Inconsistency between paper and code

[Turlan]

Thanks for releasing codes! I found several differences between the paper and the released code:

1. The batch size. In your paper, the batch size of segmentation model was set as 1, in your code, the default batch size is 4.
2. The ASPP module is different with a common design. You use concat and conv instead of sum https://github.com/RogerZhangzz/CAG_UDA/blob/b6fbea6435b2b5bc42b8043fc057878f2192dce4/models/aspp.py#L81
3. The decoder uses low-level feature as input https://github.com/RogerZhangzz/CAG_UDA/blob/b6fbea6435b2b5bc42b8043fc057878f2192dce4/models/decoder.py#L36

I wanna know if I just set the batch size to 1 as the paper described, is there a signifigent performance drop? What's the reason of the design for the decoder? Is it necessary to use low-level feature?

[Turlan]

And could you release/explain the code for updating CAC between stages? In the released code, there exists a moving-average implementation. Which one did you use? https://github.com/RogerZhangzz/CAG_UDA/blob/b6fbea6435b2b5bc42b8043fc057878f2192dce4/models/adaptation_model.py#L529

[RogerZhangzz]

Hi there, Thanks for your suggestion.

1. The batch size was set to 1 when I trained the model. Then in my re-implementation, I set it to 4 for a quick convergence. Both settings would be fine for the final performance. Plz note that you have to increase the crop size if you choose to use 1 as the batch size.

2. \ 3. Very thank you for what you have pointed out. I will check it out.

3. I have uploaded the CAC code (cac.py) and modified the instructions.