Different test accuracy

  1. In train phase, I got 99.41% accuracy on test loader and saved this model.
  2. After train process, I load above model and test -> accuracy: 99.26%

For both train and test phase, I use this test() code.

In test phase, When I remove model.eval() I can get 99.41%.
How can I understand this situation?
In train phase, model.eval() is also applied…then why test_accuracy is different?

def test(model, dataloader):
    s_correct = 0
    s_total = 0
    for batch_idx, (img, label) in enumerate(dataloader):
        img, label = img.cuda(), label.cuda()
        with torch.no_grad():
            outputs = model(img)
        _, predicted = outputs.max(1)
        s_total += label.size(0)
        s_correct += predicted.eq(label).sum().item()

    acc = s_correct / s_total
    return acc