Have someone reproduced the result of DenseNet 40 on CIFAR100 from the original paper?

I have just got the acc of 70.86%, lower than the 75.58% in paper.
RandomCrop and RandomHorizontalFlip are used with SGD which the lr is set to 0.1 and get divided by 10 in epoch 150 and 225 of the all 300 epochs.

That’s quite a low result. I often get acc of 74.3% via PyTorch. With lua Torch it is very easy to reproduce the result. I don’t know why it is so hard to achieve the same acc with PyTorch…