Reload model, accuracy changed

1

I have an interesting observation. If I save a model by using torch.save(model.state_dict(), “models.pth”) and reload it. Then the test accuracy changes. Here is the code I am using

#save the model
CUDA_DEVICE = 1
net = torch.load("wideresnet.pth")
torch.save(net.state_dict(), "wideresnetState.pth")

# reload it
net1 = wrn.WideResNet(depth=28, num_classes = 10, widen_factor=10, dropRate=0.0)
net1 = torch.nn.DataParallel(net1, device_ids=[CUDA_DEVICE]).cuda(CUDA_DEVICE)
net1.load_state_dict(torch.load("wideresnetState.pth"))

The test accuracy of “net” is 96.29% while the the test accuracy of “net1” is 94.99%.

2

Are they both in eval mode?

3

yes, they are. I guess here they just have one mode.

4

net1 over here is likely in .train() mode (they are by default). It will matter for BatchNorm.
You’ll have to set net1 to eval mode using net1.eval()