Doubt with subordinate image reconstruction

[Manjuphoenix]

Hi,
I read the paper and found it interesting and wanted to check the performance for images that are out of domain (adverse domain shifts), but after running the colab notebook that is provided in the repository the subordinate image reconstruction does not match the results on paper (qualitative result, Figure 1 of page 1)

The colab notebook link is: https://colab.research.google.com/drive/1cuT_bAgxsKHDMDNF_DUIJzaTEkGQ7ydC?usp=drive_link

Are there any hyper-parameters changes in the code or if the checkpoint that I'm using is not the right one?

[Manjuphoenix]

As seen in the highlighted part of the picture above, the checkpoint you have provided seems to be having an issue.
Using this checkpoint the subordinate reconstruction result seems to have some issue.

Please clarify if there is some issue from your end or is there something we are missing.

[zyzkevin]

Hi,

Thank you for your interest in our work. As shown in the log you have provided, the keys in the weights did not match the model perfectly. 81adb64 caused this issue, since the weight file is single-branched with different key names. I will try to get this fixed soon. Meanwhile, you can test the demo notebook in 8819825, which is working, or you can help fix this issue and open a PR.