When I change the model to eval mode, the class prediction will be the same. I tested resnet18 and resnet50.
Any explanation for this result? In my own experiment, I always get the 10% evaluation accuracy on the CIFAR-10 dataset if I set eval mode.