Has anyone reached the performance in the paper with 4 gpus in Cityscapes dataset?

[yichen928]

Thank you for the excellent work. When I run your code in cityscapes with 4 GPUs, I achieved PQ: 58.3, SQ: 79.8, RQ: 71.7, which is lower by one than your result in the paper. I have used your config file: upsnet_resnet50_cityscapes_4gpu.yaml. Did I make something wrong?

[ywher]

Thank you for the excellent work. When I run your code in cityscapes with 4 GPUs, I achieved PQ: 58.3, SQ: 79.8, RQ: 71.7, which is lower by one than your result in the paper. I have used your config file: upsnet_resnet50_cityscapes_4gpu.yaml. Did I make something wrong?

Can I ask you some questions? I am also want to reach the performance in the paper with 4 gpus in Cityscapes dataset

[yichen928]

Thank you for the excellent work. When I run your code in cityscapes with 4 GPUs, I achieved PQ: 58.3, SQ: 79.8, RQ: 71.7, which is lower by one than your result in the paper. I have used your config file: upsnet_resnet50_cityscapes_4gpu.yaml. Did I make something wrong?

Can I ask you some questions? I am also want to reach the performance in the paper with 4 gpus in Cityscapes dataset

I still have not reached the performance......

[ywher]

Thank you for the excellent work. When I run your code in cityscapes with 4 GPUs, I achieved PQ: 58.3, SQ: 79.8, RQ: 71.7, which is lower by one than your result in the paper. I have used your config file: upsnet_resnet50_cityscapes_4gpu.yaml. Did I make something wrong? Can I ask you some questions? I am also want to reach the performance in the paper with 4 gpus in Cityscapes dataset

I still have not reached the performance...... Still thanks，so your PQ is from using the model the author supply or the model trained by yourself on the default params?

[txfs1926]

Hi, with the default Cityscapes resnet 50 4 GPUs setting, I only harvested: 2020-04-19 02:07:35,845 | base_dataset.py | line 301: | PQ SQ RQ N 2020-04-19 02:07:35,845 | base_dataset.py | line 302: -------------------------------------- 2020-04-19 02:07:35,845 | base_dataset.py | line 304: All | 53.9 79.9 66.1 19

And I have tested 16 GPUs setting with this repo provided weight. The result is ~57 PQ instead of 59.3. python upsnet/upsnet_end2end_test.py --cfg upsnet/experiments/upsnet_resnet50_cityscapes_16gpu.yaml --weight_path ./model/upsnet_resnet_50_cityscapes_12000.pth What might be the problem with the performance gap? Did u trained from imagenet/coco weight or used coarse data?

[txfs1926]

I've found that some instances have been predicted to be VOID. My env: Pytorch 1.3.0/CUDA 10.2