Meet problem when reproduce the performance reported in the paper.

[siyueyu]

Hi, I find this paper fascinating. I want to follow the work. Thus, I try to reproduce the performance. However, if I directly use the provided generated data to train the deeplabv3(101), I can only obtain the performance with the highest 57 mIoU. Could you please offer me some tips when training the segmentation network?

[tranxuantuyen]

Hi @siyueyu I faced a similar problem, where the performance just around 55-57 mIoU. Have you figured out the issue yet. Thanks

[theignorantt]

Hi @siyueyu I faced a similar problem, where the performance just around 55-57 mIoU. Have you figured out the issue yet. Thanks

the same problem. Have you figured out the issue? Thanks

[quang-ngh]

Thanks for addressing the problem! We will update the training config with EMA (as we did).

[theignorantt]

Thanks for addressing the problem! We will update the training config with EMA (as we did).

looking forward to your update! I tried to modify several places, such as loading the pre-trained network of r101, etc., to increase miou to 62, but it still did not reach 64. In addition, if possible, I hope to provide the code to get the prompt file.

[zbf1991]

Thnaks for your reply, when will u update your EMA method?

Thanks for addressing the problem! We will update the training config with EMA (as we did).