预训练的pytorch模型在每个时期始终具有相同的验证精度

Question

我正在Pytorch上训练基于vGG16的模型。训练的损失每时每刻都在减少，但验证准确性却保持不变。

  

Epoch [1/5]，Val准确度：0.8498火车损耗：0.4560558137803147   -------------------时代[2/5]，Val准确度：0.8498火车损耗：0.4382643524576648   -------------------时代[3/5]，准确度：0.8498火车损耗：0.4410017436635879   -------------------时代[4/5]，准确度：0.8498火车损耗：0.4410021219067724

     

-------------------纪元[5/5]，Val准确度：0.8498火车损耗：0.4382636753429577

为什么会发生？

有什么解决方案？

我的模型配置：

from torchvision import models
model = models.vgg16(pretrained=True)
for param in model.parameters():
    param.requires_grad = True
model.classifier[6]=nn.Sequential(
                      nn.Linear(4096, 1024),
                      nn.Linear(1024, 1024), 
                      nn.ReLU(), 
                      nn.Dropout(0.5),
                      nn.Linear(1024, 2),                   
                      nn.Softmax(dim=1))
learning_rate = 0.01
criterion = nn.CrossEntropyLoss()
optimizer = torch.optim.SGD(model.parameters(), lr = 0.003, momentum= 0.9)

培训阶段：

from torch.autograd import Variable
model.double().cuda()
epochs = 5

for epoch in range(epochs):
    loss = 0
    correct = 0
    total = len(val_loader)
    losses=[]
    model.train()
    # Train the model
    for (observations, labels) in (train_loader):
        observations = Variable(observations).permute(0,3,1,2).cuda()
        labels = Variable(labels).long().cuda()

        # Forward pass
        optimizer.zero_grad()
        outputs = model(observations)

        # Backward pass
        loss = criterion(outputs, labels)
        loss.backward()

        # Optimize
        optimizer.step()
        losses.append(loss.data.item())
    model.eval()
    # Test the model on the validation data
    for observations, labels in val_loader:
        observations = Variable(observations).permute(0,3,1,2).cuda()
        labels=labels.long().cuda()
        # Forward pass
        outputs = model(observations)
        _, predicted = torch.max(outputs.data, 1)
        correct = correct+(predicted.long().cuda() == labels.long().cuda()).sum()
    accuracy = correct.item() / 273
    print('Epoch [%2d/%2d] , Val Accuracy: %.4f' % (epoch + 1, epochs, accuracy))
    print(np.mean(losses))
    print('-------------------')