Hi @AladdinPerzon, thank you for your response!
I seed like this:
torch.manual_seed(seed)
np.random.seed(seed)
random.seed(seed)
# for cuda
torch.cuda.manual_seed_all(seed)
torch.backends.cudnn.deterministic = True
torch.backends.cudnn.benchmark = False
torch.backends.cudnn.enabled = False
I am able to reproduce the results if I load the ResNet from PyTorch and use this one over and over again. But when I use my own implementation of the ResNet with the same architecture and the same number of parameters I don’t get the same results as I did with the loaded one.
In other words, loading and own implementation yield different results although the network is essentially the same…