有效果。
回答：
我注意到你在卷积层之间没有使用批量归一化。我添加了批量归一化层，似乎有效。以下是修改后的代码：
class VGG16(torch.nn.Module):def __init__(self, n_classes): super(VGG16, self).__init__() # construct model self.conv1_1 = nn.Conv2d(3, 64, 3, padding=1) self.conv11_bn = nn.BatchNorm2d(64) self.conv1_2 = nn.Conv2d(64, 64, 3, padding=1) self.conv12_bn = nn.BatchNorm2d(64) self.conv2_1 = nn.Conv2d(64, 128, 3, padding=1) self.conv21_bn = nn.BatchNorm2d(128) self.conv2_2 = nn.Conv2d(128, 128, 3, padding=1) self.conv22_bn = nn.BatchNorm2d(128) self.conv3_1 = nn.Conv2d(128, 256, 3, padding=1) self.conv31_bn = nn.BatchNorm2d(256) self.conv3_2 = nn.Conv2d(256, 256, 3, padding=1) self.conv32_bn = nn.BatchNorm2d(256) self.conv3_3 = nn.Conv2d(256, 256, 3, padding=1) self.conv33_bn = nn.BatchNorm2d(256) self.conv4_1 = nn.Conv2d(256, 512, 3, padding=1) self.conv41_bn = nn.BatchNorm2d(512) self.conv4_2 = nn.Conv2d(512, 512, 3, padding=1) self.conv42_bn = nn.BatchNorm2d(512) self.conv4_3 = nn.Conv2d(512, 512, 3, padding=1) self.conv43_bn = nn.BatchNorm2d(512) self.conv5_1 = nn.Conv2d(512, 512, 3, padding=1) self.conv51_bn = nn.BatchNorm2d(512) self.conv5_2 = nn.Conv2d(512, 512, 3, padding=1) self.conv52_bn = nn.BatchNorm2d(512) self.conv5_3 = nn.Conv2d(512, 512, 3, padding=1) self.conv53_bn = nn.BatchNorm2d(512) self.fc6 = nn.Linear(512, 512) self.fc7 = nn.Linear(512, 512) self.fc8 = nn.Linear(512, n_classes)def forward(self, x): x = F.relu(self.conv11_bn(self.conv1_1(x))) x = F.relu(self.conv12_bn(self.conv1_2(x))) x = F.max_pool2d(x, (2, 2)) x = F.relu(self.conv22_bn(self.conv2_1(x))) x = F.relu(self.conv21_bn(self.conv2_2(x))) x = F.max_pool2d(x, (2, 2)) x = F.relu(self.conv31_bn(self.conv3_1(x))) x = F.relu(self.conv32_bn(self.conv3_2(x))) x = F.relu(self.conv33_bn(self.conv3_3(x))) x = F.max_pool2d(x, (2, 2)) x = F.relu(self.conv41_bn(self.conv4_1(x))) x = F.relu(self.conv42_bn(self.conv4_2(x))) x = F.relu(self.conv43_bn(self.conv4_3(x))) x = F.max_pool2d(x, (2, 2)) x = F.relu(self.conv51_bn(self.conv5_1(x))) x = F.relu(self.conv52_bn(self.conv5_2(x))) x = F.relu(self.conv53_bn(self.conv5_3(x))) x = F.max_pool2d(x, (2, 2)) x = x.view(-1, self.num_flat_features(x)) x = F.relu(self.fc6(x)) x = F.relu(self.fc7(x)) x = self.fc8(x) return x然而，同样的更优雅的版本可以在这里找到
