How do you use the Cityscape for training?

I use your front-end code to train on the Cityscape dataset, I notice that you add a upsampling layer in dilation10_cityscapes_deploy.prototx, but the front-end_vgg_train_net.prototxt that I have generated don't have this layer, can you tell me the difference?

I am looking forward for your reply, thank you