为什么添加model.eval后测试集精度波动且较低？

Question

模型中有一个batchnorm层。在添加model.eval之前，准确率稳步上升至89%。添加model.eval后，测试集的准确率和损失变得非常不稳定，最高也只有60多。 这是代码： 另一个版本与上面的不同之处只是没有写 model.eval()

def train(epoch):
    running_loss = 0.0
    correct = 0
    total = 0
    print("---第 {} 轮训练---".format(epoch + 1))
    for batch_idx, data in enumerate(train_loader, 1):
        model.train()
        inputs, target = data
        inputs, target = inputs.to(device), target.to(device)
        optimizer.zero_grad()

        outputs = model(inputs)
        loss = criterion(outputs, target)
        loss.backward()
        predicted = outputs.data.argmax(1)
        total += target.size(0)
        correct = correct + (predicted == target).sum().item()
        optimizer.step()
        running_loss += loss.item()
        if batch_idx % 50== 0:
            print("训练次数：{}，训练loss：{}".format(batch_idx,running_loss/50))
            running_loss = 0.0
    print("训练准确率为：{} %".format(100*correct/total))


def test():
    correct = 0
    total = 0
    test_loss = 0.0
    model.eval()
    with torch.no_grad():
        for batch_idx,data in enumerate(test_loader,1):
            images, labels = data
            images, labels = images.to(device), labels.to(device)
            outputs = model(images)
            loss = criterion(outputs,labels)
            predicted = outputs.data.argmax(1)
            test_loss = test_loss + loss.item()
            # print(predicted)
            # print(labels)
            total += labels.size(0)
            correct += (predicted == labels).sum().item()
            if batch_idx % 12 == 0:
                print("测试{}次，测试集loss为：{}".format(batch_idx,test_loss/12))
                test_loss = 0.0
    print("test accurancy：{} %".format(100*correct/total))
    return correct / total