Adding new classes to a pre-trained model

[d33dler]

Assuming the model was trained on kitti/waymo dataset with 3 classes, can we add new classes and continue training with the addition of these new classes? (by changing the dataset & model configs, specifically for PVRCNN++) Thanks @jihanyang

[jihanyang]

You can have a try. But as the cls head will change with more class, I am not sure how the performance will behave.

[d33dler]

You can have a try. But as the cls head will change with more class, I am not sure how the performance will behave.

@jihanyang @sshaoshuai Thanks! So it is preferred to train with all classes from the beginning? Approximately how many classes could be used until the performance begins to degrade? Currently, adding a new class to a pre-trained model's config results in this info output and the model doesn't detect anything anymore, understandably the missing weights (are 0s?) obliterate the regressor and classifier, yes? You are not sure if this can be corrected by continuing training and updating the weights, correct? One last question, what does CLASS_AGNOSTIC mean here in the RPN head - I see that the heatmap (which relates to the abbreviation hm.) shape will differ, what is its role?

[d33dler]

@jihanyang some insight would be greatly appreciated!

[github-actions[bot]]

This issue is stale because it has been open for 30 days with no activity.

[github-actions[bot]]

This issue was closed because it has been inactive for 14 days since being marked as stale.