I would like to know if the code for the attention mechanism interaction of your MIA module has been commented out. Does this seem to be inconsistent with the description in your paper?

[L0310]

OR because the default model of the code is resnet50, so convolution is used instead of the attention mechanism? If that's the case, could you please provide me with the code for SwinTransformer?

[Ayews]

Thank you for your reminder. I have noticed that the code I uploaded may be a version of the experimental code, rather than the final version used. Due to the passage of time, I am unable to find the exact final version. Perhaps you can refer to the version before this commit: 6ec39fb#diff-8c58c5ced60bd1d98818e03732693265424cd759a2876093c6bac52560c61462
It is worth mentioning that the core idea of MIA is to let high-level features guide low-level features, and the implementation is not limited to attention mechanism.