Imbalanced dataset and class_weight hyperparameter detectnet_v2 model

I read TAO FAQ page where class_weight was mentioned, but no additional information is mentioned either in FAQ page nor in the detectnet_v2 creating a configuration file page.

class_weight: 1.0
coverage_foreground_weight: 0.05

If my classes are imbalanced I should increase the class_weight, but how much? Is there a limit?
Let’s say I have 10 000 objects in one class, but only 2000 in another, how much should I increase the second class_weight? Should I also change coverage_foreground_weight? Where canI get more information about that?

Refer to Frequently Asked Questions — TAO Toolkit 3.22.02 documentation

Hi, Yes. I mentioned before as well that I read it, but didn’t get answet to my questions:

If my classes are imbalanced I should increase the class_weight, but how much? Is there a limit?

Let’s say I have 10 000 objects in one class, but only 2000 in another, how much should I increase the second class_weight? Should I also change coverage_foreground_weight?

Please change the class_weight.
For example, 10000 objects class is set to 1.0 , another class is set to 5.0.

There is no update from you for a period, assuming this is not an issue any more.
Hence we are closing this topic. If need further support, please open a new one.
Thanks

This topic was automatically closed 14 days after the last reply. New replies are no longer allowed.