what's the difference between the official pretrained model and your provided one ?

[JosephChenHub]

I tried to reproduce the performance based on resnet18 on Cityscapes, and I set the sample number of one batch as 4000. There is a slight difference between the official pretrained model resnet18-5c106cde.pth and resnet18_v1.pth: mean iou 72.371% of the official one and 73.209% of the provided one. So what's the difference between these two pretrained model? Is the provided one trained on the ImageNet?