search

NJU-LHRS / official-CMID

The official implementation of paper "Unified Self-Supervised Learning Framework for Remote Sensing Images".
78 stars 5 forks source link

Code Question: When masking an image #18

Closed lhj-3 closed 8 months ago

lhj-3 commented 9 months ago

I'm looking at your research. Thank you for your wonderful work.

However, a question arises while looking at the code, leaving an issue. During masking processing, "Blockwise Mask Generator" confirmed that changing to mean value instead of zero resulted in discrepancy between image and mask.

Is there a process to deal with afterwards because there is something I didn't find?? Or is this error updated??

Thank you.

ayh456 commented 9 months ago

Hello, I would like to ask how the T-sne visualization mentioned in your paper is implemented, is there any code? Thank you

pUmpKin-Co commented 9 months ago

Hi~ @lhj-3 Thanks for your interest. I don't actually understand your questions. We utilize the average mean to fill the masked patches and after patch embedding we add mask token to the masked patches and regress them as training target. The mask is just used for collect the correct masked patches for loss computation. I hope my answer can help you.

pUmpKin-Co commented 9 months ago

Hi~ @ayh456. I'm sorry I don't have access to the original project environment at this moment. You can refer this implementations for t-sne visualization. I mostly borrow the code from this implementation.

pUmpKin-Co commented 8 months ago

Closed as long periods of inactivity, feel free to reopen if there is any problem.