Discrepancy Between Paper Mentioned 3x3 Convolution and Actual Code Implementation

[Shrikebird]

I have been going through the project and noticed what seems to be an inconsistency between the project paper and the actual implementation, which I would like to discuss.
According to the project's paper, there is a mention of a 3x3 convolution step that appears in the final part of the PIM. However, upon examining the project code model.py from 339 to 343, this particular convolution step does not seem to be implemented in the source code. There is only a 1x1 and then it jumps to the next level.
I am unsure whether this discrepancy is due to my misunderstanding of the codebase or if there is indeed a gap between the described methodology and its implementation. It would be great to get some insights from the community on this matter.

[StarSailor27]

I have the same question. Adding a 3x3 layer at the end of PIM seems to be consistent with the implementation of the paper, but the code uploaded to github only implements up to a 1x1 layer. I'm curious which is the correct implementation.