Unable to reproduce the good results (Acc. 56.7%) of ModelNet to ScanNet without SPST.

[Jhonve]

Using the first version of your code (using batch normalization and TransNorm), I got the results of

02-11-21 19:21:08: Test - Target 0, acc: 0.5415, avg acc: 0.4271, cls loss: 4.3767
02-11-21 19:21:08: target test accuracy: 0.5415, target test loss: 4.3767
02-11-21 19:21:08: Test confusion matrix:
02-11-21 19:21:08: 
[[ 14   2   2   0   0   1   2   0   1   4]
 [  0  23   5   0  12   5   3   1   3  33]
 [  0   4  61   0  15  14   8  15   2  27]
 [  0   0  47   4   6  18  39  16   3  16]
 [  0  19  55   1 561  31  32  44   2  56]
 [  0   1   8   3   4   8   0  12   1   4]
 [  0   0  36   0   0   3  15   6   0   1]
 [  0   0   0   0   2   0   0  23   0   0]
 [  2  15   3   0  42  11   3   8  31  19]
 [  0  12  12   3  14  24   1  13   4 218]]

And then, after you updated your code, I got the results of

04-11-21 02:37:22: Test - Target 0, acc: 0.5517, avg acc: 0.4728, cls loss: 4.2702
04-11-21 02:37:22: target test accuracy: 0.5517, target test loss: 4.9488
04-11-21 02:37:22: Test confusion matrix:
04-11-21 02:37:22: 
[[ 12   3   1   0   1   0   0   0   5   4]
 [  0  45   1   1   3   0   5   0   2  28]
 [  0  11  53   0  22  14   6   3   5  32]
 [  0  12  24   3  10  10  53   2   1  34]
 [  2  67  23   1 519  28  64  10  18  69]
 [  0   1   4   3   6  18   2   0   0   7]
 [  0   2  19   0   0   4  27   1   1   7]
 [  0   0   1   0   4   2   0  17   1   0]
 [  5  20   0   0  25   6   2   2  50  24]
 [  3  29   7   2  10   9   2   3   4 232]]

The accuracy of 0.5517 is different from your result of 56.7% (ModelNet to ScanNet without SPST) in Table 1 in your paper and the average accuracy of 10 classes is 0.4728, right? However, the reported accuracy in your paper is 52.2%.
Are there any problems with your current code or any different details from your paper?
Thanks a lot.