About accuracy

[takaniwa]

Thank you very much for your work on "PIDNet: A Real-time Semantic Segmentation Network Inspired by PID Controllers", I am very impressed. Your explanation of the three modules of PID is brilliant and I have been trying to follow your work recently. However, I'm having trouble reproducing your paper recently, I'm using imagenet pre-trained pidnet-s on four GPUs on the cityspaces training set, and several times I've obtained results around 76%, I'd like to ask you how I can achieve the accuracy that you have in your paper? Thank you very much for your work, looking forward to your reply, sorry if I disturb you!

[zhujingsong]

Me too!!! Since I only have one GPU (GTX3070, 8G), I retrained pidnet-s using pidnet_small_cityscapes_trainval.yaml with batchsize 4. But the results on val set is also lower than the reported 78.8%.

I am wondering has anyone has reproduced the results on Cityscapes???

[SugarsChieh]

i can train it to 78.22% in val,but put the result to cityscapes ,i only got 75.4%