Imbalance datasets tackling

[narges-tk]

Hi! Thanks for the nice work! While training on an imbalanced dataset, I noticed a significant improvement in the accuracy of the class with the fewest samples (the last class) around the 14th epoch. However, after that, its accuracy started to decline again. This behavior is quite interesting!
I was wondering if you could advise me on how to address data imbalance in KPConvX. Thanks in advance for your guidance!
Validation ious are as follows:
0.921 0.805 0.577 0.035 0.914 0.000
0.931 0.785 0.651 0.046 0.913 0.000
0.903 0.770 0.580 0.102 0.909 0.000
0.896 0.768 0.583 0.099 0.908 0.000
0.906 0.776 0.606 0.078 0.911 0.000
0.935 0.831 0.682 0.106 0.912 0.000
0.908 0.767 0.610 0.107 0.914 0.000
0.763 0.247 0.494 0.096 0.870 0.062
0.754 0.195 0.503 0.245 0.865 0.112
0.897 0.757 0.626 0.100 0.910 0.029
0.769 0.249 0.495 0.200 0.874 0.004
0.862 0.532 0.585 0.104 0.881 0.072
0.849 0.739 0.608 0.134 0.908 0.037
0.906 0.805 0.620 0.132 0.902 0.673
0.855 0.712 0.639 0.106 0.903 0.076
0.760 0.223 0.507 0.148 0.870 0.021
0.760 0.206 0.481 0.179 0.868 0.092
0.923 0.807 0.712 0.108 0.908 0.000
0.903 0.768 0.637 0.131 0.906 0.024
0.800 0.338 0.521 0.182 0.870 0.029
0.762 0.144 0.479 0.216 0.856 0.000
0.908 0.806 0.721 0.108 0.912 0.000
0.905 0.778 0.633 0.104 0.909 0.000
0.754 0.224 0.502 0.201 0.874 0.009
0.878 0.749 0.616 0.124 0.907 0.000
0.901 0.612 0.618 0.172 0.865 0.000
0.738 0.177 0.501 0.145 0.873 0.022
0.912 0.810 0.736 0.085 0.915 0.000
0.895 0.690 0.698 0.083 0.889 0.000
0.847 0.727 0.653 0.114 0.904 0.048
0.914 0.832 0.635 0.060 0.916 0.000
0.891 0.774 0.621 0.112 0.908 0.018
0.922 0.820 0.676 0.114 0.901 0.000
0.892 0.751 0.638 0.149 0.900 0.000
0.873 0.746 0.633 0.134 0.904 0.051
0.880 0.747 0.654 0.098 0.909 0.000
0.751 0.221 0.500 0.216 0.873 0.044
0.892 0.766 0.644 0.150 0.897 0.000
0.908 0.483 0.648 0.094 0.856 0.000
0.884 0.747 0.655 0.109 0.909 0.000
0.890 0.763 0.641 0.138 0.906 0.000
0.886 0.775 0.727 0.100 0.906 0.000
0.891 0.768 0.732 0.093 0.907 0.000
0.878 0.750 0.659 0.130 0.907 0.000

[lgleznah]

Hello! I have just uploaded a PR that will hopefully be helpful to you 🙂

[warriorsniu]

This seems to be a bug in the validation metric calculation process. You can see that this outlier value is always the average of other categories, which means (0.906, 0.805, 0.620, 0.132, 0.902). mean()=0.673. I also found this problem in my experiment on S3DIS, and I guess it should be an error in the confusion matrix calculation process.