Reproduction Issue of ResNet-50 with dim=2048 about eigenplaces HOT 4 CLOSED

Hi @chenshunpeng, thanks a lot!
The hyperparameters look correct. I just realized that there is a small logging bug at this line of the train.py, because the info.log should show the number of training images. That does not change the training, but it makes the log more clear.
Could you show me the number of images within sf_xl/processed/train? The size in MB is not really relevant as it is dependent on the file system

from eigenplaces.

Of course, here's the breakdown of the image count in each folder:

(EigenPlaces) chenshunpeng@h3c-R5300-G5:~/data/sf_xl/processed/train$ find . -maxdepth 1 -mindepth 1 -type d -exec sh -c 'echo -n "{}: "; find "{}" -maxdepth 1 -type f \( -iname \*.jpg -o -iname \*.jpeg -o -iname \*.png \) | wc -l' \;
./37.70: 256914
./37.71: 343182
./37.72: 347208
./37.73: 538302
./37.74: 482880
./37.75: 541038
./37.76: 714156
./37.77: 785682
./37.78: 753486
./37.79: 535620
./37.80: 307878
./37.81: 1074

from eigenplaces.

Ok so the issue seems to be in the number of training images, because I have this amount of images

37.70 163665
37.71 228643
37.72 235935
37.73 347857
37.74 311148
37.75 335999
37.76 412361
37.77 458764
37.78 438066
37.79 309385
37.80 188416
37.81 853

It's quite strange that you have less than me. Perhaps instead of downloading the panoramas folder you downloaded something else? You can download the all_images_paths.txt file and see which images are mismatched

from eigenplaces.

Okay, I understand. My training set is incomplete. I will re-download the whole 360° panorama and re-enact it. Thank you for your guidance.

from eigenplaces.