It might sound absurd, but the test dataset(#2 Test Phase) I downloaded before phase 2 started and the test dataset I downloaded after Phase 2 began are completely different?
For example, before:
$ md5sum ../../dev/p2/1980.tif
b2e0b42876bc4331ca682fd0fe500b08 ../../dev/p2/1980.tif
after:
$ md5sum 1980.tif
656fa7c94a72c4b150d7fd5ac5ff9c02 1980.tif
So some people have results close to zero on the phase 2 test, because they just use the old test dataset?
Also, if the test dataset was changed, it could also affect some parameters during training(eg. normalization's means and std or effect of pseudo-label).
Or did I misunderstand something, like downloading wrong data?
Posted by: OverfitMatser @ March 11, 2024, 3:09 p.m.if anyone have encountered this issue in the competition, please provide feedback :-)
AND if your test score is closed to 0, try to redownload the test dataset and check. Hope it will be helped.
Posted by: OverfitMatser @ March 11, 2024, 3:23 p.m.Dear participants, the test dataset has been reworked last week, mainly by adding new independent data, to limit the correlations reported by some participants during phase 1.
We have just sent an email to all participants to let them know.
Please download the test dataset again if you downloaded it before the start of phase 2.
We apologize for the inconvenience.
Thank you for your timely notification.
I need to rework my model now, good luck to me. (>ะด<)
The game isn't over until the last moment. ๐ถ๏ธ๐
Posted by: OverfitMatser @ March 11, 2024, 3:38 p.m.