About training accuracy

[HeQibin]

I repeated experiments on U2PL according to your settings for many times, and the accuracy in the paper can not be achieved under the condition of eight V100. Can you provide the training log to help check details?

[Haochen-Wang409]

On which dataset?

[Haochen-Wang409]

Our checkpoints here for VOC and here for Cityscapes.

[HeQibin]

Thanks for providing checkpoint, but I'm more concerned about whether the semi-supervised training can achieve the same accuracy through U2PL.

On which dataset?

Cityscapes.

[Haochen-Wang409]

On Cityscapes dataset, we apply slide-window evaluation following AEL, which means we should run sh eval.sh after training.
Therefore, the mIoU reported during training is not the final performance, which is quite different from VOC.

Did you run sh eval.sh after training?

[Haochen-Wang409]

For instance, under 1/2 partition protocol on Cityscapes, the final result is 79.12 with AEL, but the mIoU is around 75 during training.

[Haochen-Wang409]

By the way, can you reproduce the results on VOC?

[HeQibin]

For instance, under 1/2 partition protocol on Cityscapes, the final result is 79.12 with AEL, but the mIoU is around 75 during training.

Thanks. I'll try eval.sh.