What parameters should I use to reproduce the pre-training process of ResNet18

[smallbox120]

I use sgd 0.01, and scheduled 30, 60, 90 epoches with lr decay rate 0.1, momentum is set at 0.9 with nesterov enabled, but I can not get the results as good as directly downloaded model, which is 8% less in top-1. Can the owner give me some tips on how to train the pre-trained model?